A method for removing dust storm from images by integrating color cast correction

Abstract

The present invention relates to an image sandstorm removal method integrating color cast correction. The method connects multiple dust perception representation extraction DustAre modules with the same structure, first performs multiple downsampling operations on the input sandstorm image, extracts image features and fuses them; then, performs multiple upsampling operations on the extracted image features and restores them to the original image space to obtain the final sandstorm-removed image. The present invention: (1) In view of the color deviation problem of sandstorm images, the traditional color cast detection and correction strategy is adopted for the label image during network model training, so that the color of the image sandstorm removal result is more natural; (2) Based on the discovery that the chromaticity values ​​of the a* and b* color components of the sandstorm image are much larger than those of the image without sandstorm, a new loss function term is designed to better guide the network model training; (3) A sandstorm image generation model is proposed, and a sandstorm image dataset containing indoor and outdoor labeled images is synthesized for network model training.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to an image sandstorm removal method integrating color cast correction. Background Art

[0002] Dust storms have a negative impact on the capabilities of monitoring systems such as license plate detection for outdoor monitoring equipment and objective recognition technology for autonomous driving. Images taken during dust storms often exhibit color shift and low contrast due to the scattering and absorption of light when propagating through dust. Dust storm image input will bring difficulties to other computer vision tasks. Therefore, image desandstorming has become an important research topic in the field of computer vision as a preprocessing step for advanced computer vision tasks.

[0003] Existing image restoration methods can be divided into traditional color cast correction methods and deep learning-based methods. Traditional color cast correction methods usually convert the sandstorm image to the CIELAB color space first. The CIELAB color space includes two chromaticity components and one brightness component; then, the two chromaticity components are corrected for color cast in the CIELAB color space, and the brightness component is enhanced for detail to obtain the corresponding clean image. However, traditional image sandstorm removal methods only focus on image color contrast adjustment, without considering the relationship between the impact of sand and dust on the image and the image scene depth. Therefore, the restoration effect of traditional methods often has the problems of low color saturation and more residual sand and dust.

[0004] In recent years, deep learning technology has been effectively applied in image restoration. Deep learning-based methods usually design convolutional neural networks to learn image features and train the network on large-scale image datasets to generate restored images. Since most existing supervised learning neural networks are trained on synthetic image datasets, deep learning-based methods can easily produce significant results on synthetic datasets, but the results on real image datasets are poor. In addition, due to the lack of sandstorm image datasets for network training, most existing image desandstorm research still remains on traditional methods.

[0005] Similar to foggy images, the impact of sandstorms on images also depends on the depth of the image scene. When synthesizing sandstorm images, factors such as the image scene depth relationship between the sandstorm image and the clean image and the color of the sandstorm image should be considered. Therefore, the present invention describes the formation process of the sandstorm image based on the specific characteristics of the sandstorm image:

[0006] I(x)=J(x)d(x)p(x)+C(1-d(x)p(x))

[0007] Where I and J represent the sandy image and the clean image, respectively. Symbol C represents the sandstorm color map, i.e., the global masking light. Symbol p represents the image scene depth map obtained by the pre-trained monocular image scene depth estimation method. d represents the relationship between the scene depth of the sandstorm image and the scene depth of the clean image. Summary of the invention

[0008] The purpose of the present invention is to provide an image sandstorm removal method integrating color cast correction in order to more effectively remove image sandstorms. The method has a better image sandstorm removal effect than the most advanced image restoration method currently.

[0009] To achieve the above-mentioned purpose, the technical solution of the present invention is: a method for removing sandstorms from images by integrating color cast correction, constructing an image sandstorm removal network, that is, by connecting multiple dust perception representation extraction DustAre modules with the same structure, first performing multiple downsampling operations on the input sandstorm image, extracting image features and fusing them; then, performing multiple upsampling operations on the extracted image features, restoring them to the original image space, to obtain the final sandstorm-removed image.

[0010] In one embodiment of the present invention, the height and width of the image dust storm removal network are 3 and 6 respectively. They represent the input and output of the i-th DustAre module in the r-th row of the image dust storm removal network, respectively. Then, during the downsampling operation, the input of the i+1-th DustAre module in the 2nd and 3rd rows can be expressed as

[0011]

[0012] Similarly, during the upsampling operation, the input of the i+1th DustAre module in rows 2 and 1 can be expressed as,

[0013]

[0014] Among them, ↓ and ↑ represent downsampling operation and upsampling operation respectively.

[0015] In one embodiment of the present invention, the DustAre module includes two branches, wherein the first branch is used to estimate the global occlusion light, i.e., the ambient light scattered in the line of sight and the local spatial information, and the second branch is used to generate a sandstorm perception image, and fuse the sandstorm perception image, the global occlusion light and the local spatial information to obtain the output of the DustAre module.

[0016] In one embodiment of the present invention, the first branch uses a 1×1×2 matrix to encode the input features and generate a global mask light representation (β V ,γ V) and use H×W×2 matrix to generate local spatial information representation (β S ,γ S ); the second branch first uses the input x to normalize two sets of representations according to formulas (1) and (2), denoted as V and S respectively, and generates the dust storm perception map A through the instance normalization layer and the sigmoid layer; then, A, V and S are fused to obtain the output y of the DustAre module; where, for the input x, V, S and output y of the DustAre module are expressed as:

[0017]

[0018]

[0019] Here, μ and σ represent the mean and standard deviation of x, respectively.

[0020] In one embodiment of the present invention, the method further introduces a color cast correction strategy, that is, using the color cast strategy to detect and correct the label image of the training data set, so that the color of the final dust storm-free image is more natural, specifically:

[0021] Here, it is considered that when the color cast factor K ≥ 1.5, the label image itself is more likely to have color cast; therefore, the a* and b* components of these images are corrected in the CIELAB color space, that is, the final training label image The expression of is,

[0022]

[0023] in represents the original label image, and cor represents the color cast correction operation.

[0024] In one embodiment of the present invention, considering the chromaticity difference between the sandstorm image and the clean image in the CIELAB color space, the method also designs a new loss function In order to better guide the network training, specifically, the total loss function of the method is defined as:

[0025]

[0026] where λ ab , L1 and per Set to 0.3, 1.2 and 0.4 respectively;

[0027] Item 1 It is derived from the observation of sandstorm images and clean images. Since sandstorm images have serious color deviation, the average a* and b* chromaticity values ​​of sandstorm images in CIELAB color space are usually greater than the average a* and b* chromaticity values ​​of clean images. Therefore, the first loss term for,

[0028]

[0029] Where N is the total number of image pixels. In the CIELAB color space of the dust storm removal result y, y a and b Represents the a* and b* color components respectively, after color cast correction The corresponding clear image is the label image In the CIELAB color space, and Represent a* and b* color components respectively;

[0030] The second term is the smoothed L1 loss, denoted as It prevents potential gradient explosion and is defined as follows:

[0031]

[0032] The last item, perceptual loss It is used to motivate images with similar features in the Backbone network, which is defined as:

[0033]

[0034] Among them, H, W and C represent the height, weight and channel of the feature map in the i-th layer of the network respectively, φ i (y) and Represents y and The output feature map in the i-th layer, y and Represent the dust storm removal results and the corresponding clean images respectively.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention introduces deep learning technology into the field of image sandstorm removal and proposes an image sandstorm removal method (SIDNet) that integrates color correction. The algorithm of the present invention connects multiple sandstorm perception representation extraction modules with the same structure, first performs multiple downsampling operations on the input sandstorm image to extract and fuse image features; then performs multiple upsampling operations on the feature map to restore it to the image space, thereby obtaining the image sandstorm removal result. The specific innovations are as follows: (1) In response to the color deviation problem of sandstorm images, the traditional color cast detection and correction strategy is used for the label image during network model training, so that the color of the image sandstorm removal result is more natural; (2) Based on the discovery that the chromaticity values ​​of the a* and b* color components of the sandstorm image are much larger than those of the non-sandstorm image, a new loss function term is designed to better guide the network model training; (3) A sandstorm image generation model is proposed, and a sandstorm image dataset containing indoor and outdoor labeled images is synthesized for network model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The figure is a flow chart of the method of the present invention.

[0038] Figure 2 This is the network structure of the method (SIDNet) of the present invention.

[0039] Figure 3 This is the dust storm representation extraction (DustAre) module.

[0040] Figure 4 Comparison of a* and b* components of a set of sandstorm images and clean images in similar scenes; (a) real sandstorm image, (b) its a* component in CIELAB color space, (c) b* component in CIELAB, (d) its corresponding equivalent circle, (e) real clean image, (f) its corresponding a* component, (g) corresponding b* component, and (h) corresponding equivalent circle.

[0041] Figure 5 The dust storm removal results of various algorithms on outdoor synthetic images; (a) dust storm image, (b) LPNet algorithm, (c) RGNet algorithm, (d) AODNet algorithm, (e) FFANet algorithm, (f) Wang et al.'s algorithm, (g) HardGAN algorithm, (h) FFNet algorithm, (i) SIDNet algorithm, (j) label image.

[0042] Figure 6The dust storm removal results of various algorithms on indoor synthetic images; (a) dust storm image, (b) LPNet algorithm, (c) RGNet algorithm, (d) AODNet algorithm, (e) FFANet algorithm, (f) Wang et al.'s algorithm, (g) HardGAN algorithm, (h) FFNet algorithm, (i) SIDNet algorithm, (j) label image.

[0043] Figure 7 The images show the dust storm removal results of different methods on real dust storm images.

[0044] Figure 8 This is the edge extraction result of the dust storm removal sample image.

[0045] Fig. 9 The number of key point matches for sandstorm removal results using different algorithms in a group of similar scenes. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0047] like Figure 1 As shown, the present invention provides an image sandstorm removal method integrating color cast correction, and constructs an image sandstorm removal network, that is, by connecting multiple dust perception representation extraction DustAre modules with the same structure, firstly performing multiple downsampling operations on the input sandstorm image, extracting image features and fusing them; then, performing multiple upsampling operations on the extracted image features, and restoring them to the original image space, so as to obtain the final sandstorm removal image.

[0048] The following is the specific implementation process of the present invention.

[0049] 1. Network structure of the algorithm of the present invention

[0050] like Figure 2 As shown in the figure, the image dust storm removal network structure proposed in the present invention first connects multiple dust storm perception representation extraction (DustAre) modules with the same structure, performs multiple downsampling operations on the input dust storm image, extracts image features and fuses them; then, performs multiple upsampling operations on the extracted image features to generate dust storm removal results. The height and width of the network are 3 and 6 respectively. Specifically, Respectively represent the input and output of the i-th DustAre module in the r-th row of the network. Then, during the downsampling operation, the input of the i+1-th DustAre module in the 2nd and 3rd rows can be expressed as,

[0051]

[0052] Similarly, during the upsampling operation, the input of the i+1th DustAre module in rows 2 and 1 can be expressed as,

[0053]

[0054] Among them, ↓ and ↑ represent downsampling operation and upsampling operation respectively.

[0055] 2. Sandstorm Characterization Extraction Module

[0056] In order to remove dust storms from images while preserving image details, it is particularly important to estimate global masking light and local spatial information. Inspired by the existing dehazing network, this paper uses a dust storm representation extraction module to extract their representations. The structure of the dust storm perception representation extraction (DustAre) module is as follows: Figure 3 As shown in Figure 1, it mainly consists of two branches. The first branch uses a 1×1×2 matrix to encode the input features and generate a global mask light representation (denoted as β V ,γ V ), and the H×W×2 matrix generates the local spatial information representation (denoted as β S ,γ S ). The second branch first normalizes two sets of representations (respectively denoted as V and S) with input x according to formulas (1) and (2), and generates the dust storm perception map (A) through the instance normalization layer and the sigmoid layer; then, A, V and S are fused to obtain the output y of the DustAre module. Among them, for the input x, V, S and output y of the DustAre module, they are expressed as:

[0057]

[0058]

[0059] Here, μ and σ represent the mean and standard deviation of x, respectively.

[0060] 3. Color cast correction

[0061] Most existing deep learning-based image restoration methods have a common problem, that is, they have achieved remarkable results on synthetic datasets, but the results on real images are often poor. One of the main reasons is that the final image restoration result is overly dependent on the training dataset, and even due to the color cast of the training dataset itself, the color of the final real image's sandstorm removal result is unnatural. In order to solve this problem, the algorithm SIDNet of the present invention uses a color cast strategy to detect and correct the label images of the training dataset to improve the sandstorm removal effect of real sandstorm images. Specifically, the present invention believes that when the color cast factor K≥1.5, the label image itself is more likely to have color cast. Therefore, the a* and b* components of these images are corrected in the CIELAB color space, that is, the final training label image The expression of is,

[0062]

[0063] in represents the original label image, and cor represents the color cast correction operation.

[0064] 4. Loss Function

[0065] The total loss function of the algorithm SIDNet of the present invention is defined as:

[0066]

[0067] where λ ab , L1 and per Set to 0.3, 1.2 and 0.4 respectively;

[0068] Item 1 Based on the observation of sandstorm images and clean images, since sandstorm images have serious color deviation, the average a*, b* chromaticity values ​​of sandstorm images in the CIELAB color space are usually greater than the average a*, b* chromaticity values ​​of clean images. Figure 4 Two similar sandstorm images and clean images under real outdoor scenes are shown as samples to verify this observation. Figure 4 (a)-(d) show the real dust storm image, its a* chromaticity component in CIELAB color space, b* chromaticity component in CIELAB, and the corresponding equivalent circle. Figure 4 (e)-(h) show the clean image, its corresponding a* chromaticity component, b* chromaticity component and the corresponding equivalent circle. Figure 4It can be observed that the a* and b* components of the dust storm image are blurrier and brighter than those of the clean image, which indicates that the average a* and b* chromaticity values ​​of the dust storm image are larger than those of the clean image. Figure 4 The data given in (d) and (h) are consistent. In order to further verify this observation, the present invention calculated the average a* and b* chromaticity values ​​of 50 real sandstorm and clean images. Experimental data show that the average a* and b* chromaticity values ​​of 50 real sandstorm images are 13.53 and 35.69, respectively. The average a* and b* chromaticity values ​​of 50 clean images are 0.28 and 2.99, respectively. Based on these data, it can be concluded that sandstorm images usually have larger average a* and b* chromaticity values ​​in the CIELAB color space. Therefore, it can be inferred that the smaller the a* and b* chromaticity average values ​​of the image sandstorm removal result, the better the image sandstorm removal effect. Based on the above observations, the present invention designs the first loss term for,

[0069]

[0070] Where N is the total number of image pixels. In the CIELAB color space of the dust storm removal result y, y a and b Represents the a* and b* color components respectively, after color cast correction The corresponding clear image is the label image In the CIELAB color space, and Represent a* and b* color components respectively;

[0071] The second term is the smoothed L1 loss, denoted as It prevents potential gradient explosion and is defined as follows:

[0072]

[0073] The last item, perceptual loss It is used to motivate images with similar features in the Backbone network, such as the pre-trained VGG19 in this paper, which is defined as,

[0074]

[0075] Among them, H, W and C represent the height, weight and channel of the feature map in the i-th layer of the network respectively, φ i (y) and Represents y and The output feature map in the i-th layer, y and Represent the dust storm removal results and the corresponding clean images respectively.

[0076] The method of the present invention (SIDNet) is compared with seven existing image restoration methods, namely LPNet, RGNet, AODNet, FFANet, Wang et al.'s algorithm, HardGAN, and FFNet. All algorithms are trained on synthetic sandstorm datasets, and the sandstorm removal effect is tested on synthetic datasets and real sandstorm images. The quantitative comparison experiment uses PSNR (peak signal-to-noise ratio) and SSIM (structural similarity) to evaluate the sandstorm removal effect of synthetic sandstorm images; the quality based on spatial spectral entropy (SSEQ) is used to evaluate the sandstorm removal effect of real sandstorm images. In the present invention, in order to keep the correlation consistent with the peak signal-to-noise ratio (PSNR) and SSIM metrics, this paper reverses the SSEQ value, that is, the higher the value of all indicators, the better the sandstorm removal effect.

[0077] (1) Quantitative evaluation of dust storm removal results of outdoor synthetic images

[0078] In order to quantitatively compare different algorithms, the average PSNR, SSIM and SSEQ values ​​of the dust storm removal results of each algorithm on outdoor synthetic images are listed in Table 1. The quantitative evaluation results in Table 1 show that on outdoor synthetic images, the PSNR and SSIM measurement values ​​of the proposed method are significantly higher than those of other image restoration algorithms, which means that the dust storm removal algorithm is significantly better than other comparison algorithms.

[0079] Table 1 Average PSNR and SSIM values ​​of outdoor synthetic image sandstorm results

[0080]

[0081] (2) Qualitative evaluation of dust storm removal results of outdoor synthetic images

[0082] Qualitative evaluation of outdoor synthetic images and dust storm removal results is shown in Figure 5 As shown. Figure 5 It can be observed that the dust storm removal results of outdoor images obtained by the method of the present invention are closest to the labeled image situation (i.e., clean images without sandstorms). Compared with other reference methods, HardGAN and FFNet have better dust storm removal effects. For some scenes, AODNet cannot remove the dust storm. LPNet, RGNet, and FFANet cause different degrees of color distortion. The algorithm of Wang et al. corrects the color of the dust storm image, but their dust storm removal results still show low color saturation.

[0083] (3) Quantitative evaluation of dust storm removal results of indoor synthetic images

[0084] In order to adapt to image dust removal in various situations in reality, the method of the present invention is also experimented on an indoor synthetic image dataset, which is characterized by rich colors and many details. As shown in Table 2, the quantitative evaluation of the dust removal results of indoor synthetic images shows that the method of the present invention obtains the highest average PSNR (32.82) and SSIM (0.9432) values ​​in the dust removal results, indicating that its dust removal effect is better than other reference methods.

[0085] Table 2 Average PSNR and SSIM values ​​of indoor synthetic image dust removal results

[0086]

[0087] (4) Qualitative evaluation of dust storm removal results of indoor synthetic images

[0088] For the qualitative evaluation of the dust storm removal results of indoor synthetic images, Figure 6 As shown. The dust storm removal result obtained by the method of the present invention (SIDNet) is more similar to the label image, indicating that the image dust storm removal effect is the best. HardGAN and FFNet can effectively remove most of the dust storms, but there is a certain deviation in their color restoration. The dust storm removal results obtained by Wang et al. have low color saturation. LPNet, RGNet and FFANet will cause different color distortions. AODNet can remove a small amount of dust storms, but some dust storm removal results have low color saturation, such as the last image.

[0089] (5) Quantitative evaluation of dust storm removal results of real images

[0090] Table 3 shows the quantitative evaluation of the dust storm removal results of 8 representative real images. The quantitative results in Table 3 show that the proposed algorithm SIDNet obtains the highest SSEQ value on most real dust storm images (except the first image, the fifth image, and the eighth image), and obtains the second best SSEQ value on the first image. In addition, the proposed algorithm has the highest average SSEQ value (69.66), showing the best dust storm removal effect.

[0091] Table 3 SSEQ values ​​and average SSEQ values ​​of the dust storm removal results obtained by different methods on 8 real dust storm images.

[0092]

[0093] (6) Qualitative evaluation of dust storm removal results of real images

[0094] Figure 7 The dust storm removal results of 8 representative real dust storm images are shown. Figure 7It can be seen that LPNet and AODNet can remove a small amount of sandstorms, but the color of their sandstorm removal effects is dark. The sandstorm removal results of RGNet and FFANet have serious color distortion. The sandstorm removal results of Wang et al.'s algorithm have low color saturation. HardGAN and FFNet have better sandstorm removal effects, but there are still sandstorms in their sandstorm removal results, such as the 3rd, 4th and last images. In comparison, the method of the present invention can strike a good balance between better image sandstorm removal effect and high color saturation.

[0095] (7) Robustness test

[0096] In order to verify the significance of image dust storm removal for other visual tasks, the present invention carries out two potential applications, namely image edge extraction and local key point matching.

[0097] (a) Image edge extraction

[0098] Due to the color cast and low contrast of sandstorm images, it is difficult to extract the edges of sandstorm images. In order to verify the importance of SIDNet for image edge extraction, this section analyzes the effects of different methods on edge extraction of sandstorm images. Figure 8 As shown in the figure, compared with other similar methods, the edge details in the SIDNet dust storm removal effect are significantly increased. This verifies that SIDNet can significantly improve the image edge extraction effect by restoring more details.

[0099] (b) Local key point matching

[0100] The purpose of local key point matching is to find the correspondence between two similar scenes. In order to verify the importance of SIDNet in local key point matching, this paper uses the SIFT (Scale Invariant Feature Transform) algorithm to match key points on a set of similar real strong sandstorm images (all taken from a square) and the corresponding sandstorm-free images obtained by different methods. The matching results are shown in Figure 2. Fig. 9 As shown in FIG. 1 , the number of key point matches in the dust storm-free image increases significantly, which verifies that the SIDNet method of the present invention can also restore the local features of the dust storm image.

[0101] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. A method for removing sandstorm from an image by integrating color cast correction, characterized in that: Construct an image dust storm removal network, that is, by connecting multiple dust perception representation extraction DustAre modules with the same structure, first perform multiple downsampling operations on the input dust storm image, extract image features and fuse them; Then, the extracted image features are upsampled multiple times and restored to the original image space to obtain the final dust storm-free image; Considering the chromaticity difference between the sandstorm image and the clean image in the CIELAB color space, this method also designs a new loss function In order to better guide the network training, specifically, the total loss function of the method is defined as: where λ ab , L1 and per They are set to 0.3, 1.2 and 0.4 respectively; Item 1 It is derived from the observation of sandstorm images and clean images. Since sandstorm images have serious color deviation, the average a* and b* chromaticity values ​​of sandstorm images in CIELAB color space are greater than the average a* and b* chromaticity values ​​of clean images. Therefore, the first loss term for, Where N is the total number of image pixels. In the CIELAB color space of the dust storm removal result y, y a and b Represents the a* and b* color components respectively, after color cast correction The corresponding clear image is the label image In the CIELAB color space, and Represent a* and b* color components respectively; The second term is the smoothed L1 loss, denoted as Prevent potential gradient explosion, which is defined as follows: The last item, perceptual loss It is used to motivate images with similar features in the Backbone network, which is defined as: Among them, H, W and C represent the height, weight and channel of the feature map in the i-th layer of the network respectively, φ i (y) and Represents y and The output feature map in the i-th layer, y and Represent the dust storm removal results and the corresponding clean images respectively.

2. The image sandstorm removal method integrating color cast correction according to claim 1, characterized in that: The height and width of the image dust storm removal network are 3 and 6 respectively. They represent the input and output of the i-th DustAre module in the r-th row of the image dust storm removal network, respectively. Then, during the downsampling operation, the input of the i+1-th DustAre module in the 2nd and 3rd rows can be expressed as Similarly, during the upsampling operation, the input of the i+1th DustAre module in rows 2 and 1 can be expressed as, Among them, ↓ and ↑ represent downsampling operation and upsampling operation respectively.

3. The image sandstorm removal method integrating color cast correction according to claim 1 or 2, characterized in that: The DustAre module includes two branches, wherein the first branch is used to estimate the global occlusion light, i.e., the ambient light scattered in the line of sight and the local spatial information, and the second branch is used to generate a sandstorm perception image and fuse the sandstorm perception image, the global occlusion light and the local spatial information to obtain the output of the DustAre module.

4. The image sandstorm removal method integrating color cast correction according to claim 3 is characterized in that: The first branch encodes the input features using a 1×1×2 matrix and generates a global mask representation (β V ,γ V ) and use H×W×2 matrix to generate local spatial information representation (β S ,γ S ); the second branch first uses the input x to normalize two sets of representations according to formulas (1) and (2), denoted as V and S respectively, and generates the dust storm perception map A through the instance normalization layer and the sigmoid layer; then, A, V and S are fused to obtain the output y of the DustAre module; where, for the input x, V, S and output y of the DustAre module are expressed as: Here, μ and σ represent the mean and standard deviation of x, respectively.

5. The image sandstorm removal method integrating color cast correction according to claim 1, characterized in that: This method also introduces a color cast correction strategy, that is, using the color cast strategy to detect and correct the label images of the training data set, so that the final dust storm-free image has a more natural color. Specifically: Here, it is considered that when the color cast factor K ≥ 1.5, the label image itself is more likely to have color cast; therefore, the a* and b* components of these images are corrected in the CIELAB color space, that is, the final training label image The expression of is, in represents the original label image, and cor represents the color cast correction operation.

Images