CN107958073B - Particle cluster algorithm optimization-based color image retrieval method - Google Patents

Abstract

The invention discloses a color image retrieval method optimized based on particle cluster algorithm, and belongs to the technical field of image retrieval. The invention firstly extracts the underlying features of each image in an image database and saves them in the image feature database; and then describes the different image features. Assign different similarity measurement formulas; then train the PSO algorithm to obtain the weights of the database similarity measurement; when performing image retrieval processing, extract the corresponding underlying features of the query image, and compare the query image with the target database. The descriptors of the features are extracted, and based on the trained similarity measure weights, the similarity measures of different features are sorted uniformly, and the top K most similar pictures are returned as the retrieval result. Compared with the prior art, the present invention combines and optimizes various feature extraction methods, and improves the retrieval accuracy of the CBIR retrieval system by combining multiple feature descriptors.

Description

A Color Image Retrieval Method Based on Particle Cluster Algorithm Optimization

technical field

The invention belongs to the technical field of image processing and computer vision, and in particular relates to a color image retrieval method.

Background technique

Content-based image retrieval (CBIR) refers to that the query condition itself is an image, or a description of the image content, and its indexing method is to extract the underlying features, and then compare these features by computing and the distance between the query conditions to determine the similarity of the two images. Since the 1970s, content-based image retrieval has been a hot research field, and the current mainstream search engines have launched their own image search functions.

The existing CBIR technology is mainly based on knowledge in the fields of computer vision, pattern recognition, image processing, and image understanding, and then introduces new media data representation methods and data models to model image features. At the same time, in order to improve the accuracy of retrieval, CBIR also gradually involves cognitive science, artificial intelligence, human-computer interaction, information retrieval and other fields. A retrieval system with reliable performance and efficient operation is developed. In the existing technology, the underlying image features such as color, texture, outline, shape, spatial position, etc. belong to the earlier used and relatively mature technology. Among these low-level features, the most frequently used features are color, shape, and texture.

Different features have different emphasis on the description of image information. The CBIR system that only uses a single feature as an image feature descriptor has very narrow application scenarios and cannot meet the requirements of modern society for image retrieval systems. Therefore, the fusion of multiple features is introduced. in the CBIR program. The CBIR system that integrates multiple features brings another problem. Since different features have differences in similarity matching, it is inaccurate to use a unified similarity metric for similarity judgment, which may reduce the effect of retrieval. . Therefore, the use of multiple similarity metrics in the fusion of multi-feature CBIR system can greatly improve the effect of the system.

In addition to image feature extraction, another difficulty is modeling the organization of features. In the process of extracting the underlying features of digital images, there are many different modeling methods. Taking color space as an example, there are different color models such as RGB, HSV, CLEL*a*b* and CLEL*u*v* for modeling the color space. The color quantization can be roughly divided into two methods: segmentation algorithm and clustering algorithm. The basic idea of the segmentation algorithm is to use the most frequently occurring M colors in the image as the palette, and then map the rest of the colors to the palette according to the closest distance criterion, thereby reconstructing the image hierarchy. The clustering algorithm first selects several cluster centers, and then iteratively clusters the colors according to a certain combination until the most suitable cluster is found. Commonly used clustering methods include KMeans, fuzzy CMeans, etc. In addition, there is a rough division of the color space. Taking HSV as an example, the brightness is divided into multiple frequency bands according to saturation, and the number of gray levels in each frequency band is counted. The most commonly used methods are color histograms and color moments. Common methods for shape feature description extraction include Freeman chain code, curvature space descriptor, Fourier descriptor, wavelet descriptor and boundary matrix. Texture analysis method is the process of obtaining qualitative and quantitative description of texture. Texture analysis includes 4 kinds: statistical analysis method, structural analysis method, model analysis and spectrum analysis method. In recent years, Local Binary Pattern (LBP) is the most commonly used method for statistical analysis of texture features. Because LBP is relatively simple, has low computational complexity and is not affected by factors such as illumination, it is widely used in texture analysis and face analysis. Recognition is widely used. Model analysis methods include Gibbs model, Markov Random Field (MRF) and so on. Fourier transform, wavelet transform and Gabor transform in spectral texture analysis methods.

Although CBIR technology has been studied for decades, many key issues remain to be addressed. To sum up, it mainly includes three aspects: the effective extraction of image features, the definition of image similarity and dissimilarity, and the semantic gap between low-level image features and high-level semantics. Therefore, it is necessary to provide a more optimized algorithm to make up for the shortcomings of the current image retrieval, so as to meet the actual multi-faceted needs of the user's image retrieval.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a color image retrieval method optimized based on the particle cluster algorithm to improve the retrieval accuracy of the image retrieval method in view of the above existing problems.

The color image retrieval method optimized based on the particle cluster algorithm of the present invention comprises the following steps:

1. Train the similarity measure weight based on the particle cluster algorithm:

101: Use different types of query images as training samples, and extract color features, texture features and object shape features of the training samples;

The color features are: the color histogram and color moment features of the image in the HSV color space;

The texture feature is: filtering the image by using the Gabor of the real part, and extracting the mean and standard deviation of the filtering output;

The object shape feature is: extracting the object shape feature of the image based on the object shape feature extraction method based on the region shape;

102: Extract the color feature, texture feature and object shape feature of each retrieved image in the image library;

103: Use the particle cluster algorithm to optimize the weights of the similarity measure of the three types of image features:

Initialize the weight ω _c of the similarity measure D _c of the color feature, the weight ω _t of the similarity measure D _t of the texture feature, the weight ω _s of the similarity measure D _s of the object shape feature, and define the weight of each particle. The particle position is (ω _c , ω _t , ω _s ), where 0≤ω _c ,ω _t ,ω _s ≤1; initialize the number of particles k, and the local optimal position of each particle and the global optimal position of the particle swarm ;

Perform image query processing with n training samples of the same type as query images each time: Calculate the similarity measure D _c of color features and the similarity of texture features between each training sample and each retrieved image in the image database. Measure D _t and the similarity measure D _s of the object shape feature, and based on the weights ω _c , ω _t , ω _s , the three similarity measures are weighted and summed to obtain the total similarity measure D; The retrieved image corresponding to the similarity measure D is used as the retrieval result;

Iteratively update the particle position, local optimal position and global optimal position based on the current retrieval accuracy until the iteration converges, and save the latest updated particle position;

The retrieval accuracy is: the ratio of the number of retrieved images matching the query image in the retrieval result to the value K;

2. Image retrieval:

Extract the color features, texture features and object shape features of the current query image;

and calculate the similarity measure D _c of the color feature, the similarity measure D _t of the texture feature and the similarity measure D _s of the object shape feature with each retrieved image in the image library respectively;

Based on the weights ω _c , ω _t and ω _s obtained by training, the current total similarity measure D=ω _c D _c +ω _t D _t +ω _s D _s is obtained;

The retrieved images corresponding to the top K largest total similarity measures D are used as retrieval results and output.

To sum up, due to the adoption of the above-mentioned technical solutions, the beneficial effects of the present invention are:

Compared with the prior art, the present invention combines and optimizes various feature extraction methods, and improves the retrieval accuracy of the CBIR retrieval system by combining multiple feature descriptors. The image retrieval method that combines multiple features can retrieve all images that are similar in each feature, expand the user's needs to different features, and make a unified ranking of the similarity measurement results calculated by these features. The results are returned to the user in order of magnitude of the similarity measure. The present invention does not adopt a unified similarity measurement method, but adopts different measurement methods for different structural modes of different feature vectors. Different features describe images with different emphases. The present invention adopts a weighted combination method to obtain an optimal weight combination through a particle cluster algorithm to obtain an optimal similarity measurement method.

Description of drawings

Fig. 1 is the framework flow chart of the specific embodiment of the present invention;

Figure 2 is a schematic diagram of the HSV color space structure;

Figure 3 is a schematic diagram of Pseudo-Zernike Moment shape feature extraction;

Fig. 4 is a schematic diagram of Gabor texture feature extraction;

Figure 5 is a schematic diagram of the processing results, wherein (5-a) is an African inhabitant, (5-b) is a beach, and (5-c) is a building.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

The color image retrieval method optimized based on the particle cluster algorithm of the present invention normalizes various underlying features of a single image in the target database into image information descriptors of a specific length, and saves the image features in the feature database to establish an index surface. In the process of the user performing the image search, various single-low-level features are extracted from the image used as the search term to form a query vector. As shown in Figure 1, the similarity matching is performed by querying the feature vector and the feature vector in the feature database, and finally the retrieved image results are sorted and output in a unified manner. Due to the different emphases of the features described by different features, in order to avoid problems such as low accuracy of retrieval results caused by factors such as similarity calculation and incomplete matching of a single feature, content-based image retrieval uses multiple feature retrieval capabilities. Covering more matching results improves the recall rate, and using the normalization of the similarity measure increases the accuracy of the matching results during retrieval, that is, the precision rate.

The specific implementation steps of the present invention are as follows:

(1) Color feature extraction.

In this specific implementation, the HSV color space is used as a representation model of image color features. First convert the original image to HSV color space image and then perform color feature extraction.

Referring to Figure 2, the HSV color space describes the color of objects in terms of Hue (Hue, H for short), Saturation (S), and Value (V for short). This color space is more in line with the traditional RGB color space. The visual effects of human observation of things. In order to more effectively represent the color feature information of an image, the present invention adopts a color histogram (ColorHistogram) and a color moment (Color moments) to model the color features respectively.

The system of (8,2,2) is used to quantize the HSV color space, that is, the H component is evenly divided into 8 channels to represent, and S and V are quantized with 2 and 2 channels respectively, so that a histogram of 8х2х2=32 dimensions is obtained. vector.

其中p_ij是图片的j个像素点在第i个信道上的值，N是图片像素的总量。HSV是三维颜色空间，通过分别计算三个颜色空间的颜色矩，可得到6维的色矩向量。Color moment is a simple and effective color feature representation method, that is, the first-order moment μ _i (mean, mean) and the second-order moment δ _i (variance, viarance) are used to represent the color distribution of the image, which is defined as:

where p _ij is the value of the j pixels of the picture on the ith channel, and N is the total number of picture pixels. HSV is a three-dimensional color space. By calculating the color moments of the three color spaces respectively, a 6-dimensional color moment vector can be obtained.

By combining the color histogram and the color moment, a (32+6)=38-dimensional combined color feature descriptor is obtained.

(2) Extraction of texture features.

其中

x,y是图片中像素点的坐标位置，λ，θ分别代表波长和滤波的偏移方向，σ表示Gabor核函数中高斯函数的标准差，该参数决定了Gabor滤波核可接受区域的大小，σ的取值与b(Bandwidth，高低频率之差)和λ有关；γ表示纵横比，即空间纵横比，表示Gabor滤波器的椭圆度；ψ表示Gabor核函数中余弦函数的相位参数，有效值为-180度～180度，0度和180度对应的方程与原点对称，-90度和90度的方程分别与原点成中心对称。本具体实施方式中，采用0.1，0.8，2，5，11这5个参数作为滤波器的波长，0,π/4,π/2,3π/4四个参数作为Gabor滤波的偏移量。然后通过他们的组合的滤波对图像进行纹理特征提取。最后利用对滤波的结果求均值和标准差来获得纹理特征的向量：In this specific embodiment, gabor filtering is used to extract texture features of an image. Since the 2-dimensional Gabor filter function is a complex function, the computational complexity will be increased in practical application. In order to simplify the calculation while preserving the feature extraction properties of the Gabor filter function. The present invention uses the Gabor of the real part to filter, which is defined as follows:

in

x and y are the coordinate positions of the pixels in the image, λ and θ represent the wavelength and the offset direction of the filter, respectively, and σ represents the standard deviation of the Gaussian function in the Gabor kernel function. This parameter determines the size of the acceptable area of the Gabor filter kernel. The value of σ is related to b (Bandwidth, the difference between high and low frequencies) and λ; γ represents the aspect ratio, that is, the spatial aspect ratio, which represents the ellipticity of the Gabor filter; ψ represents the phase parameter of the cosine function in the Gabor kernel function, the effective value For -180 degrees to 180 degrees, the equations corresponding to 0 degrees and 180 degrees are symmetric to the origin, and the equations of -90 degrees and 90 degrees are centrosymmetric to the origin, respectively. In this specific embodiment, five parameters of 0.1, 0.8, 2, 5, and 11 are used as the wavelength of the filter, and four parameters of 0, π/4, π/2, and 3π/4 are used as the offset of the Gabor filter. The images are then subjected to texture feature extraction through their combined filtering. Finally, use the mean and standard deviation of the filtering results to obtain the vector of texture features:

where m=0,1,...,M-1; n=0,1,...,N-1. M, N are the number of wavelengths and offset parameters representing Gabor filtering, respectively, and P×Q represents the original size of the picture. E _mn (x, y) is the texture feature value processed by Gabor filtering at the coordinate point (x, y). Through the above processing, a 5×4×2=40-dimensional texture feature vector can be obtained: F _T ={μ ₀₀ ,σ ₀₀ ,μ ₀₁ ,σ ₀₁ ,...,μ _M-1N-1 ,σ _{M-1N- 1} }, as shown in Figure 4.

(3) Extraction of object shape features.

Image shape descriptors can be divided into two categories: one is the set of pixels that describe the boundary contour of the target area of the shape, called contour-based shape descriptors; the other is the set of all pixels in the target area that describes the shape, called Region-based shape descriptors. This specific implementation adopts pseudo-ZernikeMoment as the image shape feature extraction method. Pseudo-Zernick Moments is an object shape feature extraction algorithm based on region shape, which can represent images in multiple levels and has strong anti-noise ability.

其中f(x,y)是用来模拟图像的一种函数表示，(x,y)代表图片像素中的坐标位置。Pseudo-Zernike moment通过利用多项式{V_nm(x,y)}集合与极坐标变换，将f(x,y)映射到x²+y²≤1：In a digital image of size M×N, order n and repetition degree m, the Pseudo-Zernike Moment is defined as:

Where f(x, y) is a functional representation used to simulate an image, and (x, y) represents the coordinate position in the image pixel. The Pseudo-Zernike moment maps f(x,y) to x ² +y ² ≤1 by exploiting the set of polynomials {V _nm (x,y)} and polar transformations:

V _nm (x,y)=V _nm (ρ,θ)=R _nm (ρ)exp(jmθ),

θ＝arctan(y/x)。矩的阶数n是一个非负整数，重复度m的取值为|m|≤n，j。in,

θ=arctan(y/x). The order n of the moment is a non-negative integer, and the value of the repetition degree m is |m|≤n, j.

作为图像纹理特征的形状向量，最终获得21维的图像形状特征向量：The present invention uses the order of 5

As the shape vector of the image texture feature, the 21-dimensional image shape feature vector is finally obtained:

(4) Similarity measure.

The present invention uses three different image feature descriptors. In order to compare the similarity of the relevant image feature vectors with the images in the target image library. The present invention adopts different similarity measurement methods for different image features.

Color feature similarity measure:

Texture moment similarity measure:

Shape feature similarity measure:

分别代表图片Q，I的颜色特征分量；

和

分别代表Q与I图片的纹理特征矩中的均值向量和标准差向量；

代表图片在阶数为5时的Pseudo-Zernike Moment值。In the above similarity measurement formula, Q and I represent two pictures for similarity measurement. in

Represent the color feature components of pictures Q and I, respectively;

and

Represent the mean vector and standard deviation vector in the texture feature moments of the Q and I pictures, respectively;

Represents the Pseudo-Zernike Moment value of the image at order 5.

Different feature descriptors have different focus on the description of pictures. For example, if the objects to be retrieved are the blue sky and the ocean, they will pay more attention to the color descriptors of the images, but for pictures with items such as bananas or plates, they will be Pay more attention to the shape characteristics of the objects in the picture, and look for items with a sickle or round shape in the picture for comparison. In order to further improve the retrieval accuracy, the present invention uses the weighted combination of different feature similarity measurement results as the final similarity measurement value between the query image and the target image, which is expressed as: D(Q,I)=ω _c D _{color feature} +ω _t D _{texture moment} + ω _s D _shape , where ω _c , ω _t , ω _s represent the weights of the three feature components in the similarity measurement process. In order to balance the measurement results and simplify the calculation amount, let ω _c +ω _t +ω _s =1.

(5) Optimization of similarity measure using particle cluster algorithm.

Particle Swarm Optimization (PSO) is a common method used to find the optimal solution. Its velocity and particle position update formulas are as follows:

V[i]=ω*V[i]+c ₁ *rand()*(pbest[i]-present[i])+c ₂ *rand()*(gbest-present[i])present[i+ 1]=present[i]+V[i]

This functional formula is an equation for obtaining the optimal solution by moving the particle to the optimal solution, where V[i] represents the velocity of the ith particle, w represents the inertia weight, c1 and c2 represent the learning parameters, and rand() represents A random number between 0-1, pbest[i] represents the optimal value searched by the ith particle (local optimal solution), gbest represents the optimal value searched by the entire cluster (global optimal solution), present [i] represents the current position of the ith particle.

In this specific embodiment, Corel-10 is used as the experimental image database. Corel-10 is an open source image library with 10,000 images, including 10 categories, with tag names like Africa, Beaches, Mountains, Cars, Dinosaurs, Elephants, Flowers, Horses, Food, and Architecture. There are 1000 pictures for each category.

其中，0≤ω_c,ω_t,ω_s≤1。And select 3 weights in the similarity measurement process as moving particles. In order to make the weight size not produce out-of-bounds errors, let:

Among them, 0≤ω _c ,ω _t ,ω _s ≤1.

For image retrieval based on multi-feature fusion, different feature descriptors describe the same image with different emphasis. By optimizing the combination of different features, the advantages of different feature descriptors can be well utilized to optimize query results. The present invention solves this problem by using query result feedback information and PSO algorithm optimization, that is, by analyzing the results returned by the test query to guide the adjustment of the weights. Specifically, the present invention randomly selects 20 different pictures from each category of the test database (Corel database) as query pictures, and the top K (preset value) pictures returned by the CBIR system each time are used to calculate the retrieval accuracy . The retrieval accuracy is: the ratio of the number of images matching the query image in the returned K images to the value K. The present invention uses the retrieval accuracy as the adaptive function, and the higher the value, the better the value, and the closer the position of the particle is to the position of the local optimal solution.

Among them, the specific process of training the weights ω _c , ω _t , ω _s is as follows:

First, N images are randomly selected from each category in the Corel database as training data. The number of particles n=k (k is a preset value, set based on the database size), the initial particle position present[i] is generated by a random number, and 0≤present[i]≤1, for example, set it to 0.01;

其中i为粒子标识符，

表示当前粒子i对应权值ω_c、ω_t、ω_s的值；Then initialize the values of the local optimal position gbest and the global optimal position pbest[i] of the particle, and specify the initial position of any particle motion through random numbers

where i is the particle identifier,

Indicates the value of the weights ω _c , ω _t and ω _s corresponding to the current particle i;

计算查询参数为

时的检索精度，其中I_i表示查询图片的特征描述向量(颜色、纹理和形状等)；Perform a query operation on each particle

Calculate the query parameter as

Retrieval accuracy at time, where I _i represents the feature description vector (color, texture and shape, etc.) of the query image;

的值更新gbest和pbest[i]的值；According to the current highest retrieval accuracy corresponding to

update the value of gbest and pbest[i];

Determine whether the iterative convergence conditions are met, if not, continue the query operation, and update gbest and pbest[i] based on the retrieval progress; otherwise, output the current

In the PSO algorithm, this specific embodiment sets the convergence condition of the clustering algorithm as the query accuracy of querying the same image twice is less than 0.01 or the algorithm loops more than 1000 times, that is, when the deviation of the most recent two query accuracy is less than The update is stopped when the threshold (0.01) or the maximum number of iterations is reached, and the weight of the optimal combination is obtained.

That is, when the present invention performs real-time query processing, based on the currently input image to be queried, various underlying features (color histogram, color moment, texture moment, shape) of the query object are obtained, and the feature description vector of the image to be queried is obtained, Then extract vectors according to different features, use different feature measurement methods, and compare the similarity with the feature library of the image library to obtain the similarity measurement results between pictures corresponding to different features, and based on the weight of the optimal combination after training value, get the final similarity, and finally sort the obtained final similarity (such as the similarity result in increasing order and output), and return the top K (eg 35) most similar K pictures to the user, as shown in Figure 5 .

The invention integrates multiple feature vectors as feature descriptors of images, and assigns a similarity measure formula suitable for its structure to each feature vector according to the characteristics of each different feature descriptor. Since the pictures to be queried by users are unknown, different pictures have different requirements for the type of feature descriptors. A single feature cannot meet all the requirements of image retrieval. By fusing multiple feature descriptions, images that are similar in different features can be retrieved at the same time. At the same time, due to the semantic gap between the low-level feature description and the high-level semantics, different images may have the same feature descriptor in a certain aspect, and the error caused by this problem can be effectively overcome by fusing multiple features. At the same time, in the process of calculating the feature similarity, the present invention uses weighting to combine the measurement results of each component feature. The optimal weights of each similarity component are trained by the particle cluster algorithm. Thus, errors caused by incomplete similarity calculation and single feature matching description are avoided.

The present invention relates to content-based image retrieval, and mainly considers the optimization of CBIR technology. By using the PSO algorithm to optimize the multi-feature fusion CBIR system, the accuracy of image retrieval can be effectively improved, which has certain practical value.

The above descriptions are only specific embodiments of the present invention, and any feature disclosed in this specification, unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes; all the disclosed features, or All steps in a method or process, except mutually exclusive features and/or steps, may be combined in any way.

Claims (4)

1. a color image retrieval method optimized based on particle cluster algorithm, is characterized in that, comprises the following steps: 1. Train the similarity measure weight based on the particle cluster algorithm: 101: Use different types of query images as training samples, and extract color feature vectors, texture feature vectors, and object shape feature vectors of the training samples; The color feature vector is: the color histogram and the color moment feature of the image in the HSV color space; The texture feature vector is: performing Gabor filtering on the image using the Gabor of the real part, and extracting the mean and standard deviation of the filtering output; The object shape feature vector is: the object shape feature extraction method based on the region shape, extracts the object shape feature of the image; 102: Extract the color feature vector, texture feature vector and object shape feature vector of each retrieved image in the image library; 103: Use the particle cluster algorithm to optimize the weights of the similarity measure of the three types of image feature vectors: Initialize the weight ω _c of the similarity measure D _c of the color feature vector, the weight ω _t of the similarity measure D _t of the texture feature vector, the weight ω _s of the similarity measure D _s of the object shape feature vector, and define each The particle position of each particle is (ω _c , ω _t , ω _s ), where 0≤ω _c ,ω _t ,ω _s ≤1; the initialized particle number k, and the local optimal position of each particle and the global particle swarm optimal location; Perform image query processing with n training samples of the same type as query images each time: Calculate the similarity measure D _c of the color feature vector and the texture feature vector between each training sample and each retrieved image in the image database. The similarity measure D _t and the similarity measure D _s of the object shape feature vector, and based on the weights ω _c , ω _t , ω _s , the three similarity measures are weighted and summed to obtain the total similarity measure D; The retrieved images corresponding to the maximum total similarity measure D are taken as the retrieval result; Iteratively update the particle position, local optimal position and global optimal position based on the current retrieval accuracy until the iteration converges, and save the latest updated particle position; The retrieval accuracy is: the ratio of the number of retrieved images matching the query image in the retrieval result to the value K; The calculation methods of the similarity measure D _c of the color feature vector, the similarity measure D _t of the texture feature vector and the similarity measure D _s of the object shape feature vector are: It is defined that Q and I represent two images whose similarity is measured, then the similarity measures D _c , D _t and D _s between the images Q and I are specifically:

Represent the color feature components of images Q and I, respectively, and N _c represents the dimension of the color feature vector;

and

represent the mean vector and standard deviation vector in the texture feature moments of images Q and I, respectively, and N _t represents the dimension of the texture feature vector;

represent the components of the object shape feature vector of images Q and I respectively, M _s represents the order of the object shape feature vector; 2. Image retrieval: Extract the color feature vector, texture feature vector and object shape feature vector of the current query image; and calculate the similarity measure D _c of the color feature vector, the similarity measure D _t of the texture feature vector and the similarity measure D _s of the object shape feature vector with each retrieved image in the image library respectively; Based on the weights ω _c , ω _t and ω _s obtained by training, the current total similarity measure D=ω _c D _c +ω _t D _t +ω _s D _s is obtained; The retrieved images corresponding to the top K largest total similarity measures D are used as retrieval results and output. 2. method as claimed in claim 1, is characterized in that, the color histogram and color moment feature of image in HSV color space are specially: The system of (8, 2, 2) is used to quantize the HSV color space, that is, the H component is equally divided into 8-channel representation, and the S and V are equally divided into 2-channel representation; Extract the color histogram feature of each channel after quantization; The first-order moment μ _{i and the second-order moment δ i of the H, S, and V components are calculated respectively according to the formula, and the first-order moment μ i and the} second-order moment δ _i form the color moment feature of each component, where

where p _ij represents the pixel value of the j-th pixel of the image on the i-th channel, and N represents the total number of pixels in the image. 3 . The method of claim 1 , wherein the wavelengths of the Gabor filter are 0.1, 0.8, 2, 5, and 11, and the offsets are 0, π/4, π/2, and 3π/4. 4 . 4. The method according to claim 1, wherein the method for extracting the shape feature of the object is a fifth-order Pseudo-Zernick Moments algorithm.

Images