CN107180439A - A kind of colour cast feature extraction and colour cast detection method based on Lab chrominance spaces - Google Patents

Abstract

The invention discloses a color shift feature extraction and color shift detection method based on Lab chromaticity space. The color shift feature extraction method includes: 1) defining the color shift feature h in the height direction and defining the change rate of the color shift feature in the NNO region , define the color shift feature of the brightness channel, 2) extract the color shift feature h of the color shift feature h and the color shift feature of the brightness channel; the color shift detection method includes: 1) calculate the equivalent circle feature u and σ; 2) Execute the "preliminary color shift detection"process; 3) execute the "no color shift image re-detection" process, etc. The present invention more comprehensively considers the peak distribution characteristics of the color map histogram in the height direction, and further considers the change rule of the color shift feature in the original image and the NNO area and the aggregation and distribution characteristics of the color shift feature in the brightness channel, making up for the The deficiencies of existing methods in the definition and extraction of color shift features improve the accuracy of color shift detection.

Description

A color shift feature extraction and color shift detection method based on Lab chromaticity space

technical field

The invention relates to the technical field of color shift detection of color images, in particular to a color shift feature extraction and color shift detection method.

Background technique

When humans observe the world, the visual nerve's response to color, shape, surface texture and detailed features gradually weakens. Moreover, compared with text and sound, color images have richer connotations and stronger expressive capabilities. Therefore, color is the most important characteristic clue to obtain objective world information, and it is also a key factor to directly measure the imaging quality of equipment. However, when the ambient light source of the scene, the reflection characteristics of the object itself, or the light sensitivity coefficient of the acquisition device change, the color value of the object in the image recorded by the imaging device deviates from its own color, and the entire image produces a color shift phenomenon. This not only affects the visual effect of the image, but also affects subsequent processing such as image segmentation and target recognition. Therefore, color shift detection and color shift correction are often an indispensable link in the field of color image processing and machine vision.

The most classic color shift detection is a class of methods based on color constancy. The general process of this type of method is to first estimate the illumination of the scene, and then convert the input image taken under unknown illumination to standard illumination through the von Kries model, so as to achieve the purpose of color cast detection and correction. Among them, the White Patch algorithm assumes that there is always a white surface in the scene, and uses the maximum value of the RGB color channel in the image as the estimated light source. The Gray World algorithm assumes that the color information in the scene is rich enough and the reflection of all physical surfaces is color-free (ie gray), and the estimated light source is obtained by averaging the RGB color channels in the image. The Shades of Gray algorithm assumes that the entire scene is still color-free after a nonlinear reversible transformation, and uses the Minkowski-norm distance instead of simple averaging, ignoring the local correlation between pixels. The color constancy algorithm based on Bayesian inference estimates the image illumination from the posterior probability of the image color distribution by establishing a model between surface reflectance and image illumination. The color constancy algorithm based on neural network or SVM uses multi-layer neural network or SVR to establish a mapping model between scene illumination distribution and image color distribution, and predicts the illumination of a new input image according to the model. However, such color cast detection methods based on color constancy are usually only suitable for application environments that satisfy specific scene assumptions. If the assumption conditions cannot be met, the algorithm adaptability is difficult to achieve the desired effect.

In recent years, considering that there is a certain vector correlation between the RGB color channels of color images, a class of color shift detection algorithms based on Lab chromaticity histograms has been widely used. A complete description of such algorithms is given by F. Gasparini et al. First, the brightness component of the L channel and the chrominance component of the ab channel are obtained through the conversion of RGB to Lab color space. Then, using the method based on the equivalent circle and NNO area, the distribution characteristics of the chromaticity histogram are counted, and the quantitative calculation of the color shift feature and the detection of the color shift are realized. Similarly, Li Feng and Chen et al. used the two-dimensional chromaticity distribution characteristics of the Lab color space to judge color shift and non-color shift images, and proposed a Gaussian mixture clustering model to distinguish essential color shift from real color shift. Compared with the color shift detection method based on color constancy, this method does not require scene assumptions, has better scene adaptability, and can distinguish between essential color shift and real color shift.

However, in the extraction of color shift features based on equivalent circles, only the overall projection position information of the chromaticity histogram on the ab plane is considered, which makes such methods still rough in the definition and extraction of color shift features And insufficient, thus directly lead to the existing Lab color shift detection algorithm still has low color shift detection accuracy.

Contents of the invention

In view of this, the purpose of the present invention is to provide a color shift feature extraction and color shift detection method based on Lab chromaticity space, which more comprehensively considers the peak distribution characteristics of the color map histogram in the height direction, and further considers To solve the problem that the existing color shift detection algorithm based on the Lab chromaticity histogram is still in the definition and extraction of color shift features. There are deficiencies, which directly lead to the technical problem of low accuracy of color shift detection.

The present invention is based on the color shift feature extraction method of Lab chromaticity space, comprises the following steps:

1) Define the color cast feature

1) Define the two-dimensional chromaticity histogram distribution function of channel a and b as:

H=F(a,b)

Among them, a, b ∈ [-128, 127] are the chromaticity values of a and b color channels respectively, H is the number of pixels corresponding to the chromaticity value (a, b) in the image, and the distribution function F( ) is A three-dimensional spatial distribution function;

2) Define the color shift feature h of the chromaticity histogram in the height direction as:

in, and h _σ represent the mean and variance of the chromaticity histogram in the height direction, respectively;

3) Define the change rate of the color shift feature in the NNO area:

Among them, u _NNO and σ _NNO are the equivalent circle features of the chromaticity histogram of the NNO area on the ab plane, h _NNO is the color shift feature of the chromaticity histogram of the NNO area in the height direction, then u _cr , σ _cr and h _cr represents the relative change rate of the equivalent circle feature and height feature of the chromaticity histogram between the original image and the NNO area image;

4) The brightness histogram envelope fitting algorithm based on the Gaussian distribution function, defining the fitting degree R between the fitting curve and the envelope curve, and the half-width c of the fitting function as the color shift feature of the brightness channel,

The Gaussian distribution function is:

Among them, a, b, and c represent the peak size, peak center position, and half-width information of the Gaussian distribution, respectively, and x and y represent independent variables and function values, respectively;

2) Extracting color shift features

1) Extracting the color shift feature h, comprising the following steps:

a) Normalize color images of different sizes into original images I ₁ (i,j) of the same size to be tested;

b) Calculate the chromaticity histogram height matrix H ₀ of the normalized image:

Among them, n=256 represents the chromaticity level of a, b channel, h _ij =F(i-129, j-129) represents that in the two-dimensional chromaticity histogram distribution function corresponds to the chromaticity component a=i-129, b = the number of pixels of j-129;

c ₎ Filter H0 to eliminate the influence of image noise and elements with smaller peaks in the chroma histogram:

H ₁ =H ₀ ,if h _ij <T ₁ M ₁ then h _ij =0

Among them, H ₁ is the height matrix of the filtered chroma histogram, T ₁ is the filtering threshold, and M ₁ is the mean value of the height matrix H ₀ :

d) Calculate the mean value of the chromaticity histogram in the height direction and variance h _σ :

Among them, p is the number of 8-neighborhood connected regions in H ₁ , Ω _c , c=1,...,p represents the cth local connected region, S _c is the connected area of this region, is the local height mean of the area;

2) Extracting the color shift feature of the brightness channel, comprising the following steps:

a) Calculate the L-channel histogram matrix N ₀ of the normalized image,

N ₀ =[n ₀ n ₁ ... n ₁₀₀ ]

Among them, n _i =f(i), i=0～100 represents the number of pixels corresponding to the brightness L=i, f(·) is the brightness histogram function;

b) Perform filtering operation on the brightness histogram N ₀ to eliminate the influence of image noise and elements with smaller peak values in the brightness histogram, and obtain the filtered L component histogram matrix N ₁ :

N ₁ =N ₀ ,if n _i <T ₂ S ₂ then n _i =0

Wherein, S ₂ =max(n _i ) is the maximum element in the brightness histogram matrix N ₀ , and T ₂ is the filtering threshold;

c) Using the Gaussian distribution function in step 4), based on the least squares algorithm, perform curve fitting on the envelope data (x=i, y=n _i ) of the filtered brightness histogram N ₁ , i=0～100, Get the fitted L component histogram matrix N ₂ :

in, Indicates the fitting result corresponding to the brightness L=i, g ( ) is the Gaussian distribution function described in step 4), and its parameters a, b and c have all been obtained by the fitting algorithm at this moment;

d) According to the filtered and fitted L component histogram matrices N ₁ and N ₂ , use the coefficient of determination to calculate the fitting degree R of the two:

Wherein, i=0～100 represents the brightness level of the L channel, and _n represents the sample value corresponding to the brightness L ₌ i in the histogram matrix N after filtering, Indicates the fitting value corresponding to the brightness L=i in the histogram matrix _N2 after fitting, is the mean value of all elements in N ₁ , R∈[0,1] represents the joint approximation degree of the fitting value to the sample value, and the closer R is to 1, the higher the fitting accuracy is.

The present invention also discloses a color shift detection method based on Lab chromaticity space, comprising the following steps:

step 1):

According to formula (1): Formula (2): C=(u _a ,u _b ), and formula (3): Calculate the equivalent circle features u and σ of the chromaticity histogram on the ab plane; F(a, b) in the formula (1) is the chromaticity histogram distribution function, k=a, b represents on the a axis or the b axis Integrate, u _k and σ _k are the mean and variance of the chromaticity histogram on the k-axis respectively; C in formula (2) is the center of the equivalent circle, and σ is the radius of the equivalent circle;

Step 2):

Execute the "preliminary color shift detection" process: if the conditions of formula (4): (D>10and D _σ >0.6)or(D _σ >1.5) are met, the image is classified as "color shift" and go to step 5); otherwise, The image is classified as "no color cast", go to step 3);

Step 3):

Execute the process of "image re-detection without color shift": according to the formula (6): Criterion, images are divided into three categories: "color shift", "no color shift" and "unrecognizable"; for "unrecognizable" images, go to step 4); for "color shift" images, go to step 5).

Step 4):

Execute the "unrecognizable image re-detection" process:

1) according to the feature extraction algorithm described in claim 1, calculate the color shift feature h of the chromaticity histogram in the height direction;

2) according to the feature extraction algorithm described in claim 1, calculate the characteristic rate of change u _cr , σ _cr and h _cr of the chroma channel in the NNO region;

3) Use the following criteria to retest the unrecognizable image:

Step 5):

Execute the "color shift image re-detection" process:

1) according to the feature extraction algorithm described in claim 1, calculate the color cast feature R and c of brightness channel;

2) Use the following criteria to retest the color cast image:

Beneficial effects of the present invention:

The present invention is based on the color shift feature extraction and color shift detection method in Lab chromaticity space, and on the basis of the classic Lab color shift detection algorithm, more comprehensively considers the peak distribution characteristics of the color map histogram in the height direction, and further considers The change law of color shift features in the original image and NNO area and the aggregation and distribution characteristics of color shift features in the brightness channel make up for the deficiencies of existing methods in the definition and extraction of color shift features, and improve the accuracy of color shift detection.

Description of drawings

FIG. 1 is a flowchart of a color shift detection method based on Lab chromaticity space in an embodiment.

detailed description

The present invention will be further described below.

The core idea of the color shift detection algorithm based on the Lab chromaticity histogram is to examine the vector correlation between colors from the distribution characteristics of image color information. In order to quantitatively analyze the distribution characteristics of the chromaticity histogram, the classic Lab color shift detection algorithm realizes color shift feature extraction and color shift detection by introducing equivalent circles and NNO regions. The classic Lab color shift detection algorithm flow is as follows:

Step 1: Convert the color image to be detected from RGB to Lab color space to obtain image brightness component L and chrominance components a, b.

Step 2: Perform statistics on the histogram of the chroma component, and the following theoretical basis can be found according to its distribution characteristics: the chroma histogram of an image without color shift should appear as multiple discrete peaks, and most of the peaks should be distributed in the neutral Around the point (a=0, b=0); while the chromaticity histogram of the image with color shift appears to contain zero or one concentrated peak, and the peak deviates from the neutral point.

Step 3: In order to quantitatively describe the degree of concentration of the chromaticity histogram and the distance relationship between the peak distribution and the neutral point, an equivalent circle is introduced to calculate the characteristics of the chromaticity histogram.

Among them, F(a, b) is the chromaticity histogram distribution function, k=a, b means to integrate on the a-axis or b-axis, u _k and σ _k are the mean value and variance. Then according to formula (1), calculate the center C and radius σ of the equivalent circle:

C＝(u _a , u _b ),

On this basis, the following equivalent circular color shift features are defined:

Among them, u is the distance from the center of the circle to the neutral point (a=0, b=0), D is the distance from the outside of the equivalent circle to the neutral point, and D _σ indicates the degree to which the equivalent circle deviates from the neutral point. The larger the D _σ value, the more serious the chromaticity histogram of the image deviates from the neutral point or the stronger the peak aggregation, that is, the more serious the color shift.

Step 4: Analyze the equivalent circle feature of the chromaticity histogram, when the following conditions are met,

(D＞10and _Dσ ＞0.6)or( _Dσ ＞1.5) (4)

It is considered that the chromaticity histogram deviates from the neutral point and the peaks are gathered, and it is initially classified as a "color cast image"; otherwise, it is initially classified as a "no color cast image".

Step 5: Further consider the following theoretical basis: the colorless surface in the image scene (the gray surface under standard white light, that is, the neutral gray area of the color image) can completely reflect the color of the incident light in the scene, so by analyzing the NNO area in the scene The analysis can accurately estimate the illumination offset of the image. The method of obtaining the NNO area of the image is as follows:

and d is not isolated (5)

Among them, L, a, b are the three components of the image in the Lab color space, and d is the maximum value of the chromaticity radius. In order to prevent noise interference, the pixel point where d is located and the pixel point I _NNO (i, j ) is a non-isolated point.

Step 6: For the image initially classified as "no color shift" in step 4, extract its NNO region image for re-detection, and use the same method as formula (1) - formula (3) to solve the color shift feature D of its NNO region image _σNNO , and perform the following judgments:

Step 7: For the images classified as "color cast" in Step 4 and Step 6, use cluster learning algorithm to classify and identify: if the area of the image scene containing the main color of the ocean, blue sky, grass, etc. reaches the total area of the image More than 40%, it is classified as "essential color cast"; otherwise, it is considered as "true color cast".

Obviously, the accuracy of the Lab color shift detection algorithm depends on the vector correlation analysis and quantitative calculation of the distribution characteristics of the Lab color space. Therefore, the present invention mainly improves the color shift feature based on the Lab chromaticity histogram and its feature extraction algorithm, and proposes a complete improved color shift detection algorithm based on the new color shift feature and the classic Lab color shift detection algorithm.

In this embodiment, the color shift feature extraction method based on Lab chromaticity space includes the following steps:

1) Define the distribution function of the two-dimensional chromaticity histogram of channels a and b as:

H=F(a,b) (7)

Among them, a, b ∈ [-128, 127] are the chromaticity values of a and b color channels respectively, H is the number of pixels corresponding to the chromaticity value (a, b) in the image, and the distribution function F( ) is A three-dimensional spatial distribution function. In the classic Lab color shift detection algorithm, the equivalent circle is the only basis for extracting color shift features from the chromaticity histogram distribution, although the equivalent circle can better describe the overall position of the chromaticity histogram (center to neutral). However, in the process of mapping the three-dimensional spatial distribution projection of the color map histogram to the two-dimensional ab plane where the equivalent circle is located, the histogram is obviously lost Height information is an important feature clue.

In the research, it is found that the information in the height direction of the chromaticity histogram can also well reflect the degree of difference between the color shift image and the non-color shift image. The theoretical basis is: for an image without color shift, because the various chroma components in the image are relatively rich and present a more obvious dispersion, so the distribution of the main hue and other secondary hues in the height direction has a small difference; while for In the color cast image, because the chroma components show obvious aggregation, the distribution of the main hue and other secondary hues in the height direction is quite different.

2) In order to quantitatively describe and calculate the size distribution and degree of difference of the chromaticity histogram in the height direction, this embodiment defines the color shift feature h of the chromaticity histogram in the height direction as:

in, and h _σ denote the mean and variance of the chromaticity histogram in the height direction, respectively.

The specific algorithm for extracting the color shift feature h is described as follows:

Step 1: Normalize the color images of different sizes into the original image I ₁ (i,j) of the same size to be tested. Although the processed image may be distorted visually, the statistical characteristics of its color distribution are not No change.

Step 2: Calculate the chromaticity histogram height matrix H ₀ of the normalized image:

Among them, n=256 represents the chromaticity level of a, b channel, h _ij =F(i-129, j-129) represents that in the two-dimensional chromaticity histogram distribution function corresponds to the chromaticity component a=i-129, b =The number of pixels of j-129.

Step 3: Perform filtering operation on H ₀ to eliminate the influence of image noise and elements with smaller peaks in the chroma histogram:

H ₁ =H ₀ ,if h _ij <T ₁ M ₁ then h _ij =0 (10)

Among them, H ₁ is the height matrix of the filtered chroma histogram, T ₁ is the filtering threshold, and M ₁ is the mean value of the height matrix H ₀ :

Step 4: Calculate the mean value of the chromaticity histogram in the height direction and variance h _σ :

Among them, p is the number of 8-neighborhood connected regions in H ₁ , Ω _c , c=1,...,p represents the cth local connected region, S _c is the connected area of this region, is the local mean height of the region.

3) In the classic Lab color shift detection algorithm, only the parameter characteristics of the equivalent circle of the two-dimensional chromaticity histogram of the NNO region image are considered, and the quantitative analysis before and after the feature change is not combined with the relevant parameters of the original image. Therefore, in the process of re-detecting the NNO area initially judged as "no color shift image", misjudgment is easy to occur.

In our research, we found that the color shift characteristics of the chromaticity histogram in the equivalent circular plane and height direction can better reflect the color shift image and The degree of variance in non-color cast images. The theoretical basis is: for an image without color shift, the radius of the equivalent circle in the NNO area decreases greatly and is closer to the neutral point, and the characteristic change range of the chromaticity histogram in the NNO area in the height direction is small; For the color cast image, the radius of the equivalent circle in the NNO area decreases less and is farther away from the neutral point, and the characteristic change of the chromaticity histogram in the NNO area in the height direction is larger. In order to quantitatively describe and calculate the change rule of the color shift feature between the original image and the NNO area image, this embodiment further defines the following change rate of the color shift feature in the NNO area:

Among them, u _NNO and σ _NNO are the equivalent circle features of the chromaticity histogram of the NNO area on the ab plane, h _NNO is the color shift feature of the chromaticity histogram of the NNO area in the height direction, then u _cr , σ _cr and h _cr represents the relative rate of change between the original image and the NNO region image of the equivalent circle feature and height feature of the chromaticity histogram, respectively.

4) In the classic Lab color shift detection algorithm, only the distribution characteristics of the two-dimensional chroma components a and b are considered, and the channel characteristics of the luminance component L are not considered. In our research, we found that the L component in the Lab chromaticity space can better reflect the degree of difference between the essential color cast and the real color cast: for the essential color cast image, its L channel histogram presents a regional monotonous Peak aggregation distribution; for the real color cast image, its L channel histogram presents a relatively uniform discrete distribution.

Consider the following Gaussian distribution function:

Among them, a, b, and c represent the peak size, peak center position, and half-width information of the Gaussian distribution, respectively, and x and y represent independent variables and function values, respectively. This Gaussian distribution function also exhibits regional unimodal aggregation distribution characteristics, and the smaller the half-width c, the more aggregated the Gaussian distribution is.

In order to quantitatively describe and calculate the distribution characteristics of the L channel histogram, this embodiment proposes a Gaussian distribution-based brightness histogram envelope fitting algorithm, and defines the fitting degree R between the fitting curve and the envelope curve, and the fitting The half-width c of the function is used as the color shift characteristic of the luma channel. Obviously, the higher the fitting degree (the larger R) and the better the regional aggregation (the smaller c), the more obvious the unimodal aggregation of the L channel histogram, and the more likely the corresponding color image belongs to the essential color cast.

The specific algorithm for extracting the color cast feature of the brightness channel is described as follows:

Step 1: Calculate the L-channel histogram matrix N ₀ of the normalized image,

N ₀ =[n ₀ n ₁ ... n ₁₀₀ ] (15)

Wherein, n _i =f(i), i=0-100 represents the number of pixels corresponding to the brightness L=i, and f(·) is a brightness histogram function.

Step 2: Perform filtering operation on the brightness histogram N ₀ to eliminate the influence of image noise and elements with smaller peaks in the brightness histogram, and obtain the filtered L component histogram matrix N ₁ :

N ₁ =N ₀ ,if n _i <T ₂ S ₂ then n _i =0 (16)

Wherein, S ₂ =max(n _i ) is the maximum element in the luminance histogram matrix N ₀ , and T ₂ is the filtering threshold, and T ₂ =0.2 in this embodiment.

Step 3: Using the Gaussian distribution function described in formula (14), based on the least squares algorithm, curve the envelope data (x=i, y=n _i ) of the filtered brightness histogram N ₁ , i=0～100 Fitting to get the fitted L component histogram matrix N ₂ :

in, Indicates the fitting result corresponding to the brightness L=i, g(·) is the Gaussian distribution function described by formula (14), and its parameters a, b and c have been obtained by the fitting algorithm at this time.

Step 4: According to the filtered and fitted L component histogram matrices N ₁ and N ₂ , use the coefficient of determination to calculate the fitting degree R of the two:

Wherein, i=0～100 represents the brightness level of the L channel, and _n represents the sample value corresponding to the brightness L ₌ i in the histogram matrix N after filtering, Indicates the fitting value corresponding to the brightness L=i in the histogram matrix _N2 after fitting, is the mean value of all elements in N ₁ , R∈[0,1] represents the joint approximation degree of the fitting value to the sample value, and the closer R is to 1, the higher the fitting accuracy is.

The color shift detection method based on the Lab chromaticity space in this embodiment includes the following steps:

Step 1: Calculate the equivalent circle features u and σ of the chromaticity histogram on the ab plane according to formulas (1)-(3).

Step 2: Execute the "preliminary color shift detection" process: if the condition of formula (4) is met, the image is classified as "color shift" and go to step 5; otherwise, the image is classified as "no color shift" and go to step 3.

Step 3: Execute the process of "re-detection of images without color shift": according to the criterion of formula (6), images are divided into three categories: "color shift", "no color shift" and "unrecognizable"; for "unrecognizable" images, Go to step 4; for the "color cast" image, go to step 5.

Step 4: Execute the "unrecognizable image re-detection" process:

1) According to the extraction algorithm of the aforementioned color shift feature h, calculate the color shift feature h of the chromaticity histogram in the height direction;

2) According to the feature extraction algorithm of the color shift feature of the aforementioned luma channel, calculate the feature change rates u _cr , σ _cr and h _cr of the chroma channel in the NNO region;

3) Use the following criteria to retest the unrecognizable image:

Step 5: Execute the "color shift image re-detection" process:

1) According to the feature extraction algorithm of the color shift feature of the aforementioned brightness channel, calculate the color shift features R and c of the brightness channel;

2) Use the following criteria to retest the color cast image:

In this embodiment, in order to verify the effectiveness of the color shift feature proposed by the present invention, a test data set consisting of 480 color images is constructed. Among them, the numbers of color cast and non-color cast images are 191 and 289, respectively. In this experiment, the color shift feature extraction method based on the Lab chromaticity space in this embodiment is firstly used to obtain improved color shift features. Then, use the color shift detection method based on the Lab chromaticity space in this embodiment to detect the color shift. For the unrecognized images in step 6) of the classic Lab color shift detection algorithm, do the following processing to obtain higher accuracy:

As can be seen from the detection results listed in Table 1, the color shift feature and feature extraction method proposed by the present invention are all feasible; in addition, compared with the classic color shift detection algorithm based on Lab, the feature and detection method proposed by the present invention with higher precision.

Table 1. Detection results of all test images

Finally, it is noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be carried out Modifications or equivalent replacements without departing from the spirit and scope of the technical solution of the present invention shall be covered by the claims of the present invention.

Claims (2)

1. A color cast characteristic extraction method based on Lab chromaticity space is characterized by comprising the following steps: the method comprises the following steps:

one) defining color shift characteristics

1) The two-dimensional chromaticity histogram distribution function defining the channels a and b is as follows:

H＝F(a,b)

wherein, a, b ∈ [ -128,127] is the chroma value of a, b color channel respectively, H is the number of pixel points corresponding to chroma value (a, b) in the image, the distribution function F (-) is a three-dimensional space distribution function;

2) defining the color cast characteristic h of the chromaticity histogram in the height direction as follows:

wherein,and h_σRespectively representing the mean value and the variance of the chroma histogram in the height direction;

3) defining the color cast characteristic change rate of the NNO area:

wherein u is_NNOAnd σ_NNOFor the circle-equivalent feature of the NNO region chromaticity histogram on the ab plane, h_NNOIs the color shift characteristic of the NNO area chromaticity histogram in the height direction, then u_cr、σ_crAnd h_crRespectively representing the relative change rate of the equivalent circle feature and the height feature of the chromaticity histogram between the original image and the NNO area image;

4) defining the fitting degree R of a fitting curve and an envelope curve and the half width c of the fitting function as the color cast characteristic of a brightness channel by a brightness histogram envelope fitting algorithm based on a Gaussian distribution function,

the gaussian distribution function is:

wherein, a, b and c respectively represent the peak value size, the peak value center position and the half width information of Gaussian distribution, and x and y respectively represent independent variables and function values;

II) extracting color cast characteristics

1) Extracting the color cast characteristic h, comprising the following steps:

a) standardizing color images with different sizes into to-be-detected original images with the same sizeStarting picture I₁(i,j)；

b) Computing a chroma histogram height matrix H for the normalized image₀：

Where n-256 denotes the chroma level of the a and b channels, and h_ijF (i-129, j-129) represents the number of pixels corresponding to the chrominance component a-i-129 and b-j-129 in the two-dimensional chrominance histogram distribution function;

c) to H₀And performing filtering operation to eliminate the influence of image noise and small elements of peak values in the chrominance histogram:

H₁＝H₀,if h_ij＜T₁M₁then h_ij＝0

wherein H₁For the filtered chrominance histogram height matrix, T₁To filter the threshold, M₁Is a height matrix H₀Average value of (d):

d) calculating the mean value of the chrominance histogram in the height directionSum variance h_σ：

Wherein p is H₁The number of middle 8 neighborhood connected regions, omega_cWhere c is 1, …, p represents the c-th local connected domain, S_cIs the area of communication of the region,is the local height mean of the area;

2) extracting color cast characteristics of a brightness channel, comprising the following steps:

a) computing an L-channel histogram matrix N of a normalized image₀，

N₀＝[n₀n₁… n₁₀₀]

Wherein n is_iF (i), wherein i is 0-100 to represent the number of pixels corresponding to the brightness L is i, and f (·) is a brightness histogram function;

b) for luminance histogram N₀Filtering to eliminate the influence of image noise and small peak value elements in the brightness histogram and obtain the filtered L-component histogram matrix N₁：

N₁＝N₀,if n_i＜T₂S₂then n_i＝0

Wherein S is₂＝max(n_i) Is a luminance histogram matrix N₀Maximum element in (1), T₂Is a filtering threshold;

c) adopting the Gaussian distribution function in the step four), and based on a least square algorithm, carrying out on the filtered brightness histogram N₁Is (x ═ i, y ═ n), and is used as a reference signal_i) And (5) carrying out curve fitting on the i-0-100 to obtain a fitted L-component histogram matrix N₂：

Wherein,representing the fitting result corresponding to the brightness L ═ i, g (·) is the gaussian distribution function described in step four), whose parameters a, b, and c have been obtained by the fitting algorithm;

d) according to the L component histogram matrix N after filtering and fitting₁And N₂And calculating the fitting degree R of the two by adopting a decision coefficient:

wherein i is 0 to 100 and represents the brightness level of L channel, and n_iRepresenting the filtered histogram matrix N₁The sample value corresponding to the luminance L ═ i,representing the fitted histogram matrix N₂Corresponding to the fitted value of luminance L ═ i,is N₁Average of all elements in (A), R ∈ [0,1 ]]The joint approximation of the fitting values to the sample values is shown, with the fitting accuracy being higher the closer R is to 1.

2. A color cast detection method based on Lab chromaticity space is characterized in that: the method comprises the following steps:

step 1):

according to formula (1):formula (2):and formula (3):calculating equivalent circle features u and sigma of the chromaticity histogram on an ab plane; in formula (1), F (a, b) is a chromaticity histogram distribution function, k is a, b represents integration on the a-axis or b-axis, and u is_kAnd σ_kRespectively representing the mean value and the variance of the chromaticity histogram on a k axis; c in the formula (2) is the center of the equivalent circle, and sigma is the radius of the equivalent circle;

step 2):

executing a 'preliminary color cast detection' process: if formula (4) is satisfied: (D > 10and D_σ＞0.6)or(D_σ> 1.5), the image is classified as "color cast", shiftedStep 5); otherwise, the image is classified as 'no color cast', and the step 3) is carried out;

step 3):

and executing a non-color deviation image re-detection flow: according to equation (6):according to the criterion, the image is divided into three types of color cast, no color cast and unrecognizable; turning to step 4) for the 'unrecognizable' image; for "color cast" images, go to step 5).

Step 4):

executing a 'recognizable image retest' process:

1) the feature extraction algorithm according to claim 1, calculating a color shift feature h of the chromaticity histogram in the height direction;

2) the feature extraction algorithm of claim 1, calculating the feature change rate u of the chroma channel in the NNO region_cr、σ_crAnd h_cr；

3) Adopting the following criteria to detect the unrecognized image again:

step 5):

and executing a color cast image redetection flow:

1) the feature extraction algorithm of claim 1, calculating color shift features R and c of a luminance channel;

2) and adopting the following criteria to detect the color cast image again: