CN103886344A - Image type fire flame identification method - Google Patents

Abstract

The invention discloses an image-type fire flame recognition method, which comprises the following steps: 1. Image acquisition; 2. Image processing: 201. Image preprocessing; The classification model is a support vector machine model for classifying two categories with flames and no flames; the establishment process of the binary classification model is as follows: Ⅰ, image information collection; Ⅱ, feature extraction; Ⅲ, training sample acquisition; Ⅳ, binary classification model Establishment: Ⅳ-1. Kernel function selection; Ⅳ-2. Classification function determination: use conjugate gradient method to optimize parameters C and D, and then convert optimized parameters C and D into γ and σ ² ; Ⅴ. Two classification Model training. The method of the present invention has simple steps, convenient implementation, simple operation, high reliability and good use effect, and can effectively solve the problems of low reliability, high rate of false positives and false positives and poor use effect in the complex environment existing in the existing video fire detection system. Bad question.

Description

An Image-based Fire Flame Recognition Method

technical field

The invention belongs to the technical field of image acquisition and processing, and in particular relates to an image-based fire flame identification method.

Background technique

Fire is one of the major disasters in mines, which seriously threatens human health, natural environment and safe production of coal mines. With the advancement of science and technology, automatic fire detection technology has gradually become an important means of monitoring and fire warning. Nowadays, in coal mines, fire prediction and detection are mainly based on monitoring the temperature effect of fire, combustion products (effects of smoke and gas) and electromagnetic radiation effects, but the above-mentioned existing detection methods are in terms of sensitivity and reliability. Both have yet to be improved, and cannot respond to early fires, so they are not compatible with the increasingly stringent fire safety requirements. Especially when there are occluders in a large space, the spread of fire combustion products in the space will be affected by the height and area of the space. The ordinary point-type smoke-sensing and temperature-sensing fire detection and alarm system cannot quickly collect the information of the smoke temperature change emitted by the fire. , Only when the fire develops to a certain extent, will it respond, so it is difficult to meet the requirements of early detection of fire. With the rapid development of video processing technology and pattern recognition technology, fire detection and early warning methods are developing in the direction of image, digitization, scale and intelligence. The fire detection technology based on video surveillance has the advantages of wide detection range, short response time, low cost, and not affected by the environment. Combined with computer intelligence technology, it can provide more intuitive and richer information, which is of great significance to the safe production of coal mines.

At present, video fire detection technology is still in its infancy at home and abroad, and there are differences in detection methods, working principles, system structures, and practical occasions. The Alarm Eye VISFD distributed intelligent image fire detection system developed by Intellvision Company, the dual-band monitoring of infrared and ordinary cameras of Bosque Company of the United States, and the VSD-8 system jointly developed by Swiss ISL Company and Magnox Electric Company for power station fire monitoring. Domestic research on fire detection and automatic fire extinguishing is currently done by the State Key Laboratory of Fire Science at the University of Science and Technology of China. In addition, Tianjin University, Xi'an Jiaotong University, Shenyang University of Technology and Shanghai Jiaotong University have also done active research, but the above image fire The detection system is mostly used for fire detection in power stations, buildings, warehouses, etc., and its application in underground coal mines is relatively small. In recent years, many researchers at home and abroad have carried out in-depth research on the key technology of this image-based fire detection system, that is, the flame image analysis algorithm, and made great contributions, mainly in the following aspects: ①Based on the static characteristics of the flame such as pixels Video flame detection method based on spectral characteristics such as brightness and chromaticity and regional structure such as shape and outline; ② Video flame detection method based on flame color movement area; ③ Video detection method based on flame stroboscopic characteristics and time-frequency characteristics. However, when the research on the above-mentioned flame image analysis algorithm is applied to the existing video fire detection system, there are certain limitations to varying degrees. It cannot effectively remove interference in complex scenes, and the system has serious false positives and negative negatives. Therefore, there is a lack of an image-based fire flame recognition method with simple steps, convenient implementation, easy operation, high reliability, and good use effect, which can effectively solve the low reliability in the complex environment of the existing video fire detection system. , high rate of false positives and false negatives, poor use effect and other issues.

Contents of the invention

The technical problem to be solved by the present invention is to provide an image-type fire flame identification method for the above-mentioned deficiencies in the prior art. The video fire detection system has problems such as low reliability in complex environments, high rate of false positives and false positives, and poor use effect.

In order to solve the above-mentioned technical problems, the technical solution adopted in the present invention is: an image-type fire flame recognition method, which is characterized in that the method includes the following steps:

Step 1. Image collection: using an image collection unit to collect digital images of the area to be detected according to a preset sampling frequency f _s , and synchronously transmitting the digital images collected at each sampling moment to the processor; the images The acquisition unit is connected with the processor;

Step 2, image processing: the processor performs image processing on the digital images collected at each sampling moment in step 1 according to the time sequence, and the processing methods for the digital images collected at each sampling moment are the same; When the digital image collected at any sampling moment is processed, the following steps are included:

Step 201, image preprocessing, the process is as follows:

Step 2011, image reception and synchronous storage: the processor synchronously stores the received digital image collected at the current sampling time in the data memory, and the data memory is connected to the processor;

Step 2012, image enhancement: the digital image collected at the current sampling moment is enhanced by the processor to obtain the enhanced digital image;

Step 2013, image segmentation: segment the digital image enhanced in step 2012 by the processor to obtain the target image;

Step 202, fire identification: use the pre-established binary classification model to process the target image described in step 2013, and obtain the fire state category of the area to be detected at the current sampling moment; the fire state category includes flame and no flame Two categories, the two classification model is a support vector machine model for classifying two categories with flame and no flame;

The establishment process of the two classification models is as follows:

Step 1. Image information acquisition: using the image acquisition unit to acquire multi-frame digital images 1 of the area to be detected when a fire occurs and multi-frame digital images 2 of the area to be detected when no fire occurs;

Step II, feature extraction: perform feature extraction on the multi-frame digital image one and the multi-frame digital image respectively, and extract a group of characteristic parameters that can represent and distinguish the digital image from each digital image, and the The group feature parameters include M feature quantities, and the M feature quantities are numbered, and the M feature quantities form a feature vector, where M≥2;

Step III, acquisition of training samples: from the feature vectors of multiple frames of the digital image one and multiple frames of the digital image two obtained after feature extraction in step II, respectively select the feature vectors and The eigenvectors of the second digital image in frame m2 form a training sample set; wherein m1 and m2 are positive integers and m1=40～100, m2=40～100; the number of training samples in the training sample set is m1+m2 indivual;

Step Ⅳ, two-category model establishment, the process is as follows:

Step IV-1, kernel function selection: select radial basis function as the kernel function of the binary classification model;

Step IV-2, determination of classification function: After the factor γ to be penalized and the kernel parameter ^σ2 of the radial basis function selected in step IV-1 are determined, the classification function of the binary classification model is obtained, and the binary classification is completed Model establishment process; where, γ=C ^-2 , σ=D ^-1 , 0.01<C≤10, 0.01<D≤50;

When determining the penalty factor γ and the kernel parameter σ ² , the parameters C and D are optimized using the conjugate gradient method first, and the optimized parameters C and D are obtained, and then according to γ=C ^-2 and σ=D ^-1 The optimized parameters C and D are converted into penalty factor γ and kernel parameter σ ² ;

Step V, binary classification model training: input the m1+m2 training samples in the training sample set mentioned in step III to the binary classification model established in step IV for training.

The above-mentioned image type fire flame recognition method is characterized in that: the total number of training samples in the training sample set described in step III is N and N=m1+m2; N training samples in the sample set are numbered, the numbering of the pth training sample in the training sample set is p, p is a positive integer and p=1, 2, ..., N; the pth training sample is written as (x _p ,y _p ), where x _p is the feature parameter of the pth training sample, y _p is the category number of the pth training sample and y _p =1 or -1, where the category number is 1, which means there is flame, and the category number is -1 means no flame;

When using the conjugate gradient method to optimize parameters C and D in step IV-2, use the m1+m2 training samples in the training sample set mentioned in step III to optimize, and the optimization process is as follows:

为矩阵s中除去第p个元素后的其余元素组成的向量；s(p)为矩阵s的第p个元素，(A^-1)(p^-,p)为矩阵A^-1的第p列除去第p个元素后的其余元素组成的列向量，(A^-1)(p,p)为矩阵A^-1的第p列的第p个元素；

为矩阵

的第p列除去第p个元素后的其余元素组成的列向量，矩阵为矩阵K的增广矩阵，其中矩阵 $K=\left|\begin{array}{cccc}K(x_1,x_1)& K(x_1,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_1,x_N)\\ K(x_2,x_1)& K(x_2,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_2,x_N)\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ K(x_N,x_1)& K(x_N,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_N,x_N)\end{array}\right|;$ 矩阵A^-1表示矩阵A的逆矩阵，矩阵 $A=\left[\begin{array}{cc}K+C^2\&CenterDot;I& I_N\\ {I_N}^T& 0\end{array}\right],$ 其中矩阵I为单位矩阵，矩阵I_N=[1,1,…,1]^T，T表示矩阵的转置，矩阵I_N中包含N个元素且N个元素均为1；矩阵s=A^-1Y，矩阵 $Y=\left[\begin{array}{c}{y}_1\\ y_2\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ y_N\\ 0\end{array}\right],$ 其中y₁、y₂、…、y_N分别为所述训练样本集中N个训练样本的类别；Step 1, objective function determination: $\mathrm{sse}(C,\mathrm{D.})=\frac{1}{2}\underset{p=1}{\overset{\mathrm{<figure-callout\; id="2"\; label="N\; e\; p"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">N</figure-callout>}}{\mathrm{\&Sigma;}}}{e}_p^{}{}_2^{}=\frac{1}{2}\underset{p=1}{\overset{N}{\mathrm{\&Sigma;}}}{[\tilde{K}(p,p^{-})\&CenterDot;{\mathrm{the\; s}}_p-\mathrm{the\; y}\left(p\right)]}^2$ (1), where sse(C, D) is the sum of the squares of the leave-one-out prediction error, p is the serial number of each training sample in the training sample set, and e _p is the p-th training result of the binary classification model established in step IV. The sample prediction error and $e_p=\tilde{K}(p,p^{-})\&Center\; Dot;{\mathrm{the\; s}}_p-\mathrm{the\; y}\left(p\right);$ in, ${\mathrm{the\; s}}_p=\mathrm{the\; s}\left(p^{-}\right)-\frac{\mathrm{the\; s}\left(p\right)}{\left(A^{-1}\right)(p,p)}\left(A^{-1}\right)(p^{-},p),$ In the formula

is a vector composed of the remaining elements after removing the p-th element in the matrix s; s(p) is the p-th element of the matrix s, (A ^-1 )(p ^- ,p) is the p-th column of the matrix A ^-1 A column vector composed of the remaining elements after removing the p-th element, (A ^-1 )(p,p) is the p-th element of the p-th column of the matrix A ^-1 ;

for the matrix

The column vector and matrix composed of the remaining elements after removing the p-th column of the p-th column is the augmented matrix of matrix K, where the matrix $K=\left|\begin{array}{cccc}K(x_1,x_1)& K(x_1,x_2)& \&Center\; Dot;\&CenterDot;\&CenterDot;& K(x_1,x_N)\\ K(x_2,x_1)& K(x_2,x_2)& \&Center\; Dot;\&Center\; Dot;\&Center\; Dot;& K(x_2,x_N)\\ \&Center\; Dot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&Center\; Dot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ K(x_N,x_1)& K(x_N,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_N,x_N)\end{array}\right|;$ Matrix A ^-1 represents the inverse matrix of matrix A, matrix $A=\left[\begin{array}{cc}K+{\mathrm{<figure-callout\; id="2"\; label="C"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">C</figure-callout>}}^2\&Center\; Dot;I& I_N\\ {{\mathrm{<figure-callout\; id="0"\; label="I\; N\; T"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">I</figure-callout>}}_N}^T& 0\end{array}\right],$ Among them, the matrix I is the identity matrix, the matrix I _N =[1,1,…,1] ^T , T represents the transposition of the matrix, the matrix I _N contains N elements and all N elements are 1; the matrix s=A ^{- 1} Y, matrix $Y=\left[\begin{array}{c}{\mathrm{the\; <figure-callout\; id="1"\; label="y"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">y</figure-callout>}}_1\\ {\mathrm{the\; <figure-callout\; id="2"\; label="y"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">y</figure-callout>}}_2\\ \&Center\; Dot;\\ \&Center\; Dot;\\ \&Center\; Dot;\\ {\mathrm{the\; <figure-callout\; id="0"\; label="y\; N"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">y</figure-callout>}}_N\\ 0\end{array}\right],$ Wherein y ₁ , y ₂ , ..., y _N are respectively the categories of the N training samples in the training sample set;

Step ii. Initial parameter setting: determine the initial values C ₁ and D ₁ of parameters C and D respectively, and set the recognition error threshold ε and ε>0;

计算得出步骤ⅰ中目标函数对C_k和D_k的梯度g_k，k为迭代次数且k=1、2、…；若||g_k||≤ε，停止计算，此时C_k和D_k分别为优化后的参数C与D；否则，进入步骤ⅳ；Step Ⅲ, the gradient g _k calculation of the current iteration: according to the formula

Calculate the gradient g _k of the objective function to C _k and D _k in step i, k is the number of iterations and k=1, 2, ...; if ||g _k ||≤ε, stop the calculation, at this time C _k and D _k are the optimized parameters C and D respectively; otherwise, go to step ⅳ;

in, $\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k}=\underset{p=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_p\&CenterDot;\tilde{K}(p,p^{-})\&CenterDot;\frac{\&PartialD;{\mathrm{the\; s}}_p}{\&PartialD;C_k};$

的第p列除去第p个元素后的其余元素组成的列向量；s(p^-)为矩阵s中除去第p个元素后的其余元素组成的向量，e_p为步骤Ⅳ中所建立的二分类模型对第p个训练样本的预测误差； $\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k}=\underset{p=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{e}_p\&Center\; Dot;[\frac{\&PartialD;\tilde{K}(p,p^{-})}{\&PartialD;{\mathrm{D.}}_k}\&CenterDot;\mathrm{the\; s}\left(p^{-}\right)+\tilde{K}(p,p^{-})\&Center\; Dot;\frac{\&PartialD;{\mathrm{the\; s}}_p}{\&PartialD;{\mathrm{D.}}_k}],$ In the formula for the matrix

s(p ^- ) is a vector composed of the remaining elements after removing the p-th element in the matrix s, and e _p is the two elements established in step IV The prediction error of the classification model for the pth training sample;

Step ⅳ. Calculation of the search direction d _k of the current iteration: according to the formula $d_k=\left\{\begin{array}{cc}-g_k& k=1\\ -g_k+{\mathrm{\&beta;}}_k{d}_{k-1}& k\&Greater\; Equal;2\end{array}\right.,$ Calculate the search direction d _k of the current iteration, where d _k-1 is the search direction of the k-1th iteration, β _k =||g _k || ² /||g _k-1 || ² , g _{k- 1} is the gradient of the k-1th iteration;

Step ⅳ, determine the search step size λ _k of the current iteration: search along the search direction d _k determined in step ⅳ, and find out that satisfies the formula

表示在（0，+∞）中寻找使

达到最小值的步长λ_k； $\mathrm{sse}(C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k},{\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k})=\underset{{\mathrm{\&lambda;}}_k>0}{\min }\mathrm{sse}(C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k},{\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k})$ The search step size λ _k , where

Indicates that in (0, +∞) to find the

the step size λ _k to reach the minimum;

Step ⅵ, according to the formula $C_{k+1}=C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k}$ and ${\mathrm{D.}}_{k+1}={\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k},$ Calculate C _k+1 and D _k+1 ;

Step ⅶ, set k=k+1, and then return to step Ⅲ for the next iteration;

该径向基函数的回归函数为

式中α_t和b均为回归参数，s为正整数且s=1、2、…、N，t为正整数且t=1、2、…、N。The radial basis function selected in step IV-1 is

The regression function of the radial basis function is

In the formula, α _t and b are regression parameters, s is a positive integer and s=1, 2, ..., N, t is a positive integer and t = 1, 2, ..., N.

The above-mentioned image-based fire flame recognition method is characterized in that: in step II, M=6, and the six feature quantities are area, similarity, moment characteristic, density, texture characteristic and stroboscopic characteristic.

The above-mentioned image-type fire flame recognition method is characterized in that: when C1 _{and D1} are determined in step ii, _C1 and _D1 are determined by grid search method or random value extraction; When determining C ₁ and D ₁ by numerical method, C ₁ is a value randomly selected in (0.01, 1], and D ₁ is a value randomly selected in (0.01, 50]; grid search method is used to determine C ₁ and D ₁ are determined, first divide the grid with a step size of 10 ^-3 , then use C and D as independent variables and the objective function described in step i as the dependent variable to make a three-dimensional grid map, and then use grid search Find multiple sets of parameters of C and D, and finally take the average value of multiple sets of parameters as C ₁ and D ₁ .

The above-mentioned image-based fire flame recognition method is characterized in that: when image enhancement is performed in step 2012, an image enhancement method based on fuzzy logic is used for enhancement processing.

Above-mentioned a kind of image type fire flame identification method is characterized in that: when adopting the image enhancement method based on fuzzy logic to carry out enhancement processing, the process is as follows:

Step 20121, transform from the image domain to the fuzzy domain: according to the membership function ${\mathrm{\&mu;}}_{\mathrm{gh}}=T\left(x_{\mathrm{gh}}\right)=\left\{\begin{array}{cc}{x}_{\mathrm{gh}}/x_T& x_{\mathrm{gh}}\&le;x_T\\ x_{\mathrm{gh}}/x_{\max }& x_{\mathrm{gh}}>x_T\end{array}\right.$ (7), the gray value of each pixel of the image to be enhanced is mapped to the fuzzy membership degree of the fuzzy set, and the fuzzy set of the image to be enhanced is obtained accordingly; x _gh is in the image to be enhanced The gray value of any pixel point (g, h), X _T is the gray threshold selected when the image to be enhanced is enhanced using the image enhancement method based on fuzzy logic, and X _max is the image to be enhanced The maximum gray value of ;

Step 20122, use the fuzzy enhancement operator in the fuzzy domain to perform fuzzy enhancement processing: the adopted fuzzy enhancement operator is μ' _gh =I _r (μ _gh )=I _r (I _r-1 μ _gh ), where r is The number of iterations and it is a positive integer, r=1, 2, ...; where $I_1\left({\mathrm{\&mu;}}_{\mathrm{gh}}\right)=\left\{\begin{array}{cc}{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="2"\; label="gh"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">gh</figure-callout>}}^2/{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="0"\; label="c"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">c</figure-callout>}}& 0\&le;{\mathrm{\&mu;}}_{\mathrm{gh}}\&le;{\mathrm{\&mu;}}_c\\ 1-{(1-{\mathrm{\&mu;}}_{\mathrm{gh}})}^2/(1-{\mathrm{\&mu;}}_c)& {\mathrm{\&mu;}}_c\&le;{\mathrm{\&mu;}}_{\mathrm{gh}}\&le;1\end{array}\right.,$ In the formula, μ _c =T(X _C ), where X _C is the transition point and X _C =X _T ;

(6)，将模糊增强处理后得到的μ'_gh进行逆变换，获得增强处理后数字图像中各像素点的灰度值，并获得增强处理后的数字图像。Step 20123, inverse transformation from fuzzy domain to image domain: according to the formula

(6) Perform inverse transformation on the _μ'gh obtained after the fuzzy enhancement processing, obtain the gray value of each pixel in the digital image after the enhancement processing, and obtain the digital image after the enhancement processing.

The above-mentioned image-type fire flame recognition method is characterized in that: before transforming from the image domain to the fuzzy domain in step 20121, first adopt the maximum inter-class variance method to select the gray scale threshold X _T ; Before selecting the threshold value X _T of the image to be enhanced, find out all the gray values whose number of pixels is 0 from the gray range of the image to be enhanced, and use the processor (3) to average all the gray values found Marked as calculation-free grayscale value; when using the maximum inter-class variance method to select the grayscale threshold _XT , other grayscales in the grayscale variation range of the image to be enhanced except for the calculation-free grayscale value Calculate the inter-class variance value when the value is used as the threshold value, and find the maximum inter-class variance value from the calculated inter-class variance value, and the gray value corresponding to the found maximum inter-class variance value is gray degree threshold X _T .

The above-mentioned image-type fire flame recognition method is characterized in that: in step 1, the size of the digital image collected at each sampling moment is M1×N1 pixels;

When step 2013 performs image segmentation, the process is as follows:

Step 20131, two-dimensional histogram establishment: using a processor to establish a two-dimensional histogram of the pixel gray value and the neighborhood average gray value of the image to be segmented; any point in the two-dimensional histogram is denoted as (i , j), where i is the abscissa value of the two-dimensional histogram and it is the gray value of any pixel point (m, n) in the image to be segmented, and j is the ordinate value of the two-dimensional histogram And it is the neighborhood average gray value of the pixel point (m, n); the frequency of occurrence of any point (i, j) in the established two-dimensional histogram is recorded as C(i, j), and The frequency of occurrence of j) is denoted as h(i,j), where

Step 20132, fuzzy parameter combination optimization: the processor calls the fuzzy parameter combination optimization module, and uses the particle swarm optimization algorithm to optimize the fuzzy parameter combination used in the image segmentation method based on two-dimensional fuzzy partition maximum entropy, and obtains the optimized Fuzzy parameter combination;

In this step, before optimizing the combination of fuzzy parameters, according to the two-dimensional histogram established in step 20131, the functional relational expression of the two-dimensional fuzzy entropy when the image to be segmented is segmented is calculated, and the calculated The obtained two-dimensional fuzzy entropy function relation is used as the fitness function when the particle swarm optimization algorithm is used to optimize the combination of fuzzy parameters;

Step 20133, image segmentation: the processor uses the fuzzy parameter combination optimized in step 20132, and classifies each pixel in the image to be segmented according to the image segmentation method based on two-dimensional fuzzy partition maximum entropy, and correspondingly Complete the image segmentation process and obtain the segmented target image.

The above-mentioned image-type fire flame recognition method is characterized in that: the image to be segmented in step 20131 is composed of a target image O and a background image P; wherein the membership function of the target image O is μ _o (i, j)=μ _ox (i;a,b)μ _oy (j;c,d)(1); membership function of background image P μ _b (i,j)=μ _bx (i;a,b)μ _oy (j; c,d)+μ _ox (i;a,b)μ _by (j;c,d)+μ _bx (i;a,b)μ _by (j;c,d)(2);

In formulas (1) and (2), μ _ox (i;a,b) and μ _oy (j;c,d) are both one-dimensional membership functions of the target image O and both are S functions, μ _bx (i;a,b) and μ _by (j;c,d) are the one-dimensional membership functions of the background image P and both are S functions, μ _bx (i;a,b)=1-μ _ox (i;a,b), μ _by (j;c,d)=1-μ _oy (j;c,d), where a, b, c, and d are a combination of the target image O and the background image P Parameters that control the shape of the dimension membership function;

When calculating the functional relational expression of the two-dimensional fuzzy entropy in step 20132, first, according to the two-dimensional histogram established in step I, the minimum value g _min and the maximum value g _max and the minimum value s _min and maximum value s _max of the neighborhood average gray value are determined respectively;

The functional relational expression of the two-dimensional fuzzy entropy calculated in step 20132 is:

$h\left(P\right)=-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)})-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)})$ (3), in formula (3) $p\left(o\right)=\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_o(i,j,)h_{\mathrm{ij}},p\left(B\right)=\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_b(i,j,)h_{\mathrm{ij}},$ where h(i,j) is the frequency of occurrence of the point (i,j) mentioned in step I;

When the particle swarm optimization algorithm is used to optimize the fuzzy parameter combination in step 20132, the optimized fuzzy parameter combination is (a, b, c, d).

The above-mentioned image-type fire flame recognition method is characterized in that: when performing parameter combination optimization of two-dimensional fuzzy division maximum entropy in step 20132, the following steps are included:

Step Ⅱ-1. Particle swarm initialization: take a value of the parameter combination as a particle, and form multiple particles into an initialized particle swarm; denoted as (a _k , b _k , c _k , d _k ), where k is a positive integer and its k=1, 2, 3, ~, K, wherein K is a positive integer and it is the number of particles contained in the particle group, a _k is a random value of parameter a, and b _k is A random value of parameter b, c _k is a random value of parameter c, d _k is a random value of parameter d, a _k < b _k and c _k < d _k ;

Step Ⅱ-2, fitness function determination:

Will $h\left(P\right)=-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)})-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)})$ (3), as a fitness function;

Step Ⅱ-3. Particle fitness evaluation: evaluate the fitness of all particles at the current moment, and the fitness evaluation methods of all particles are the same; when evaluating the fitness of the kth particle at the current moment, first According to the fitness function determined in step Ⅱ-2, the fitness value of the kth particle at the current moment is calculated and recorded as fitnessk, and the difference between the calculated fitnessk and Pbestk is compared: when the comparison results in fitnessk > When Pbestk, Pbestk=fitnessk, and set Update to the position of the kth particle at the current time, where Pbestk is the maximum fitness value achieved by the kth particle at the current moment and it is the individual extremum of the kth particle at the current moment, is the individual optimal position of the kth particle at the current moment; among them, t is the current iteration number and it is a positive integer;

更新为当前时间适应度值最大的粒子的位置，其中gbest为当前时刻的全局极值，

为当前时刻的群体最优位置；After the fitness value of all particles at the current moment is calculated according to the fitness function determined in step II-2, the fitness value of the particle with the largest fitness value at the current moment is recorded as fitnesskbest, and fitnesskbest is compared with gbest Difference comparison: when the comparison shows that fitnesskbest>gbest, gbest=fitnesskbest, and the

Update to the position of the particle with the largest fitness value at the current time, where gbest is the global extremum at the current moment,

is the optimal position of the group at the current moment;

Step Ⅱ-4. Judging whether the iteration termination condition is satisfied: when the iteration termination condition is satisfied, the parameter combination optimization process is completed; otherwise, the position and speed of each particle at the next moment are obtained according to the update of the particle swarm optimization algorithm, and return to step Ⅱ -3; The iteration termination condition in step Ⅱ-4 is that the current iteration number t reaches the preset maximum iteration number I _max or Δg≤e, where Δg=|gbest-gmax|, where is the global extremum value of gbest at the current moment , gmax is the originally set target fitness value, e is a positive number and it is a preset deviation value.

Compared with the prior art, the present invention has the following advantages:

1. The steps of the method are simple, the design is reasonable, the implementation is convenient, and the input cost is low.

2. The image enhancement method adopted is simple in steps, reasonable in design and good in enhancement effect. According to the characteristics of low illumination in coal mines and poor image quality caused by all-weather artificial lighting, on the basis of analyzing and comparing traditional image enhancement processing algorithms, a new method is proposed. An image enhancement preprocessing method based on fuzzy logic, which adopts a new membership function, which can not only reduce the pixel information loss in the low gray area of the image, overcome the problem of contrast decrease caused by fuzzy enhancement, and improve the adaptability . At the same time, a fast maximum inter-class variance method is used for threshold selection, which realizes adaptive and rapid selection of fuzzy enhancement thresholds, improves the algorithm operation speed, enhances real-time performance, and can enhance images in different environments, and It can effectively improve the detailed information of the image, improve the image quality, and the calculation speed is fast to meet the real-time requirements

3. The image segmentation method adopted has simple steps, reasonable design and good segmentation effect. Since the segmentation effect of the one-dimensional maximum entropy method is not ideal for images with low signal-to-noise ratio and low illumination, the maximum segmentation method based on two-dimensional fuzzy division is adopted. The entropy segmentation method is used for segmentation. This segmentation method considers the characteristics of gray level information, spatial neighborhood information and its own fuzziness, but has the disadvantage of slow operation speed. The optimization makes it easy, fast and accurate to obtain the optimized fuzzy parameter combination, thus greatly improving the efficiency of image segmentation. Moreover, the particle swarm optimization algorithm adopted is reasonably designed and easy to implement. It adaptively adjusts the size of the local space according to the current state of the particle swarm and the number of iterations, and obtains a higher search success rate and a higher search rate without affecting the convergence speed. Higher-quality solution, better segmentation effect, strong robustness, and improved computing speed to meet real-time requirements.

4. Because the segmentation method based on the maximum entropy of two-dimensional fuzzy partition can quickly and accurately segment the flame image, it overcomes the problem that the traditional algorithm uses a single threshold noise point to be misclassified. Optimization, which solves the nonlinear integer programming problem, makes the segmented target better maintain its shape while overcoming the influence of noise. Therefore, the present invention combines the segmentation method based on the maximum entropy of two-dimensional fuzzy division with the particle swarm optimization algorithm to realize the rapid segmentation of infrared images, and sets the parameter combination (a, b, c, d) as particles, and the two-dimensional fuzzy division entropy as The fitness function determines the search direction of the particle in the solution space. Once the two-dimensional histogram of the image is obtained, the PSO algorithm is used to search for the optimal parameter combination (a, b, c, d) that maximizes the fitness function, and finally according to the maximum membership The degree principle classifies the pixels in the image, so as to realize the segmentation of the image. Moreover, the segmentation method of the present invention has a very good segmentation effect on infrared images with large noise, low contrast and small targets.

5. In actual feature extraction, area, similarity, moment characteristics, density, texture features, and stroboscopic characteristics are selected as the feature basis for fire image recognition, which not only retains the features that contribute greatly to the classification, but discards redundant features. The feature dimension is reduced, and the optimal selection of features is completed.

6. The modeling method of the binary classification model adopted is simple. The design is reasonable and easy to implement, and the use effect is good, and the hyperparameters of the kernel function are optimized by using the conjugate gradient method. Based on the characteristics of artificial neural network suitable for dealing with imperfect and fuzzy information and the advantages of small samples, nonlinear and high-dimensional models of support vector machine, the fire point identification is carried out separately to achieve the purpose of complementary advantages of each criterion, thus overcoming the traditional The disadvantage of using a single criterion to judge fire hazards is likely to cause false alarms. The commonly used cross-validation optimization parameter method is quite time-consuming, and it does not guarantee that the selected parameters will ensure that the classifier has the best classification performance. However, other existing hyperparameter selection algorithms cannot select the penalty factor and kernel function parameters at the same time. Deficiency, for the small sample LS-SVM model classification problem, the present invention adopts to minimize the sum of the prediction error squares as the goal, and uses the method of gradient descent to simultaneously select two super Parameters Kernel function parameters and penalty factors. The binary classification model established by the present invention not only has a high recognition rate and high classification accuracy, but also takes a short time and can easily and quickly complete the fire recognition process. When it is recognized that the category of the currently collected image is flame, it means that a fire has occurred. Make an alarm prompt in time and take corresponding measures. Aiming at the small-sample and non-linear problems of fire recognition in the complex and special environment of coal mines, and the advantages of support vector machines in high dimensions, the present invention proposes a fire image recognition method based on least squares support vector machines, and quickly stays one On the basis of the method, the hyperparameters are optimized using the conjugate gradient method, and the FR-LSSVM model is constructed.

In summary, the method of the present invention has simple steps, convenient implementation, easy operation, high reliability, and good application effect, and can effectively solve the problem of low reliability and high false alarm and false negative rate in the complex environment existing in the existing video fire detection system. High, poor use effect and other issues.

The technical solutions of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments.

Description of drawings

Fig. 1 is a flow chart of the method of the present invention.

Fig. 2 is a schematic circuit block diagram of the image acquisition and processing system used in the present invention.

FIG. 3 is a schematic structural diagram of a two-dimensional histogram established in the present invention.

Fig. 4 is a schematic diagram of the segmentation state when performing image segmentation in the present invention.

Explanation of reference signs:

1—CCD camera; 2—video capture card; 3—processor; 4—data memory.

Detailed ways

A kind of image type fire flame identification method as shown in Figure 1, comprises the following steps:

Step 1. Image acquisition: the image acquisition unit is used to acquire the digital image of the area to be detected according to the preset sampling frequency f _s , and the digital image acquired at each sampling moment is synchronously transmitted to the processor 3 . The image acquisition unit is connected to the processor 3 .

In this embodiment, the image acquisition unit includes a CCD camera 1 and a video acquisition card 2 connected with the CCD camera 1, the CCD camera 1 is connected with the video acquisition card 2, and the video acquisition card 2 is connected with the processor 3 connect.

In this embodiment, the size of the digital image collected at each sampling moment is M1×N1 pixels. Where M1 is the number of pixels on each row in the collected digital image, and N1 is the number of pixels on each column in the collected digital image.

Step 2, image processing: the processor 3 performs image processing on the digital images collected at each sampling time in step 1 according to the time sequence, and the processing methods for the digital images collected at each sampling time are the same; for step 1 When processing the digital images collected at any sampling moment, the following steps are included:

Step 201, image preprocessing, the process is as follows:

Step 2011, image reception and synchronous storage: the processor 3 synchronously stores the received digital image collected at the current sampling moment in the data memory 4, and the data memory 4 is connected to the processor 3;

In this embodiment, the CCD camera 1 is an infrared CCD camera, and the CCD camera 1, video capture card 2, processor 3 and data storage 4 form an image acquisition and preprocessing system, see FIG. 2 for details.

Step 2012, image enhancement: the digital image collected at the current sampling moment is enhanced by the processor 3 to obtain an enhanced digital image.

Step 2013, image segmentation: the processor 3 performs segmentation processing on the enhanced digital image in step 2012 to obtain a target image.

Step 202, fire identification: use the pre-established binary classification model to process the target image described in step 2013, and obtain the fire state category of the area to be detected at the current sampling moment; the fire state category includes flame and no flame Two categories, the binary classification model is a support vector machine model for classifying two categories with flame and without flame.

The establishment process of the two classification models is as follows:

Step 1. Image information collection: using the image collection unit to separately collect multiple frames of digital image 1 of the area to be detected when a fire occurs and multiple frames of digital image 2 of the area to be detected when there is no fire.

Step II, feature extraction: perform feature extraction on the multi-frame digital image one and the multi-frame digital image respectively, and extract a group of characteristic parameters that can represent and distinguish the digital image from each digital image, and the The group feature parameters include M feature quantities, and the M feature quantities are numbered, and the M feature quantities form a feature vector, where M≥2.

Step Ⅲ, training sample acquisition: from the feature vectors of multiple frames of the digital image 1 and multiple frames of the digital image 2 obtained after the feature extraction in step Ⅱ, select the features of the digital image 1 in the m1 frame respectively The feature vector of the digital image two in the vector and the m2 frame forms a training sample set; wherein, m1 and m2 are positive integers and m1=40～100, m2=40～100; the number of training samples in the training sample set is m1 +m2 pieces.

In this embodiment, when obtaining the training samples, the image acquisition unit is used to collect a digital image sequence 1 when a fire occurs within a period of time t1 and a digital image sequence 2 of a region to be detected when no fire occurs; the digital image sequence The number of frames of digital images included in one is n1=t1×f, where t1 is the sampling time of said digital image sequence one; the number of frames of digital images included in said digital image sequence two is n2=t2×f, Wherein t2 is the sampling time of the second digital image sequence. Wherein, n1 is not less than m1, and n2 is not less than m2. Afterwards, m1 digital images are selected from the digital image sequence one as flame samples, and m2 digital images are selected from the digital image sequence two as non-flame samples.

In this embodiment, m1=m2.

Step Ⅳ, two-category model establishment, the process is as follows:

Step IV-1, kernel function selection: select radial basis function as the kernel function of the binary classification model;

Step IV-2, determination of classification function: After the factor γ to be penalized and the kernel parameter ^σ2 of the radial basis function selected in step IV-1 are determined, the classification function of the binary classification model is obtained, and the binary classification is completed Model establishment process; where, γ=C ^-2 , σ=D ^-1 , 0.01<C≤10, 0.01<D≤50.

When determining the penalty factor γ and the kernel parameter σ ² , the parameters C and D are optimized using the conjugate gradient method first, and the optimized parameters C and D are obtained, and then according to γ=C ^-2 and σ=D ^-1 The optimized parameters C and D are transformed into penalty factor γ and kernel parameter σ ² .

When the conjugate gradient method is used to optimize parameters C and D in step IV-2, the m1+m2 training samples in the training sample set described in step III are used for optimization.

Step V, binary classification model training: input the m1+m2 training samples in the training sample set mentioned in step III to the binary classification model established in step IV for training.

In this embodiment, the total number of training samples in the training sample set in step III is N and N=m1+m2; before the binary classification model is established in step IV, number the N training samples in the training sample set , the number of the pth training sample in the training sample set is p, p is a positive integer and p=1, 2, ..., N; the pth training sample is recorded as (x _p , y _p ), where x _p is The feature parameter of the pth training sample (ie, the feature vector), _yp is the category number of the pth training sample and _yp =1 or -1, where the category number is 1 means there is flame, and the category number is -1 Indicates no flame.

When using the conjugate gradient method to optimize parameters C and D in step IV-2, use the m1+m2 training samples in the training sample set mentioned in step III to optimize, and the optimization process is as follows:

为矩阵

的第p列除去第p个元素后的其余元素组成的列向量，矩阵

为矩阵K的增广矩阵，其中矩阵 $K=\left|\begin{array}{cccc}K(x_1,x_1)& K(x_1,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_1,x_N)\\ K(x_2,x_1)& K(x_2,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_2,x_N)\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ K(x_N,x_1)& K(x_N,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_N,x_N)\end{array}\right|;$ 矩阵A^-1表示矩阵A的逆矩阵，矩阵 $A=\left[\begin{array}{cc}K+C^2\&CenterDot;I& I_N\\ {I_N}^T& 0\end{array}\right],$ 其中矩阵I为单位矩阵，矩阵I_N=[1,1,…,1]^T，T表示矩阵的转置，矩阵I_N中包含N个元素且N个元素均为1；矩阵s=A^-1Y，矩阵 $Y=\left[\begin{array}{c}{y}_1\\ y_2\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ y_N\\ 0\end{array}\right],$ 其中y₁、y₂、…、y_N分别为所述训练样本集中N个训练样本的类别。Step 1, objective function determination: $\mathrm{sse}(C,\mathrm{D.})=\frac{1}{2}\underset{p=1}{\overset{\mathrm{<figure-callout\; id="2"\; label="N\; e\; p"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">N</figure-callout>}}{\mathrm{\&Sigma;}}}{e}_p^{}{}_2^{}=\frac{1}{2}\underset{p=1}{\overset{N}{\mathrm{\&Sigma;}}}{[\tilde{K}(p,p^{-})\&Center\; Dot;{\mathrm{the\; s}}_p-\mathrm{the\; y}\left(p\right)]}^2$ (1), where sse(C, D) is the sum of the squares of the leave-one-out prediction error, p is the serial number of each training sample in the training sample set, and e _p is the p-th training result of the binary classification model established in step IV. The sample prediction error and $e_p=\tilde{K}(p,p^{-})\&Center\; Dot;{\mathrm{the\; s}}_p-\mathrm{the\; y}\left(p\right);$ in, ${\mathrm{the\; s}}_p=\mathrm{the\; s}\left(p^{-}\right)-\frac{\mathrm{the\; s}\left(p\right)}{\left(A^{-1}\right)(p,p)}\left(A^{-1}\right)(p^{-},p),$ In the formula, s(p ^- ) is a vector composed of the remaining elements after removing the p-th element in the matrix s; s(p) is the p-th element of the matrix s, and (A ^-1 )(p ^- ,p) is the matrix A column vector composed of the remaining elements after the p-th column of A ^-1 except the p-th element, (A ^-1 )(p,p) is the p-th element of the p-th column of matrix A ^-1 ;

for the matrix

A column vector composed of the remaining elements after removing the p-th column of the p-th column, matrix

is the augmented matrix of matrix K, where the matrix $K=\left|\begin{array}{cccc}K(x_1,x_1)& K(x_1,x_2)& \&CenterDot;\&CenterDot;\&Center\; Dot;& K(x_1,x_N)\\ K(x_2,x_1)& K(x_2,x_2)& \&CenterDot;\&Center\; Dot;\&Center\; Dot;& K(x_2,x_N)\\ \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;\\ \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;\\ \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;\\ K(x_N,x_1)& K(x_N,x_2)& \&Center\; Dot;\&Center\; Dot;\&Center\; Dot;& K(x_N,x_N)\end{array}\right|;$ Matrix A ^-1 represents the inverse matrix of matrix A, matrix $A=\left[\begin{array}{cc}K+C^2\&Center\; Dot;I& I_N\\ {I_N}^T& 0\end{array}\right],$ Among them, the matrix I is the identity matrix, the matrix I _N =[1,1,…,1] ^T , T represents the transposition of the matrix, the matrix I _N contains N elements and all N elements are 1; the matrix s=A ^{- 1} Y, matrix $Y=\left[\begin{array}{c}{\mathrm{the\; y}}_1\\ {\mathrm{the\; y}}_2\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ {\mathrm{the\; y}}_N\\ 0\end{array}\right],$ Wherein y ₁ , y ₂ , . . . , y _N are respectively the categories of the N training samples in the training sample set.

为矩阵K的右侧增加一列元素且各元素均为1后的矩阵。Among them, the matrix $\tilde{K}=[K,1]=\left|\begin{array}{ccccc}K(x_1,x_1)& K(x_1,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_1,x_N)& 1\\ K(x_2,x_1)& K(x_2,x_2)& \&Center\; Dot;\&Center\; Dot;\&Center\; Dot;& K(x_2,x_N)& 1\\ \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;\\ \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;\\ \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;& \&Center\; Dot;\\ K(x_N,x_1)& K(x_N,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_N,x_N)& 1\end{array}\right|.$ Among them, the matrix K is the kernel function matrix. matrix

Add a column of elements to the right side of the matrix K and each element is 1.

Since the constraints of the least squares support vector machine (LS-SVM) are expressed in the following form:

(5.22)，式中w^T·φ(x_p)+b为高维特征空间中的分类超平面，w和b为分类超平面的参数；e_p为第p个训练样本的训练误差，

为经验风险；w^T·w=||w||²衡量了学习机器的复杂性。

(5.22), where w ^T φ(x _p )+b is the classification hyperplane in the high-dimensional feature space, w and b are the parameters of the classification hyperplane; e _p is the training error of the pth training sample,

is the empirical risk; w ^T ·w=||w|| ² measures the complexity of the learning machine.

After the training sample set is determined, the performance of the LS-SVM model depends on the type of its kernel function and the selection of two hyperparameters, the two hyperparameters are the penalty factor γ and the kernel parameter σ ² , the LS-SVM model The classification accuracy is related to the selection of hyperparameters. The kernel parameter σ ² represents the width of the radial basis function and has a great relationship with the smooth performance of the LS-SVM model; the penalty factor γ is also called a regularization parameter. Controls the degree of punishment for wrong samples, which is closely related to the complexity of the LS-SVM model and the matching degree of training samples.

In this embodiment, the radial basis function selected in step IV-1 is $K(x_{\mathrm{the\; s}},x_t)=\exp [-\frac{{|x_{\mathrm{the\; s}}-x_t|}^2}{{\mathrm{\&sigma;}}^2}],$ The regression function of the radial basis function is $\mathrm{the\; y}=\underset{t=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{t}K(x,x_t)+b,$ In the formula, α _t and b are regression parameters, s is a positive integer and s=1, 2,..., N, t is a positive integer and t=1, 2,..., N.

Equation (5.22) can be written as: $\left\{\begin{array}{c}\min J(w,b,e)=w^{T}w+\frac{1}{2{\mathrm{<figure-callout\; id="2"\; label="C"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">C</figure-callout>}}^2}\underset{t=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{\mathrm{<figure-callout\; id="2"\; label="e\; t"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">e</figure-callout>}}_t^{}\\ \mathrm{the\; s}.t.{\mathrm{the\; y}}_t=w^T\mathrm{\&phi;}\left(x_t\right)+b+e_t,t=\mathrm{1,2},...,N\end{array}\right.---\left(5.23\right);$

In the formula (5.23), C ^-2 replaces the penalty factor γ, but it can also play a role in balancing the complexity of the LS-SVM model and the empirical risk; σ is replaced by D ^-1 , and the radial basis function $K(x_{\mathrm{the\; s}},x_t)=\exp [-\frac{{|x_{\mathrm{the\; s}}-x_t|}^2}{{\mathrm{\&sigma;}}^2}]$ It is expressed by the following formula: K(x _s ,x _t )=exp(-D ² ·||x _s -x _t || ² ).

According to the principle of least squares support vector machine, the formula (5.23) is transformed into a linear equation system $\left[\begin{array}{cc}K+{\mathrm{<figure-callout\; id="2"\; label="C"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">C</figure-callout>}}^2\&Center\; Dot;I& I_N\\ {{\mathrm{<figure-callout\; id="0"\; label="I\; N\; T"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">I</figure-callout>}}_N}^T& 0\end{array}\right]\&CenterDot;\mathrm{the\; s}=Y=\left[\begin{array}{c}{\mathrm{the\; <figure-callout\; id="1"\; label="y"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">y</figure-callout>}}_1\\ {\mathrm{the\; <figure-callout\; id="2"\; label="y"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">y</figure-callout>}}_2\\ \&Center\; Dot;\\ \&Center\; Dot;\\ \&Center\; Dot;\\ {\mathrm{the\; y}}_N\\ 0\end{array}\right]$ (5.25), the derivation process of formula (5.25) refers to "Online Recursive Least Squares Support Vector Based on Fast Leave-One-Out Cross-Validation Method" published in October 2012 in the Journal of Qingdao University of Science and Technology (Natural Science Edition), Volume 33, Issue 5 Machine Modeling Method" (authored by Shao Weiming, Tian Xuemin).

由公式(5.25)可以得出：矩阵s=A^-1Y(5.28)。Solving the formula (5.25), the regression function of the radial basis function can be obtained

From formula (5.25), it can be drawn: matrix s=A ^-1 Y(5.28).

The established binary classification model is verified N times through the N training samples in the training sample set. When performing the pth verification, the pth training sample is used as the prediction set, and the remaining N-1 samples are used as the training set. Set, after solving the LS-SVM parameters a _p and b through the training set, classify the pth training sample as the prediction set, and record whether the classification result is correct or not; in this way, after N times of verification, you can calculate the leave-one-out The predicted misclassification rate e _LOO is calculated as (5.29). For each set of given hyperparameters (including C and D), the corresponding _eLOO can be calculated, so that the hyperparameter combination with the smallest _eLOO can be selected as the optimized parameter.

because ${\mathrm{the\; s}}_p=\mathrm{the\; s}\left(p^{-}\right)-\frac{\mathrm{the\; s}\left(p\right)}{\left(A^{-1}\right)(p,p)}\left(A^{-1}\right)(p^{-},p)$ (5.30), so for each set of given hyperparameters, when performing a leave-one-out cross-validation, you only need to solve A ^-1 once, and then calculate _sp in each iteration, thus saving a lot of cross-validation time, calculating volume is greatly reduced.

In order to minimize sse(C,D), the formula ${\mathrm{the\; s}}_p=\mathrm{the\; s}\left(p^{-}\right)-\frac{\mathrm{the\; s}\left(p\right)}{\left(A^{-1}\right)(p,p)}\left(A^{-1}\right)(p^{-},p)$ To optimize to search for C and D, first, the gradient of sse(C,D) to C and D, according to the definition of matrix derivation and inverse matrix derivation, it is obtained: $\frac{\&PartialD;k(i,j)}{\&PartialD;\mathrm{\&sigma;}}=-2\mathrm{D.}\&Center\; Dot;K(i,j)\&CenterDot;{\left|\right|x_i-x_j\left|\right|}^2$ (5.32); $\frac{\&PartialD;\tilde{k}}{\&PartialD;\mathrm{D.}}=[\frac{\&PartialD;k}{\&PartialD;\mathrm{D.}},0]$ (5.33); $\frac{\&PartialD;A}{\&PartialD;C}=\left[\begin{array}{cc}2C\&Center\; Dot;I& 0\\ 0^T& 0\end{array}\right]$ (5.34) and $\frac{\&PartialD;A}{\&PartialD;\mathrm{D.}}=\left[\begin{array}{cc}\frac{\&PartialD;k}{\&PartialD;\mathrm{D.}}& 0\\ 0^T& 0\end{array}\right]$ (5.35), matrix 0 in formula (5.35) represents an N-dimensional column vector whose elements are all 0.

(5.36)；According to A·A ^-1 =I (I is the identity matrix), it can be derived:

(5.36);

According to the formula $\mathrm{sse}(C,\mathrm{D.})=\frac{1}{2}\underset{p=1}{\overset{\mathrm{<figure-callout\; id="2"\; label="N\; e\; p"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">N</figure-callout>}}{\mathrm{\&Sigma;}}}{e}_p^{}{}_2^{}=\frac{1}{2}\underset{p=1}{\overset{N}{\mathrm{\&Sigma;}}}{[\tilde{K}(p,p^{-})\&CenterDot;{\mathrm{the\; s}}_p-\mathrm{the\; y}\left(p\right)]}^2,$ Two partial derivatives can be derived as follows: $\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C}=\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C}=\underset{p=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{e}_p\&Center\; Dot;\tilde{K}(p,p^{-})\&CenterDot;\frac{\&PartialD;{\mathrm{the\; s}}_p}{\&PartialD;C}---\left(5.37\right),$

$\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; sse</span>}}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; e</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; [</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\stackrel{<span\; class="notranslate">\; ~</span>}{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5.38</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

According to formula (5.30),

$\frac{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\frac{<span\; class="notranslate">\; \&PartialD;</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; C</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{{<span\; class="notranslate">\; [</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&CenterDot;</span>\frac{<span\; class="notranslate">\; \&PartialD;</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5.39</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

$\frac{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\frac{<span\; class="notranslate">\; \&PartialD;</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{{<span\; class="notranslate">\; [</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&Center\; Dot;</span>\frac{<span\; class="notranslate">\; \&PartialD;</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5.40</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

而可由式(5.32)-(5.35)计算得出。obviously,

and It can be calculated by formula (5.32)-(5.35).

For each set of hyperparameters C and D, the gradient of sse(C,D) to them can be calculated according to formulas (5.37) and (5.38). According to the principle of LS-SVM, the selection of hyperparameters of LS-SVM is converted from a constrained optimization problem to an unconstrained optimization problem, replacing γ with C ^-2 , and replacing σ with D ^-1 . This transformation will not affect the LS-SVM model On the other hand, the values of C and D do not affect the calculation of the gradient.

Step ii. Initial parameter setting: determine the initial values C ₁ and D ₁ of parameters C and D respectively, and set the recognition error threshold ε and ε>0.

计算得出步骤ⅰ中目标函数对C_k和D_k的梯度g_k，k为迭代次数且k=1、2、…；若||g_k||≤ε，停止计算，此时C_k和D_k分别为优化后的参数C与D；否则，进入步骤ⅳ。Step Ⅲ, the gradient g _k calculation of the current iteration: according to the formula

Calculate the gradient g _k of the objective function to C _k and D _k in step i, k is the number of iterations and k=1, 2, ...; if ||g _k ||≤ε, stop the calculation, at this time C _k and D _k are the optimized parameters C and D respectively; otherwise, go to step ⅳ.

in, $\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k}=\underset{p=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_p\&Center\; Dot;\tilde{K}(p,p^{-})\&Center\; Dot;\frac{\&PartialD;{\mathrm{the\; s}}_p}{\&PartialD;C_k};$

为矩阵

的第p列除去第p个元素后的其余元素组成的列向量；s(p^-)为矩阵s中除去第p个元素后的其余元素组成的向量，e_p为步骤Ⅳ中所建立的二分类模型对第p个训练样本的预测误差。 $\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k}=\underset{p=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{e}_p\&Center\; Dot;[\frac{\&PartialD;\tilde{K}(p,p^{-})}{\&PartialD;{\mathrm{D.}}_k}\&Center\; Dot;\mathrm{the\; s}\left(p^{-}\right)+\tilde{K}(p,p^{-})\&CenterDot;\frac{\&PartialD;{\mathrm{the\; s}}_p}{\&PartialD;{\mathrm{D.}}_k}],$ In the formula

for the matrix

s(p ^- ) is a vector composed of the remaining elements after removing the p-th element in the matrix s, and e _p is the two elements established in step IV The prediction error of the classification model for the pth training sample.

Step ⅳ. Calculation of the search direction d _k of the current iteration: according to the formula $d_k=\left\{\begin{array}{cc}-g_k& k=1\\ -g_k+{\mathrm{\&beta;}}_k{d}_{k-1}& k\&Greater\; Equal;2\end{array}\right.,$ Calculate the search direction d _k of the current iteration, where d _k-1 is the search direction of the k-1th iteration, β _k =||g _k || ² /||g _k-1 || ² , g _{k- 1} is the gradient of the k-1th iteration.

Step ⅳ, determine the search step size λ _k of the current iteration: search along the search direction d _k determined in step ⅳ, and find out that satisfies the formula

表示在（0，+∞）中寻找使达到最小值的步长λ_k。 $\mathrm{sse}(C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k},{\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k})=\underset{{\mathrm{\&lambda;}}_k>0}{\min }\mathrm{sse}(C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k},{\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k})$ The search step size λ _k , where

Indicates that in (0, +∞) to find the The step size λ _k to reach the minimum.

Step ⅵ, according to the formula $C_{k+1}=C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k}$ and ${\mathrm{D.}}_{k+1}={\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k},$ Calculations are performed for C _k+1 and D _k+1 .

Step ⅶ, set k=k+1, and then return to step Ⅲ for the next iteration.

Finally, the matrix $\mathrm{the\; s}=A^{-1}Y=\left[\begin{array}{c}{\mathrm{\&alpha;}}_1\\ {\mathrm{\&alpha;}}_2\\ \&Center\; Dot;\\ \&CenterDot;\\ \&Center\; Dot;\\ {\mathrm{\&alpha;}}_{\mathrm{<figure-callout\; id="0"\; label="N"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">N</figure-callout>}}\\ 0\end{array}\right].$

In this embodiment, after the binary classification model is established, the final classifier used is: $f\left(x\right)=\mathrm{sgn}[\underset{t=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t{\mathrm{the\; y}}_{t}K(x\&Center\; Dot;x_t)+b].$

T表示矩阵的转置，H为自相关矩阵且H=A^T·A。In this embodiment, in step v

T represents the transposition of the matrix, H is the autocorrelation matrix and H=A ^T ·A.

In the actual operation process, when determining C ₁ and D ₁ in step ii, use the grid search method or the method of randomly extracting values to determine C ₁ and D ₁ ; use the method of randomly extracting values to determine C ₁ and D ₁ When determining, C ₁ is a value randomly selected in (0.01, 1], D ₁ is a value randomly selected in (0.01, 50]; when using the grid search method to determine C ₁ and D ₁ , first 10 ^-3 is the step size to divide the grid, and then use C and D as independent variables and the objective function described in step i as the dependent variable to make a three-dimensional grid map, and then find out multiple sets of parameters of C and D through grid search , and finally take the average value of multiple groups of parameters as C ₁ and D ₁ .

In this embodiment, the grid search method is used to determine C ₁ and D ₁ , and the group B parameters of C and D are found through the grid search, where B is a positive integer and B=5-20.

The conjugate gradient method has the characteristics of simple algorithm, small storage required, and fast convergence speed. It transforms the multidimensional problem into a series of one-dimensional search methods along the negative gradient direction. It can effectively reduce the number of iterations and reduce the running time.

When the grid search method is used to optimize C and D, only when the step size is set to a small value can the optimal hyperparameter value with high precision be found, which will be very time-consuming.

其中，AR表示相邻帧间高亮区域的面积变化率，A(n)和A(n+1)分别表示当前帧和下一帧中可疑区域的面积。为防止相邻两帧图像中都不存在可疑火焰区域而使得计算出的面积变化率成为无穷大，在分母上加一个极小值eps。另外为了实现归一化，取两帧中高亮区域面积的最大值作为上式的分母，这样可使最终计算出的结果介于(0,1)之间。Combustion is a continuous and typically unstable physical process with various characterization parameters. The flame image in the early stage of the fire mainly has the characteristics of enlarged flame area, edge jitter, irregular shape, and basically stable position. Among them, the criterion calculations used for the area growth characteristics are all developed and implemented under the Visual C++ ^[115] platform, where the area change rate is defined as:

Among them, AR represents the area change rate of highlighted regions between adjacent frames, and A(n) and A(n+1) represent the areas of suspicious regions in the current frame and the next frame, respectively. In order to prevent the calculated area change rate from becoming infinite because there is no suspicious flame area in two adjacent frames of images, a minimum value eps is added to the denominator. In addition, in order to achieve normalization, the maximum value of the highlighted area in the two frames is taken as the denominator of the above formula, so that the final calculated result can be between (0,1).

求得若干个相似度后，使用连续几帧图像的相似度ξ_h的平均值

作为判据。The shape similarity of images is usually carried out by means of similarity with known descriptors, and this method can establish corresponding similarity measures at any degree of complexity. According to the background difference method, let the known image sequence be f _h (x, y), h=1,2,..., N ₀ , where (x, y) is the coordinate of each pixel in the image, and N ₀ is the number of frames , let the reference image be f _o (x,y), so that a difference image sequence can be defined as: δ _h (x,y)=|f _h (x,y)-f _o (x,y)|, this The difference image sequence represents the difference between each frame of the original image sequence and the reference image. Next, binarize the obtained difference image sequence to obtain an image sequence {b _h (x, y)}. Pixels marked as 1 in this image sequence represent areas where there is a significant difference between the original image sequence and the reference image. We consider this area to be the possible flame area, and mark the pixels with 1 in each frame in the image sequence after filtering out the influence of isolated points, and obtain the possible flame area Ω _h in each frame in the sequence image. When the suspicious flame area is found, the flame and the disturbance are classified by calculating the similarity of the continuous frame change images. The similarity ξ _h of continuous frame changing images is defined as:

After obtaining several similarities, use the average value of the similarity ξ _h of several consecutive frames of images

as a criterion.

和

方向的一阶矩(M₁₀,M₀₁)，再计算得到其质心。Moment characteristics, from the perspective of flame recognition, use the center of mass feature of the flame image, and use the center of mass to represent its stability. For a flame image, first calculate its centroid, such as the formula: M ₀₀ is the zero-order moment of the target area, that is, the area of the target area. Computational image

and

The first moment of the direction (M ₁₀ , M ₀₁ ), and then calculate its center of mass.

The edge change of the early fire flame has its own unique rules, and the simple and practical characteristic parameters such as density and eccentricity are used as one of the fire criteria to identify the edge change of the flame.

Density is usually used to describe the complexity of object boundaries, also known as circularity or dispersion, which is a feature quantity that calculates the shape complexity of objects or regions on the basis of area and perimeter. It is defined as follows: k=1, 2,...n1, in the formula (4.7), C _k represents the compactness of the primitive numbered k, and P _k is the perimeter of the kth primitive, that is, the boundary length of the suspicious primitive, which can be obtained by Calculate the boundary chain code to get. A _k is the area of the k-th graphic element. For grayscale images, it can be obtained by calculating the number of bright spots of suspicious graphic elements. For binary images, it can be obtained by calculating the number of pixels with a pixel value of 1. n is the suspicious flame in the image The number of primitives. The calculation of the perimeter is relatively complicated, but it can be determined by extracting the boundary chain code.

The steps to calculate the density of the flame area are as follows:

① Calculate the area of the suspected flame area on the basis of image segmentation;

②Detect continuous perimeter pixels in the vertical direction, and record the number Nx of continuous perimeter pixels; detect continuous boundary pixels in the horizontal direction, and record the number N _y of continuous boundary pixels, and calculate the total number of boundary pixels S _N ;

③The number of chain codes of even numbers is N _E =N _x +N _y , the number of chain codes of odd numbers N ₀ =S _N —N _E , using the perimeter formula $P=N_{\mathrm{E.}}+\sqrt{2}{\mathrm{<figure-callout\; id="0"\; label="N"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">N</figure-callout>}}_0$ Calculate the circumference.

④ Substitute the results of ① and ③ into formula (4.7) to calculate the density.

Using the gray-level co-occurrence matrix to extract image texture features, Haralick et al. extracted 14 features from the gray-level co-occurrence matrix. In this embodiment, the extracted image texture features include five features of contrast, entropy, energy, uniformity and correlation.

The burning flame has a flickering phenomenon, which reflects the distribution of pixels of a frame of image at different gray levels over time. By calculating the changes of the edge pixels, the flickering law of the target pattern can be obtained. The frequency of flame flickering is generally in the low frequency region of 10-20Hz. Since the general frequency of video image acquisition is 25Hz (2 frames/s), it cannot meet the sampling requirements for obtaining flicker frequency without distortion, so it is difficult to obtain its characteristic spectrum directly through video information. For the flickering law, Toreyin proposed to use wavelet to analyze the R component of this point under the RGB model by using the color value of a fixed pixel point in each frame to change with time. If there is a flame, it will cause a drastic change in the value at this point. Wavelet The high frequency components of the decomposition will be non-zero. Wang Zhenhua ^[73] proposed discrete wavelet transform to decompose and reconstruct the time series of flame characteristics, and use the change of area to express its scintillation law. Zhang Jinhua et al. pointed out that the height of the flame will change greatly when the flame flickers, and there is a direct relationship between the change law and the flicker frequency, and there is a big difference between the interference source, so it is proposed to use the change of the flame height instead of the flame flicker feature to Methods for performing flame identification. In this embodiment, the stroboscopic characteristics are extracted by using the method pointed out by Zhang Jinhua et al. that the height of the flame will change greatly when it flickers.

In this embodiment, M=6 in step II, and the 6 feature quantities are area, similarity, moment characteristic, density, texture characteristic and stroboscopic characteristic respectively.

In the actual operation process, in order to test the performance of the established binary classification model, a total of 81 training samples including flame samples and non-flame samples were selected, and each sample was 7-dimensional. For 81 training samples, a set of data is taken out for prediction and classification each time, and the remaining 80 sets of data are used to optimize the selection of hyperparameters. Among them, the initial value C ₁ =1 and D ₁ =1, using the conjugate gradient method to search, the mean value of C obtained is 0.1386, the mean square error is 0.0286, the mean value of D is 0.2421, the mean square error is 0.0273, it can be seen that the optimal hyperparameter comparison Stablize. Compare the binary classification model (FR-LSSVM model) used in the present invention with BP (neural network model), LS-SVM (least squares support vector machine model) and standard SVM (support vector machine model) three classification models , the recognition results are shown in Table 1:

Table 1 Comparison table of recognition results of different classification models

It can be seen from Table 1 that in terms of recognition rate, BP is the worst, and LS-SVM with initial values selected only by grid search is also poor, while FR-LSSVM and standard SVM are significantly better than them. In terms of training time, FR-LSSVM and LS-SVM are significantly superior, standard SVM is slightly superior, and standard SVM is difficult to find the optimal hyperparameters, BP neural network training is time-consuming, and the recognition rate is slightly low. The reason is that the number of training samples is small, the number of samples has a great impact on the recognition rate, and the amount of feature information is insufficient, and the neural network has insufficient convergence and local minimization. The selection of each parameter depends on experience, and there is a large gap in the parameter setting. Uncertainty, the recognition rate can be improved by supplementing the training samples and further modifying the weights of the BP neural network. The recognition rate of standard SVM is higher than that of LS-SVM, but the training time and recognition time are longer. The FR-LSSVM hyperparameter algorithm is more standardized, less time-consuming, more stable, and reduces uncertainty. It is especially suitable for modeling small samples and nonlinear problems, and has significant advantages in speed and accuracy. In addition, these algorithms have high requirements on image quality. If the image pixels are low, the target area in the image is blocked by obstacles in a large area or covered or surrounded by dust, the extracted target is incomplete or there is noise in the extracted target, etc., which may lead to poor recognition rate. reduce.

In this embodiment, when image enhancement is performed in step 2012, an image enhancement method based on fuzzy logic is used for enhancement processing.

In the actual enhancement process, when the image enhancement method based on fuzzy logic (specifically, the classic Pal-King fuzzy enhancement algorithm, namely the Pal algorithm) is used for image enhancement processing, there are the following defects:

①Pal algorithm uses complex power function as fuzzy membership function when performing fuzzy transformation and its inverse transformation, which has the defects of poor real-time performance and large amount of computation;

②In the process of fuzzy enhancement transformation, quite a lot of low grayscale values in the original image are hard set to zero, resulting in the loss of low grayscale information;

③The selection of the fuzzy enhancement threshold (transition point X _c ) is generally obtained by experience or multiple comparison attempts, which lacks theoretical guidance and is arbitrary; the parameters F _d and F _e in the membership function are adjustable, and the parameter values F _d , The reasonable selection of F _e is closely related to the effect of image processing;

④In the process of fuzzy enhancement transformation, multiple iterative operations are used to repeatedly enhance the image, and the selection of the number of iterations is not guided by relevant theoretical principles, and the edge details will be affected when the number of iterations is large.

In order to overcome the above-mentioned defects in the classic Pal-King fuzzy enhancement algorithm, in the present embodiment, when the digital image is enhanced in step 2012, the image to be enhanced, the process is as follows:

Step 20121, transform from the image domain to the fuzzy domain: according to the membership function ${\mathrm{\&mu;}}_{\mathrm{gh}}=T\left(x_{\mathrm{gh}}\right)=\left\{\begin{array}{cc}{x}_{\mathrm{gh}}/x_T& x_{\mathrm{gh}}\&le;x_T\\ x_{\mathrm{gh}}/x_{\max }& x_{\mathrm{gh}}>x_T\end{array}\right.$ (7), the gray value of each pixel of the image to be enhanced is mapped to the fuzzy membership degree of the fuzzy set, and the fuzzy set of the image to be enhanced is obtained accordingly; x _gh is in the image to be enhanced The gray value of any pixel point (g, h), X _T is the gray threshold selected when the image to be enhanced is enhanced using the image enhancement method based on fuzzy logic, and X _max is the image to be enhanced The maximum gray value of .

After the gray value of each pixel of the image to be enhanced is mapped to the fuzzy membership of the fuzzy set, the corresponding fuzzy membership mapped to the gray value of all the pixels of the image to be enhanced forms the fuzzy membership of the fuzzy set matrix.

Since μ _gh ∈ [0,1] in formula (7), it overcomes the defect that many low gray values of the original image are cut to zero after fuzzy transformation in the classic Pal-King fuzzy enhancement algorithm, and the threshold X _T is used as the dividing line , defining the membership degree of the gray level x _gh in different regions, this method of defining the membership degree in the low gray level area and the high gray level area of the image separately also ensures that the information loss of the image in the low gray level area is minimal, thereby ensuring that the image Enhanced effect.

In this embodiment, before transforming from the image domain to the fuzzy domain in step 20121, the grayscale threshold X _T is selected by using the method of maximum variance between classes.

Step 20122, use the fuzzy enhancement operator in the fuzzy domain to perform fuzzy enhancement processing: the adopted fuzzy enhancement operator is μ' _gh =I _r (μ _gh )=I _r (I _r-1 μ _gh ), where r is The number of iterations and it is a positive integer, r=1, 2, ...; where $I_1\left({\mathrm{\&mu;}}_{\mathrm{gh}}\right)=\left\{\begin{array}{cc}{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="2"\; label="gh"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">gh</figure-callout>}}^2/{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="0"\; label="c"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">c</figure-callout>}}& 0\&le;{\mathrm{\&mu;}}_{\mathrm{gh}}\&le;{\mathrm{\&mu;}}_c\\ 1-{(1-{\mathrm{\&mu;}}_{\mathrm{gh}})}^2/(1-{\mathrm{\&mu;}}_c)& {\mathrm{\&mu;}}_c\&le;{\mathrm{\&mu;}}_{\mathrm{gh}}\&le;1\end{array}\right.,$ where μ _c =T(X _C ), where X _C is the transition point and X _C =X _T .

The above formula $I_1\left({\mathrm{\&mu;}}_{\mathrm{gh}}\right)=\left\{\begin{array}{cc}{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="2"\; label="gh"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">gh</figure-callout>}}^2/{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="0"\; label="c"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">c</figure-callout>}}& 0\&le;{\mathrm{\&mu;}}_{\mathrm{gh}}\&le;{\mathrm{\&mu;}}_c\\ 1-{(1-{\mathrm{\&mu;}}_{\mathrm{gh}})}^2/(1-{\mathrm{\&mu;}}_c)& {\mathrm{\&mu;}}_c\&le;{\mathrm{\&mu;}}_{\mathrm{gh}}\&le;1\end{array}\right.$ The nonlinear transformation of increases the value of μ _gh larger than μ _c , and decreases the value of μ _gh smaller than μ _c . Here μ _c has evolved into a generalized transition point.

(6)，将模糊增强处理后得到的μ'_gh进行逆变换，获得增强处理后数字图像中各像素点的灰度值，并获得增强处理后的数字图像。Step 20123, inverse transformation from fuzzy domain to image domain: according to the formula

(6) Perform inverse transformation on the _μ'gh obtained after the fuzzy enhancement processing, obtain the gray value of each pixel in the digital image after the enhancement processing, and obtain the digital image after the enhancement processing.

Since the selection of the fuzzy enhancement threshold (transition point X _c ) in the Pal algorithm is the key to image enhancement, it needs to be obtained by experience or multiple attempts in practical applications. Among them, the more classic method is the method of maximum between-class variance (Ostu), which is simple, stable and effective, and is often used in practical applications. The Ostu threshold selection method gets rid of the limitation of multiple attempts of manual intervention, and the computer can automatically determine the optimal threshold according to the gray information of the image. The principle of the Ostu method is to use the variance between classes as the criterion, and select the gray value with the largest variance between classes as the optimal threshold to realize the automatic selection of the fuzzy enhancement threshold, thereby avoiding manual intervention in the enhancement process.

In this embodiment, before using the maximum inter-class variance method to select the grayscale threshold _XT , first find out all the grayscale values whose number of pixels is 0 from the grayscale variation range of the image to be enhanced, and use processing All the gray values found by the device 3 are marked as calculation-free gray values; when the gray threshold X _T is selected by the maximum inter-class variance method, the range of gray changes of the image to be enhanced is excluded from the Calculate the inter-class variance value when other gray-scale values other than the gray-scale value are used as the threshold value, and find the maximum inter-class variance value from the calculated inter-class variance value, and find the maximum inter-class variance value The grayscale value corresponding to the variance value is the grayscale threshold X _T .

所采集的数字图像各个灰度级出现的概率

阈值X_T将图像中的像素点按其灰度级划分为两类C₀和C₁，C₀={0,1,…t}，C₁={t+1,t+2,…L－1}，并假定类C₀和C₁的像素点数占总像素点数的比率分别为w₀(t)和w₁(t)且二者平均灰度值分别为μ₀(t)和μ₁(t)。When using the traditional maximum inter-class variance method (Ostu) to select fuzzy enhancement, if the number of pixels with gray value s is n _s , the total number of pixels

The probability of occurrence of each gray level of the collected digital image

Threshold X _T divides the pixels in the image into two categories C ₀ and C ₁ according to their gray levels, C ₀ ={0,1,…t}, C ₁ ={t+1,t+2,…L -1}, and assume that the ratio of the number of pixels of class C ₀ and C ₁ to the total number of pixels is w ₀ (t) and w ₁ (t) respectively, and the average gray value of the two is μ ₀ (t) and μ ₁ (t).

For C ₀ there is: $w_0\left(t\right)=\underset{i_2=0}{\overset{\mathrm{<figure-callout\; id="2"\; label="t\; P\; i"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">t</figure-callout>}}{\mathrm{\&Sigma;}}}{P}_{i_{}}{{}_2}_{}=w\left(t\right),{\mathrm{\&mu;}}_0\left(t\right)=\frac{1}{{\mathrm{<figure-callout\; id="0"\; label="w"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">w</figure-callout>}}_0}\underset{i_2=0}{\overset{t}{\mathrm{\&Sigma;}}}{i}_2{\mathrm{<figure-callout\; id="2"\; label="P\; i"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_{i_{}}=\frac{\mathrm{\&mu;}\left(t\right)}{w\left(t\right)};$

For _C1 there is: $w_1\left(t\right)=\underset{i_3={\mathrm{<figure-callout\; id="2"\; label="t"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">t</figure-callout>}}_2+1}{\overset{L-1}{\mathrm{\&Sigma;}}}{\mathrm{<figure-callout\; id="3"\; label="P\; i"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_{i_{}}=1-w\left(t\right),{\mathrm{\&mu;}}_1\left(t\right)=\frac{1}{{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">w</figure-callout>}}_1}\underset{i_3={\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">i</figure-callout>}}_2+1}{\overset{L-1}{\mathrm{\&Sigma;}}}{i}_3{\mathrm{<figure-callout\; id="3"\; label="P\; i"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_{i_{}}=\frac{\mathrm{\&mu;}-\mathrm{\&mu;}\left(t\right)}{1-w\left(t\right)};$

是整体图像灰度的统计均值，则μ=w₀μ₀+w₁μ₁；in

is the statistical mean of the overall image grayscale, then μ=w ₀ μ ₀ +w ₁ μ ₁ ;

Thus the optimal threshold $x_T={\mathrm{\&sigma;}}_{\max }^2\left(t\right)=\mathrm{Arg}\underset{t\&Element;L}{\max }(w_0\left(t\right)\&times;w_1\left(t\right)\&times;{({\mathrm{\&mu;}}_1\left(t\right)-{\mathrm{\&mu;}}_0\left(t\right))}^2)---\left(8\right),$

The process of automatically extracting the optimal fuzzy enhancement threshold X _T is as follows: traverse all gray levels from gray level 0 to L-1 level, and find the value of X _T that satisfies formula (8) to take the maximum value, which is the threshold value X _T . Because the number of pixels of the image may be zero on some gray levels, in order to reduce the number of calculation variances, the present invention proposes an improved fast Ostu method;

$\begin{array}{c}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="0"\; label="w"\; filenames="HDA0000490950320000021.png,HDA0000490950320000022.png"\; state="\{\{state\}\}">w</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="w"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">w</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&times;</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ <span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&times;</span>{<span\; class="notranslate">\; [</span>\frac{\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ <span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}\end{array}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.32</span>}<span\; class="notranslate">\; )</span>$

Assuming that the number of pixels with gray level t' is zero, then P _t' =0

If t'-1 is selected as the threshold, then:

$\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ;</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; IP</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ;</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; IP</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.33</span>}<span\; class="notranslate">\; )</span>$

When t' is selected as the threshold:

$\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.34</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; IP</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; IP</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}{\mathrm{<span\; class="notranslate">\; P</span>}}_{{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; IP</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.35</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; IP</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2.36</span>}<span\; class="notranslate">\; )</span>$

It can be seen from this:

σ ² (t'-1)=σ ² (t') (2.37)

Also assuming that there are continuous gray levels t ₁ , t ₂ ,..., t _n , it can also be pushed upwards:

σ ² (t ₁ -1)=σ ² (t ₁ )=σ ² (t ₂ -1)=σ ² (t ₂ )=…=σ ² (t _n -1)=σ ² (t _n ) ( 2.38)

It can be seen from the above that if the number of pixels of a gray level is zero, it is not necessary to calculate the inter-class variance value when it is used as the threshold, but only need to calculate the value corresponding to the smaller gray level whose nearest neighbor pixel number is not zero The inter-class variance of is used as the inter-class variance value. Therefore, in order to quickly find the maximum value of the inter-class variance, multiple gray levels with equal inter-class variance can be regarded as the same gray level, that is, the number of those pixels is zero The gray value of σ 2 is regarded as non-existent, and the inter-class variance σ ² (t) is directly assigned as zero when it is used as the threshold, without calculating their variance value, which has no influence on the selection of the final result of the threshold, but Increased speed of adaptive selection of enhancement thresholds.

In this embodiment, before the fuzzy enhancement processing in step 20122, the low-pass filtering method is used to smooth the fuzzy set of the image to be enhanced obtained in step 20121; when the low-pass filtering processing is actually performed, the adopted The filter operator is $\frac{1}{16}\left[\begin{array}{ccc}1& 2& 1\\ 2& 4& 2\\ 1& 2& 1\end{array}\right].$

Because images are susceptible to noise pollution during generation and transmission, before image enhancement, the fuzzy set of the image should be smoothed to reduce noise. In this embodiment, the smoothing of the image fuzzy set is realized through the convolution operation of the 3×3 spatial domain low-pass filter operator and the image fuzzy set matrix.

In the present embodiment, when step 2013 performs image segmentation, the process is as follows:

Step 20131, establishment of two-dimensional histogram: use processor 3 to establish a two-dimensional histogram about pixel gray value and neighborhood average gray value of the image to be segmented; any point in the two-dimensional histogram is denoted as ( i, j), wherein i is the abscissa value of the two-dimensional histogram and it is the gray value of any pixel point (m, n) in the image to be segmented, and j is the ordinate of the two-dimensional histogram value and it is the neighborhood average gray value of the pixel point (m, n); ,j) is recorded as h(i,j), where

In this embodiment, when calculating the neighborhood average gray value of a pixel point (m, n), according to the formula $g(m,\mathrm{no})=\frac{1}{d\&times;d}\underset{i1=-(d-1)/2}{\overset{(d-1)/2}{\mathrm{\&Sigma;}}}\underset{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">j</figure-callout>}1=-(d-1)/2}{\overset{(d-1)/2}{\mathrm{\&Sigma;}}}f(m+\mathrm{<figure-callout\; id="1"\; label="i"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">i</figure-callout>}1,\mathrm{no}+j1)$ (6) Calculate, where f(m+i1,n+j1) is the gray value of the pixel point (m+i1,n+j1), where d is the width of the pixel square neighborhood window, which is usually an odd number.

Moreover, the neighborhood average gray value g(m,n) and the pixel gray value f(m,n) have the same gray scale variation range, and the gray scale variation range of both is [0, L), so the step The two-dimensional histogram established in Ⅰ is a square area, see Figure 3 for details, where L-1 is the maximum value of the neighborhood average gray value g(m,n) and pixel gray value f(m,n) value.

In FIG. 3 , the established two-dimensional histogram is divided into four regions by using the threshold vector (i, j). Due to the strong correlation between the pixels inside the target image or the background image, the gray value of the pixel is very close to the average gray value of its neighborhood; and the pixel near the boundary of the target image and the background image, its The difference between the pixel gray value and the neighborhood average gray value is obvious. Therefore, in Figure 3, the 0# area corresponds to the background image, the 1# area corresponds to the target image, and the 2# area and the 3# area represent the distribution of pixels and noise points near the boundary, so it should be used in the 0# and 1# area The gray value of the pixel point and the average gray value of the neighborhood are determined by the two-dimensional fuzzy division of the maximum entropy segmentation method to determine the optimal threshold, so that the amount of information that truly represents the target and the background is maximized.

Step 20132, fuzzy parameter combination optimization: the processor 3 calls the fuzzy parameter combination optimization module, and uses the particle swarm optimization algorithm to optimize the fuzzy parameter combination used in the image segmentation method based on two-dimensional fuzzy partition maximum entropy, and obtains the optimized combination of fuzzy parameters.

In this step, before optimizing the combination of fuzzy parameters, according to the two-dimensional histogram established in step 20131, the functional relational expression of the two-dimensional fuzzy entropy when the image to be segmented is segmented is calculated, and the calculated The obtained two-dimensional fuzzy entropy function relation is used as the fitness function when the particle swarm optimization algorithm is used to optimize the combination of fuzzy parameters.

In this embodiment, the image to be segmented in step 20131 is composed of the target image O and the background image P; wherein the membership function of the target image O is μ _o (i, j) = μ _ox (i; a, b) μ _oy (j;c,d)(1).

The membership function of background image P μ _b (i,j)=μ _bx (i;a,b)μ _oy (j;c,d)+μ _ox (i;a,b)μ _by (j;c, d)+μ _bx (i;a,b)μ _by (j;c,d)(2).

In formulas (1) and (2), μ _ox (i;a,b) and μ _oy (j;c,d) are both one-dimensional membership functions of the target image O and both are S functions, μ _bx (i;a,b) and μ _by (j;c,d) are the one-dimensional membership functions of the background image P and both are S functions, μ _bx (i;a,b)=1-μ _ox (i;a,b), μ _by (j;c,d)=1-μ _oy (j;c,d), where a, b, c, and d are a combination of the target image O and the background image P Parameters that control the shape of the dimension membership function.

in, ${\mathrm{\&mu;}}_{\mathrm{ox}}(i;a,b)=\left\{\begin{array}{cc}0& i\&le;a\\ 2\&times;{\left(\frac{i-a}{b-a}\right)}^2& a<i\&le;\frac{a+\mathrm{<figure-callout\; id="2"\; label="b"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">b</figure-callout>}}{2}\\ 1-2\&times;{\left(\frac{i-b}{b-a}\right)}^2& \frac{a+\mathrm{<figure-callout\; id="2"\; label="b"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">b</figure-callout>}}{2}<i\&le;b\\ 1& b<i\&le;L-1\end{array}\right.;$

${\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; oy</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ;</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="c"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">c</figure-callout></span>}\\ \mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; \&times;</span>{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; c</span>}}{\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}& \mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&le;</span>\frac{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="d"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">d</figure-callout></span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ \mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; \&times;</span>{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; d</span>}}{\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}& \frac{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="d"\; filenames="HDA0000490950320000012.png,HDA0000490950320000021.png"\; state="\{\{state\}\}">d</figure-callout></span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; d</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right.<span\; class="notranslate">\; .</span>$

When calculating the functional relational expression of the two-dimensional fuzzy entropy in step 20132, first, according to the two-dimensional histogram established in step 20131, the minimum value g _min and the maximum value g of the pixel gray value of the image to be segmented are calculated _max and the minimum value s _min and maximum value s _max of the neighborhood average gray value are determined respectively. In this embodiment, g _max =s _max =L-1, and g _min =s _min =0. Among them, L-1=255.

The functional relational expression of the two-dimensional fuzzy entropy calculated in step 20132 is:

$h\left(P\right)=-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)})-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)})$ (3), in formula (3) $p\left(o\right)=\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_o(i,j,)h_{\mathrm{ij}},p\left(B\right)=\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_b(i,j,)h_{\mathrm{ij}},$ where h(i,j) is the frequency at which point (i,j) occurs in step I.

When the particle swarm optimization algorithm is used to optimize the fuzzy parameter combination in step 20132, the optimized fuzzy parameter combination is (a, b, c, d).

In this embodiment, when performing parameter combination optimization of two-dimensional fuzzy partition maximum entropy in step 20132, the following steps are included:

Step Ⅱ-1. Particle swarm initialization: take a value of the parameter combination as a particle, and form multiple particles into an initialized particle swarm; denoted as (a _k , b _k , c _k , d _k ), where k is a positive integer and its k=1, 2, 3, ~, K, wherein K is a positive integer and it is the number of particles contained in the particle group, a _k is a random value of parameter a, and b _k is A random value of parameter b, c _k is a random value of parameter c, d _k is a random value of parameter d, a _k <b _k and c _k <d _k .

In this embodiment, K=15.

In actual use, K can be set to a value between 10 and 100 according to specific needs.

Step Ⅱ-2, fitness function determination:

Will $h\left(P\right)=-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)})-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)})$ (3), as a fitness function.

更新为当前时间第k个粒子的位置，其中Pbestk为当前时刻第k个粒子所达到的最大适应度值且其为当前时刻第k个粒子的个体极值，

为当前时刻第k个粒子的个体最优位置；其中，t为当前迭代次数且其为正整数。Step Ⅱ-3. Particle fitness evaluation: evaluate the fitness of all particles at the current moment, and the fitness evaluation methods of all particles are the same; where, when evaluating the fitness of the kth particle at the current moment, first According to the fitness function determined in step Ⅱ-2, the fitness value of the kth particle at the current moment is calculated and recorded as fitnessk, and the difference between the calculated fitnessk and Pbestk is compared: when the comparison results in fitnessk > When Pbestk, Pbestk=fitnessk, and set

Update to the position of the kth particle at the current time, where Pbestk is the maximum fitness value achieved by the kth particle at the current moment and it is the individual extremum of the kth particle at the current moment,

is the individual optimal position of the kth particle at the current moment; among them, t is the current iteration number and it is a positive integer.

更新为当前时间适应度值最大的粒子的位置，其中gbest为当前时刻的全局极值，

为当前时刻的群体最优位置。After the fitness value of all particles at the current moment is calculated according to the fitness function determined in step II-2, the fitness value of the particle with the largest fitness value at the current moment is recorded as fitnesskbest, and fitnesskbest is compared with gbest Difference comparison: when the comparison shows that fitnesskbest>gbest, gbest=fitnesskbest, and the

Update to the position of the particle with the largest fitness value at the current time, where gbest is the global extremum at the current moment,

is the optimal position of the group at the current moment.

Step Ⅱ-4. Judging whether the iteration termination condition is satisfied: when the iteration termination condition is satisfied, the parameter combination optimization process is completed; otherwise, the position and speed of each particle at the next moment are obtained according to the update of the particle swarm optimization algorithm, and return to step Ⅱ -3.

The iteration termination condition in step II-4 is that the current iteration number t reaches the preset maximum iteration number I _max or Δg≤e, where Δg=|gbest-gmax|, where is the global extreme value of gbest at the current moment, and gmax is The originally set target fitness value, e is a positive number and it is a preset deviation value.

In this embodiment, the maximum number of iterations I _max =30. In actual use, the maximum number of iterations I _max can be adjusted between 20 and 200 according to specific needs.

In this embodiment, when the particle swarm is initialized in step II-1, (a k , c k ) among particles (a _k , b _k , c _k , d _{k )} is the initial velocity vector of the kth particle, ₍ b _k , d _k ) is the initial position of the kth particle.

When the position and velocity of each particle at the next moment are obtained according to the update of the particle swarm optimization algorithm in step II-4, the updating methods of the position and velocity of all particles are the same; among them, the velocity and velocity of the kth particle at the next moment are When the position is updated, first calculate the velocity vector of the kth particle at the next moment according to the velocity vector, position, individual extremum Pbestk and global extremum of the kth particle at the current moment, and calculate the velocity vector of the kth particle at the current moment according to the kth particle at the current moment position and the calculated velocity vector of the kth particle at the next moment to calculate the position of the kth particle at the next moment.

和位置

公式(4)和(5)中

为当前时刻第k个粒子的位置，公式(4)中

为当前时刻第k个粒子的速度矢量，c₁和c₂均为加速度系数且c₁+c₂=4，r₁和r₂为[0,1]之间的均匀分布的随机数；ω为惯性权重且其随迭代次数的增加线性减小，

式中ω_max和ω_min分别为预先设定的惯性权重最大值和最小值，t为当前迭代次数，I_max为预先设定的最大迭代次数。And, when updating the velocity and position of the kth particle at the next moment in step II-4, according to $v_k^{t+1}=\mathrm{\&omega;}{v}_k^t+c_1{r}_1(g_{\mathrm{kbest}}^t-x_k^t)+c_2{r}_2(g_{\mathrm{gbest}}^t-x_k^t)$ (4) and the formula $x_k^{t+1}=x_k^t+v_k^{t+1}$ (5) Calculate the velocity vector of the kth particle at the next moment

and location

In formulas (4) and (5)

is the position of the kth particle at the current moment, in the formula (4)

is the velocity vector of the kth particle at the current moment, c ₁ and c ₂ are both acceleration coefficients and c ₁ +c ₂ =4, r ₁ and r ₂ are uniformly distributed random numbers between [0,1]; ω is the inertia weight and it decreases linearly with the increase of the number of iterations,

where ω _max and ω _min are the preset maximum and minimum inertia weights respectively, t is the current iteration number, and I _max is the preset maximum iteration number.

In this embodiment, ω _max =0.9, ω _min =0.4, c ₁ =c ₂ =2.

In this embodiment, before the initialization of the particle swarm in step II-1, the search ranges of a _k , b _k , c _k and d _k need to be determined first, where the pixel grayscale of the image to be segmented in step I is The minimum value is g _min and its minimum value is g _max ; the neighborhood size of a pixel point (m, n) is d×d pixels, and the average gray value of its neighborhood is s _min and its average gray value is maximum s _max , a _k =g _min ,..., g _max -1, b _k =g _min +1,..., g _max , c _k =s _min ,..., s _max -1, d _k =s _min +1, ..., s _max .

In this embodiment, d=5.

During actual use, the value of d can be adjusted accordingly according to specific needs.

Step 20133, image segmentation: the processor 3 uses the fuzzy parameter combination optimized in step 20132 to classify each pixel in the image to be segmented according to the image segmentation method based on two-dimensional fuzzy partition maximum entropy, and The image segmentation process is completed correspondingly, and the target image after segmentation is obtained.

In this embodiment, after obtaining the optimized combination of fuzzy parameters (a, b, c, d), the pixels are classified according to the principle of maximum membership degree: when μ _o (i, j) ≥ 0.5, the class The pixels are divided into the target area, otherwise they are divided into the background area, see Figure 4 for details. In FIG. 4 , the square where μ _o (i,j)≥0.5 is represented as the target area after image segmentation.

The above are only preferred embodiments of the present invention, and do not limit the present invention in any way. All simple modifications, changes and equivalent structural changes made to the above embodiments according to the technical essence of the present invention still belong to the technical aspects of the present invention. within the scope of protection of the scheme.

Claims (10)

1. An image type fire flame recognition method is characterized in that the method may further comprise the steps: Step 1. Image collection: using an image collection unit to collect digital images of the area to be detected according to a preset sampling frequency f _s , and synchronously transmitting the digital images collected at each sampling moment to the processor (3); The image acquisition unit is connected to the processor (3); Step 2, image processing: the processor (3) performs image processing on the digital images collected at each sampling time in step 1 in chronological order, and the processing methods for the digital images collected at each sampling time are the same; When processing the digital image collected at any sampling moment in step 1, the following steps are included: Step 201, image preprocessing, the process is as follows: Step 2011, image reception and synchronous storage: the processor (3) synchronously stores the received digital image collected at the current sampling moment in the data storage (4), and the data storage (4) and the processor (3) connected; Step 2012, image enhancement: the processor (3) performs enhancement processing on the digital image collected at the current sampling moment to obtain an enhanced digital image; Step 2013, image segmentation: Segment the enhanced digital image in step 2012 through the processor (3) to obtain the target image; Step 202, fire identification: use the pre-established binary classification model to process the target image described in step 2013, and obtain the fire state category of the area to be detected at the current sampling moment; the fire state category includes flame and no flame Two categories, the two classification model is a support vector machine model for classifying two categories with flame and no flame; The establishment process of the two classification models is as follows: Step 1. Image information acquisition: using the image acquisition unit to acquire multi-frame digital images 1 of the area to be detected when a fire occurs and multi-frame digital images 2 of the area to be detected when no fire occurs; Step II, feature extraction: perform feature extraction on the multi-frame digital image one and the multi-frame digital image respectively, and extract a group of characteristic parameters that can represent and distinguish the digital image from each digital image, and the The group feature parameters include M feature quantities, and the M feature quantities are numbered, and the M feature quantities form a feature vector, where M≥2; Step III, acquisition of training samples: from the feature vectors of multiple frames of the digital image one and multiple frames of the digital image two obtained after feature extraction in step II, respectively select the feature vectors and The eigenvectors of the second digital image in frame m2 form a training sample set; wherein m1 and m2 are positive integers and m1=40～100, m2=40～100; the number of training samples in the training sample set is m1+m2 indivual; Step Ⅳ, two-category model establishment, the process is as follows: Step IV-1, kernel function selection: select radial basis function as the kernel function of the binary classification model; Step IV-2, determination of classification function: After the factor γ to be penalized and the kernel parameter ^σ2 of the radial basis function selected in step IV-1 are determined, the classification function of the binary classification model is obtained, and the binary classification is completed Model establishment process; where, γ=C ^-2 , σ=D ^-1 , 0.01<C≤10, 0.01<D≤50; When determining the penalty factor γ and the kernel parameter σ ² , the parameters C and D are optimized using the conjugate gradient method to obtain the optimized parameters C and D, and then according to γ=C ^-2 and σ=D ^-1 The optimized parameters C and D are converted into penalty factor γ and kernel parameter σ ² ; Step V, binary classification model training: input the m1+m2 training samples in the training sample set mentioned in step III to the binary classification model established in step IV for training. 2. according to a kind of image type fire flame recognition method described in claim 1, it is characterized in that: the total number of training samples in the training sample set described in step III is N and N=m1+m2; Carry out two classification models in step IV Before setting up, the N training samples in the training sample set are numbered, the number of the pth training sample in the training sample set is p, p is a positive integer and p=1, 2, ..., N; A training sample is recorded as (x _p , y _p ), where x _p is the characteristic parameter of the pth training sample, y _p is the category number of the pth training sample and y _p =1 or -1, where the category number is 1 means there is flame, the category number is -1 means no flame; When using the conjugate gradient method to optimize parameters C and D in step IV-2, use the m1+m2 training samples in the training sample set mentioned in step III to optimize, and the optimization process is as follows: Step 1, objective function determination: $\mathrm{sse}(C,\mathrm{D.})=\frac{1}{2}\underset{p=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_p^2=\frac{1}{2}\underset{p=1}{\overset{N}{\mathrm{\&Sigma;}}}{[\tilde{K}(p,p^{-})\&Center\; Dot;{\mathrm{the\; s}}_p-\mathrm{the\; y}\left(p\right)]}^2$ (1), where sse(C, D) is the sum of the squares of the leave-one-out prediction error, p is the serial number of each training sample in the training sample set, and e _p is the p-th training result of the binary classification model established in step IV. The sample prediction error and $e_p=\tilde{K}(p,p^{-})\&Center\; Dot;{\mathrm{the\; s}}_p-\mathrm{the\; y}\left(p\right);$ in, ${\mathrm{the\; s}}_p=\mathrm{the\; s}\left(p^{-}\right)-\frac{\mathrm{the\; s}\left(p\right)}{\left(A^{-1}\right)(p,p)}\left(A^{-1}\right)(p^{-},p),$ In the formula, s(p ^- ) is a vector composed of the remaining elements after removing the p-th element in the matrix s; s(p) is the p-th element of the matrix s, and (A ^-1 )(p ^- ,p) is the matrix A column vector composed of the remaining elements after the p-th column of A ^-1 except the p-th element, (A ^-1 )(p,p) is the p-th element of the p-th column of matrix A ^-1 ;

for the matrix

The column vector and matrix composed of the remaining elements after removing the p-th column of the p-th column

is the augmented matrix of matrix K, where the matrix $K=\left|\begin{array}{cccc}K(x_1,x_1)& K(x_1,x_2)& \&CenterDot;\&CenterDot;\&CenterDot;& K(x_1,x_N)\\ K(x_2,x_1)& K(x_2,x_2)& \&Center\; Dot;\&Center\; Dot;\&Center\; Dot;& K(x_2,x_N)\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&Center\; Dot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ K(x_N,x_1)& K(x_N,x_2)& \&Center\; Dot;\&Center\; Dot;\&Center\; Dot;& K(x_N,x_N)\end{array}\right.|;$ Matrix A ^-1 represents the inverse matrix of matrix A, matrix $A=\left[\begin{array}{cc}K+C^2\&Center\; Dot;I& I_N\\ {I_N}^T& 0\end{array}\right],$ Among them, the matrix I is the identity matrix, the matrix I _N =[1,1,…,1] ^T , T represents the transposition of the matrix, the matrix I _N contains N elements and all N elements are 1; the matrix s=A ^{- 1} Y, matrix $Y=\left[\begin{array}{c}{\mathrm{the\; y}}_1\\ {\mathrm{the\; y}}_2\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ {\mathrm{the\; y}}_N\\ 0\end{array}\right],$ Wherein y ₁ , y ₂ , ..., y _N are respectively the categories of the N training samples in the training sample set; Step ii. Initial parameter setting: determine the initial values C ₁ and D ₁ of parameters C and D respectively, and set the recognition error threshold ε and ε>0; Step Ⅲ, the gradient g _k calculation of the current iteration: according to the formula

Calculate the gradient g _k of the objective function to C _k and D _k in step i, k is the number of iterations and k=1, 2, ...; if ||g _k ||≤ε, stop the calculation, at this time C _k and D _k are the optimized parameters C and D respectively; otherwise, go to step ⅳ; in, $\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k}=\underset{p=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_p\&CenterDot;\tilde{K}(p,p^{-})\&Center\; Dot;\frac{\&PartialD;{\mathrm{the\; s}}_p}{\&PartialD;C_k};$ $\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k}=\underset{p=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{e}_p\&CenterDot;[\frac{\&PartialD;\tilde{K}(p,p^{-})}{\&PartialD;{\mathrm{D.}}_k}\&Center\; Dot;\mathrm{the\; s}\left(p^{-}\right)+\tilde{K}(p,p^{-})\&CenterDot;\frac{\&PartialD;{\mathrm{the\; s}}_p}{\&PartialD;{\mathrm{D.}}_k}],$ In the formula

for the matrix

s(p ^- ) is a vector composed of the remaining elements after removing the p-th element in the matrix s, and e _p is the two elements established in step IV The prediction error of the classification model for the pth training sample; Step ⅳ. Calculation of the search direction d _k of the current iteration: according to the formula $d_k=\left\{\begin{array}{cc}-g_k& k=1\\ -g_k+{\mathrm{\&beta;}}_k{d}_{k-1}& k\&Greater\; Equal;2\end{array}\right.,$ Calculate the search direction d _k of the current iteration, where d _k-1 is the search direction of the k-1th iteration, β _k =||g _k || ² /||g _k-1 || ² , g _{k- 1} is the gradient of the k-1th iteration; Step ⅳ, determine the search step size λ _k of the current iteration: search along the search direction d _k determined in step ⅳ, and find out that satisfies the formula $\mathrm{sse}(C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k},{\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k})=\underset{{\mathrm{\&lambda;}}_k>0}{\min }\mathrm{sse}(C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k},{\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k})$ The search step size λ _k , where

Indicates that in (0, +∞) to find the

the step size λ _k to reach the minimum; Step ⅵ, according to the formula $C_{k+1}=C_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;C_k}$ and ${\mathrm{D.}}_{k+1}={\mathrm{D.}}_k-{\mathrm{\&lambda;}}_k\frac{\&PartialD;\mathrm{sse}}{\&PartialD;{\mathrm{D.}}_k},$ Calculate C _k+1 and D _k+1 ; Step ⅶ, set k=k+1, and then return to step Ⅲ for the next iteration; The radial basis function selected in step IV-1 is

The regression function of the radial basis function is

In the formula, α _t and b are regression parameters, s is a positive integer and s=1, 2,..., N, t is a positive integer and t=1, 2,..., N. 3. An image-type fire flame recognition method according to claim 1 or 2, characterized in that: M=6 in step II, and the 6 feature quantities are respectively area, similarity, moment characteristic, density, texture features and strobe characteristics. 4. according to a kind of image type fire flame identification method described in claim 2, it is characterized in that: when _C1 and _D1 are determined in the step ii, adopt grid search method or the method for randomly extracting numerical value to C1 and _D1 D ₁ is used to determine; when C ₁ and D ₁ are determined by randomly selecting values, C ₁ is a value randomly selected in (0.01, 1], and D ₁ is a value randomly selected in (0.01, 50] ; When using the grid search method to determine C ₁ and D ₁ , first divide the grid with a step size of 10 ^-3 , then use C and D as independent variables and take the objective function described in step i as the dependent variable to make a three-dimensional network Grid map, and then find multiple sets of parameters of C and D through grid search, and finally take the average value of multiple sets of parameters as C ₁ and D ₁ . 5. An image-based fire flame recognition method according to claim 1 or 2, characterized in that: when image enhancement is performed in step 2012, an image enhancement method based on fuzzy logic is used for enhancement processing. 6. according to a kind of image type fire flame identification method described in claim 5, it is characterized in that: when adopting the image enhancement method based on fuzzy logic to carry out enhancement processing, process is as follows: Step 20121, transform from the image domain to the fuzzy domain: according to the membership function ${\mathrm{\&mu;}}_{\mathrm{gh}}=T\left(x_{\mathrm{gh}}\right)=\left\{\begin{array}{cc}{x}_{\mathrm{gh}}/x_T& x_{\mathrm{gh}}\&le;x_T\\ x_{\mathrm{gh}}/x_{\max }& x_{\mathrm{gh}}>x_T\end{array}\right.$ (7), the gray value of each pixel of the image to be enhanced is mapped to the fuzzy membership degree of the fuzzy set, and the fuzzy set of the image to be enhanced is correspondingly obtained; x _gh is in the image to be enhanced The gray value of any pixel point (g, h), X _T is the gray threshold selected when the image to be enhanced is enhanced using the image enhancement method based on fuzzy logic, and X _max is the image to be enhanced The maximum gray value of ; Step 20122, use the fuzzy enhancement operator in the fuzzy domain to perform fuzzy enhancement processing: the adopted fuzzy enhancement operator is μ' _gh =I _r (μ _gh )=I _r (I _r-1 μ _gh ), where r is The number of iterations and it is a positive integer, r=1, 2, ...; where $I_1\left({\mathrm{\&mu;}}_{\mathrm{gh}}\right)=\left\{\begin{array}{cc}{\mathrm{\&mu;}}_{\mathrm{gh}}^2/{\mathrm{\&mu;}}_c& 0\&le;{\mathrm{\&mu;}}_{\mathrm{gh}}\&le;{\mathrm{\&mu;}}_c\\ 1-{(1-{\mathrm{\&mu;}}_{\mathrm{gh}})}^2/(1-{\mathrm{\&mu;}}_c)& {\mathrm{\&mu;}}_c\&le;{\mathrm{\&mu;}}_{\mathrm{gh}}\&le;1\end{array}\right.,$ In the formula, μ _c =T(X _C ), where X _C is the transition point and X _C =X _T ; Step 20123, inverse transformation from fuzzy domain to image domain: according to the formula

(6) Perform inverse transformation on the _μ'gh obtained after the fuzzy enhancement processing, obtain the gray value of each pixel in the digital image after the enhancement processing, and obtain the digital image after the enhancement processing. 7. According to a kind of image-type fire flame recognition method according to claim 6, it is characterized in that: before transforming from image domain to fuzzy domain in step 20121, first adopt the method of maximum variance between classes to select the gray scale threshold value X _T ; Before using the maximum inter-class variance method to select the grayscale threshold _XT , all grayscale values with a pixel number of 0 are found from the grayscale variation range of the image to be enhanced, and the processor (3) is used to All the gray values found are marked as calculation-free gray values; when using the maximum inter-class variance method to select the gray threshold X _T , the calculation-free gray values are excluded from the gray-scale variation range of the image to be enhanced. Calculate the inter-class variance value when the gray value other than the gray value is used as the threshold value, and find the maximum inter-class variance value from the calculated inter-class variance value, and find the maximum inter-class variance value The corresponding gray value is the gray threshold X _T . 8. According to claim 1 or 2 described a kind of image type fire flame recognition method, it is characterized in that: in step 1, the size of the digital image collected at each sampling moment is M1 * N1 pixel points; When step 2013 performs image segmentation, the process is as follows: Step 20131, two-dimensional histogram establishment: use the processor (3) to establish a two-dimensional histogram of the pixel gray value and the neighborhood average gray value of the image to be segmented; record any point in the two-dimensional histogram Be (i, j), wherein i is the abscissa value of the two-dimensional histogram and it is the gray value of any pixel point (m, n) in the image to be segmented, and j is the value of the two-dimensional histogram The ordinate value is the neighborhood average gray value of the pixel point (m, n); the frequency of occurrence of any point (i, j) in the established two-dimensional histogram is recorded as C (i, j), and the point The frequency of occurrence of (i,j) is denoted as h(i,j), where

Step 20132, fuzzy parameter combination optimization: the processor (3) calls the fuzzy parameter combination optimization module, and uses the particle swarm optimization algorithm to optimize the fuzzy parameter combination used in the image segmentation method based on two-dimensional fuzzy partition maximum entropy, and obtains Optimized fuzzy parameter combination; In this step, before optimizing the combination of fuzzy parameters, according to the two-dimensional histogram established in step 20131, the functional relational expression of the two-dimensional fuzzy entropy when the image to be segmented is segmented is calculated, and the calculated The obtained two-dimensional fuzzy entropy function relation is used as the fitness function when the particle swarm optimization algorithm is used to optimize the combination of fuzzy parameters; Step 20133, Image Segmentation: The processor (3) utilizes the fuzzy parameter combination optimized in step 20132, and classifies each pixel in the image to be segmented according to the image segmentation method based on two-dimensional fuzzy division maximum entropy , and correspondingly complete the image segmentation process to obtain the segmented target image. 9. An image-type fire flame recognition method according to claim 8, characterized in that: the image to be segmented in step 20131 is composed of a target image O and a background image P; wherein the membership function of the target image O is μ _o (i,j)=μ _ox (i;a,b)μ _oy (j;c,d)(1); the membership function of the background image P μ _b (i,j)=μ _bx (i;a ,b)μ _oy (j;c,d)+μ _ox (i;a,b)μ _by (j;c,d)+μ _bx (i;a,b)μ _by (j;c,d) (2); In formulas (1) and (2), μ _ox (i;a,b) and μ _oy (j;c,d) are both one-dimensional membership functions of the target image O and both are S functions, μ _bx (i;a,b) and μ _by (j;c,d) are the one-dimensional membership functions of the background image P and both are S functions, μ _bx (i;a,b)=1-μ _ox (i;a,b), μ _by (j;c,d)=1-μ _oy (j;c,d), where a, b, c, and d are a combination of the target image O and the background image P Parameters that control the shape of the dimension membership function; When calculating the functional relational expression of the two-dimensional fuzzy entropy in step 20132, first, according to the two-dimensional histogram established in step I, the minimum value g _min and the maximum value g _max and the minimum value s _min and maximum value s _max of the neighborhood average gray value are determined respectively; The functional relational expression of the two-dimensional fuzzy entropy calculated in step 20132 is: $h\left(P\right)=-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)})-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)})$ (3), in formula (3) $p\left(o\right)=\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_o(i,j,)h_{\mathrm{ij}},p\left(B\right)=\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_b(i,j,)h_{\mathrm{ij}},$ where h(i,j) is the frequency of occurrence of the point (i,j) mentioned in step I; When the particle swarm optimization algorithm is used to optimize the fuzzy parameter combination in step 20132, the optimized fuzzy parameter combination is (a, b, c, d). 10. according to a kind of image type fire flame recognition method described in claim 9, it is characterized in that: when carrying out the parameter combination optimization of two-dimensional fuzzy division maximum entropy in the step 20132, comprise the following steps: Step Ⅱ-1. Particle swarm initialization: take a value of the parameter combination as a particle, and form multiple particles into an initialized particle swarm; denoted as (a _k , b _k , c _k , d _k ), where k is a positive integer and its k=1, 2, 3, ~, K, wherein K is a positive integer and it is the number of particles contained in the particle group, a _k is a random value of parameter a, and b _k is A random value of parameter b, c _k is a random value of parameter c, d _k is a random value of parameter d, a _k < b _k and c _k < d _k ; Step Ⅱ-2, fitness function determination: Will $h\left(P\right)=-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_o(i,j)h_{\mathrm{ij}}}{p\left(o\right)})-\underset{i=g_{\min }}{\overset{g_{\max }}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{the\; s}}_{\min }}{\overset{{\mathrm{the\; s}}_{\max }}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)}\exp (1-\log \frac{{\mathrm{\&mu;}}_b(i,j)h_{\mathrm{ij}}}{p\left(B\right)})$ (3), as a fitness function; Step Ⅱ-3. Particle fitness evaluation: evaluate the fitness of all particles at the current moment, and the fitness evaluation methods of all particles are the same; where, when evaluating the fitness of the kth particle at the current moment, first According to the fitness function determined in step Ⅱ-2, the fitness value of the kth particle at the current moment is calculated and recorded as fitnessk, and the difference between the calculated fitnessk and Pbestk is compared: when the comparison results in fitnessk > When Pbestk, Pbestk=fitnessk, and set

Update to the position of the kth particle at the current time, where Pbestk is the maximum fitness value achieved by the kth particle at the current moment and it is the individual extremum of the kth particle at the current moment,

is the individual optimal position of the kth particle at the current moment; among them, t is the current iteration number and it is a positive integer; After the fitness value of all particles at the current moment is calculated according to the fitness function determined in step II-2, the fitness value of the particle with the largest fitness value at the current moment is recorded as fitnesskbest, and fitnesskbest is compared with gbest Difference comparison: when the comparison shows that fitnesskbest>gbest, gbest=fitnesskbest, and the Update to the position of the particle with the largest fitness value at the current time, where gbest is the global extremum at the current moment,

is the optimal position of the group at the current moment; Step Ⅱ-4. Judging whether the iteration termination condition is satisfied: when the iteration termination condition is satisfied, the parameter combination optimization process is completed; otherwise, the position and speed of each particle at the next moment are obtained according to the update of the particle swarm optimization algorithm, and return to step Ⅱ -3; The iteration termination condition in step Ⅱ-4 is that the current iteration number t reaches the preset maximum iteration number I _max or Δg≤e, where Δg=|gbest-gmax|, where is the global extremum value of gbest at the current moment , gmax is the originally set target fitness value, e is a positive number and it is a preset deviation value.

Images