CN111277809A - An image color correction method, system, device and medium - Google Patents

Abstract

The application discloses an image color correction method, system, device and medium, comprising: dividing the collected image data into color distortion images and high-quality color image sets corresponding to the color distortion images; inputting the color distortion image into a generation model to obtain a corrected image; fixing parameters of a generated model, inputting the corrected image and the high-quality color image into a discrimination model, judging the image, and optimizing the parameters of the discrimination model until the discrimination model can distinguish two groups of images; and substituting the corrected image and the high-quality color image into a loss function, calculating loss, optimizing parameters of a generated model, inputting the corrected image and the high-quality color image into a trained discrimination model, and distinguishing the images until the discrimination model cannot distinguish two groups of images. The method for generating the countermeasure network corrects the color image, and the obtained model has the advantages of high correction speed, high precision and strong noise resistance.

Description

An image color correction method, system, device and medium

technical field

The present application relates to the technical field of color correction, and in particular, to an image color correction method, system, device and medium.

Background technique

Color is one of the important parameters in the field of computer vision. With the development of computer technology in the world, the requirements for color in various image application fields are getting higher and higher. Because the spectral sensitivity of the image sensor cannot best simulate the human visual system's ability to perceive color, and due to the uneven spectral distribution of the light source, the color of the collected image has a certain deviation from the real color of the object. Correction is made to restore true colors as much as possible.

Existing color correction schemes include spectral response-based methods and target color-based methods.

The principle of color correction based on spectral response is to measure the spectral response of the image sensor with a special instrument, and find the relationship between the spectral response of the image sensor to be corrected and the CIE color matching function.

The principle of the method based on the target color is to establish a calibration model with the color standard value and the measured value containing a certain amount of color samples, and obtain the functional relationship between the two, which is simple and practical. Its typical algorithms include: three-dimensional look-up table method, polynomial method, least square method with constraints, and pattern search method. Generally, the table lookup method is more accurate, but the disadvantage is that it requires a large number of standard tristimulus values of colors, and the table building speed is slow and the accuracy is not high. The polynomial rule is based on the functional relationship between the two, and the correction coefficient is obtained by the least square method, but it is only suitable for a very small color gamut space, and it is easy to amplify the noise. Even if the constraint term is added to the least squares method, although the accuracy of white can be improved, it will further enlarge the error after correction of other colors. For the polynomial method and the least square method, when constructing the correction model, it has been proved that the correction coefficient solved by using the 1-norm is more accurate than the 2-norm. However, since the 1-norm is not differentiable, the pattern search method can be used to find the global optimal solution to suppress noise amplification. The pattern search method may fall into the local optimum problem, the initial value influence is too large, and even the image brightness difference before and after correction is too large.

SUMMARY OF THE INVENTION

Embodiments of the present application provide an image color correction method, system, device, and medium, so that image color correction can be performed quickly, with high precision, and with strong noise resistance.

In view of this, a first aspect of the present application provides an image color correction method, the method comprising:

101. Divide the collected image data into two data sets, including a color-distorted image set and a high-quality color image set corresponding to the color-distorted image set;

102. Input the color distortion image into the constructed generative model to obtain a corrected image;

103. Train a discriminant model: fix the parameters of the generation model, and input the corrected image and the high-quality color image into the discriminant model, so that the discriminant model is responsible for the corrected image and the high-quality color image. The image is judged, the parameters of the discriminant model are optimized according to the judgement result, and step 103 is repeated until the discriminant model distinguishes the corrected image and the high-quality color image, that is, the discriminant model training is completed;

104. Train a generative model: substitute the corrected image and the high-quality color image into a loss function, calculate the loss between the corrected image and the high-quality color image, and calculate the loss between the corrected image and the high-quality color image. The parameters of the generation model are optimized, and the corrected image and the high-quality color image are input into the trained discriminant model, so that the discriminant model can determine the corrected image and the high-quality color image. The images are distinguished, and step 104 is repeated until the discriminant model cannot distinguish between the corrected image and the high-quality color image, that is, the training of the generative model is completed.

Preferably, the pair of collected image data is divided into two data sets, including a color-distorted image set and a high-quality color image set corresponding to the color-distorted image set, and further comprising:

preprocessing the set of color-distorted images and the set of high-quality color images;

The preprocessing includes: dividing the color-distorted image set and the high-quality color image set into a training set and a test set; randomly cropping the training images in the training set to obtain a plurality of image blocks, and performing the processing on the image blocks. Flip up and down, left and right to get a variety of training images.

Preferably, the loss function is specifically:

表示满足1-Lipschitz限制的函数。In the formula, E represents the expectation of distance; z represents the number of samples; P _g represents the sample distribution generated by the generator; D represents the probability of whether the real input and the label are fused and input into the discriminator to obtain whether it is real data;

represents a function satisfying the 1-Lipschitz limit.

Preferably, the activation function of the generative model is the same as the activation function of the discriminant model, specifically:

In the formula, b _{x, y} are the normalized feature vectors, a _{x, y} are the original feature vectors, and N is the number of feature maps; e is a tiny positive number used to avoid dividing by 0.

Preferably, before the inputting the corrected image and the high-quality color image corresponding to the corrected image into the loss function, the method further comprises: performing a regional analysis on the corrected image and the high-quality color image. Sampling, the area sampling is specifically:

Random difference sampling on the line connecting the corrected image and the high-quality color image samples:

In the formula, α represents a random number between 0-1, x _r represents the generated sample area; x _g represents the real sample area.

A second aspect of the present application provides an image color correction system, the system comprising:

a data acquisition module, which is configured to divide the collected image data into two data sets, including a color-distorted image set and a high-quality color image set corresponding to the color-distorted image set;

an image correction module, which is used for inputting the color-distorted image into the constructed generative model to obtain a corrected image;

A discriminant model training module, which is used to fix the parameters of the generated model, and input the corrected image and the high-quality color image into the discriminant model, so that the discriminant model can correct the corrected image. The image and the high-quality color image are judged, the parameters of the discriminant model are optimized according to the judgement result, and the operations in the discriminant model training module are repeated until the discriminant model distinguishes the corrected image and the high-quality color image;

a generative model training module, which is used to substitute the corrected image and the high-quality color image into a loss function, and calculate the loss between the corrected image and the high-quality color image, The parameters of the generation model are optimized according to the loss, and the corrected image and the high-quality color image are substituted into the loss function and input into the trained discriminant model, so that the discriminant model can affect the The corrected image and the high-quality color image are distinguished, and the operations in the generation model training module are repeated until the discrimination model cannot distinguish the corrected image from the high-quality color image.

Preferably, it also includes:

A preprocessing module, the preprocessing module is configured to perform preprocessing on the set of color-distorted images and the set of high-quality color images, and the preprocessing includes Divide into a training set and a test set; randomly crop the training images in the training set to obtain a plurality of image blocks, and flip the image blocks up and down and left and right to obtain a variety of training images.

Preferably, it also includes:

A sampling module, the sampling module is used to perform regional sampling on the corrected image and the high-quality color image, and the regional sampling is specifically:

Random difference sampling on the line connecting the corrected image and the high-quality color image samples:

In the formula, α represents a random number between 0-1, x _r represents the generated sample area; x _g represents the real sample area.

A third aspect of the present application provides an image color correction device, the device includes a processor and a memory:

the memory is used to store program code and transmit the program code to the processor;

The processor is configured to execute the steps of the image color correction method according to the first aspect above according to the instructions in the program code.

A fourth aspect of the present application provides a computer-readable storage medium, where the computer-readable storage medium is used to store program codes, and the program codes are used to execute the method described in the first aspect.

As can be seen from the above technical solutions, the present application has the following advantages:

In this application, an image color correction method is provided, which includes dividing the collected image data into color-distorted images and their corresponding high-quality color image sets; inputting the color-distorted images into a generation model to obtain a corrected image; The parameters of the model, input the corrected image and high-quality color image into the discriminant model, judge the image, and optimize the parameters of the discriminant model until the discriminant model can distinguish two sets of images; The color image is substituted into the loss function, the loss is calculated, and the parameters of the generation model are optimized. The corrected image and high-quality color image are input into the trained discriminant model, and the images are distinguished until the discriminant model cannot distinguish the two groups. image.

The method of the generative adversarial network of the present application integrates the optimization of the discriminant model and the generative model, through the mutual confrontation between the generative model and the discriminant model, and finally obtains the optimal weight parameters through the back-propagation algorithm, and finally obtains a stable The reliable model can correct any color distortion image, and the model obtained by the training of the present invention has high anti-noise interference ability, and only needs to be trained once, and the obtained model can be used repeatedly, which is efficient and convenient, and The correction of image color has certain robustness.

Description of drawings

1 is a method flowchart of an embodiment of an image color correction method of the present application;

2 is a schematic diagram of a system in an embodiment of an image color correction system of the application;

FIG. 3 is a schematic diagram of a model training process of a specific embodiment of an image color correction method of the present application.

Detailed ways

In order to make those skilled in the art better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only It is a part of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

Please refer to FIG. 1. FIG. 1 is a method flowchart of an embodiment of an image color correction method of the present application, and FIG. 1 includes:

101. Divide the collected image data into two data sets, including a color-distorted image set and a high-quality color image set corresponding to the color-distorted image set.

It should be noted that, in the color-distorted image set, each color-distorted image can find a corresponding high-quality image in the high-quality color image set.

In a specific implementation, the color distortion image set and the high-quality color image set also need to be preprocessed, and the preprocessing may include: dividing the color-distorted image set and the high-quality color image set into a training set and a test set; randomly The training images in the training set are cropped to obtain a plurality of image blocks, and the image blocks are flipped up and down and left and right to obtain various training images.

It should be noted that, in a specific implementation manner, the resolution of the image needs to be higher than 500×400, and the content of the image may include landscape, portrait, food, and the like. In addition, some images can be selected from the image set as a training set and some as a test set for training and testing the generative adversarial network. In addition, due to the different image sizes of the training set sample data, it cannot be directly input into the generative adversarial network for training. Therefore, before the training image, 150 image blocks of 256×256 size can be randomly cropped from the training image, and some images The blocks are randomly flipped up and down and left and right to increase the diversity of training images. At the same time, the size of the test images is adjusted to 900×600.

102. Input the color-distorted image into the constructed generative model to obtain a corrected image.

103. Train the discriminant model: fix the parameters of the generation model, input the corrected image and the high-quality color image into the discriminant model, so that the discriminant model judges the corrected image and the high-quality color image, and optimizes the discriminant model according to the judgment result , repeat step 103 until the discriminant model distinguishes the corrected image and the high-quality color image, that is, the discriminant model training is completed.

It should be noted that before training the generative model, the discriminative model needs to be trained first. Since the purpose of training the discriminant model is to enable the discriminant model to recognize the corrected image and high-quality color image, the parameters of the generative model can be fixed, and then input to the corrected image and high-quality output of the generative model. The color image is input into the discriminant model, the corrected image and the high-quality color image are judged, the parameters of the discriminant model are optimized according to the judgment result, and different images are input repeatedly, and the parameters of the discriminant model are continuously optimized until the discriminant The model is able to distinguish between corrected images and high-quality color images.

104. Train the generative model: Substitute the corrected image and the high-quality color image into the loss function, calculate the loss between the corrected image and the high-quality color image, optimize the parameters of the generative model according to the loss, and transfer the corrected image to the loss function. And the high-quality color image is input into the trained discriminant model, so that the discriminant model distinguishes the corrected image and the high-quality color image, repeating step 104 until the discriminant model cannot distinguish the corrected image and the high-quality color image, that is, The generative model training is complete.

It should be noted that the generation model is a convolutional neural network. When the color distortion image is input into the generation model, the generation model corrects the color distortion image so that the corrected image is close to the original high-quality color image. The image and the original high-quality color image are input to the loss function, the gap between the corrected image and the high-quality color image is calculated, and the parameters of the generative model are optimized according to the gap, so that the image corrected by the generative model is more and more The closer to the high-quality color image, the more the generative model can learn the features of the high-quality color image. In addition, it is also necessary to input the corrected image and high-quality color image into the trained discriminant model, so that the discriminant model can distinguish the corrected image and high-quality color image, and finally when the discriminant model cannot distinguish the corrected image. images and high-quality color images, the generative model has been trained.

The method of the generative adversarial network of the present application integrates the optimization of the discriminant model and the generative model, through the mutual confrontation between the generative model and the discriminant model, and finally obtains the optimal weight parameters through the back-propagation algorithm, and finally obtains a stable The reliable model can correct any color distortion image, and the model obtained by the training of the present invention has high anti-noise interference ability, and only needs to be trained once, and the obtained model can be used repeatedly, which is efficient and convenient, and The correction of image color has certain robustness.

The application also provides a specific implementation, as shown in Figure 3, specifically:

201. Randomly sample from the color-distorted image set, and take out a total of m image samples as a batch (batch), denoted as X.

202. Establish a generative model. The generative model is a convolutional neural network that takes a color-distorted image, corrects it and outputs it. When the color-distorted image is input into the constructed generative model, the generative model corrects the color-distorted image so that the corrected image is close to the original high-quality color image, and then the corrected image and the original high-quality color image are input To the loss function, the gap between the corrected image and the high-quality color image is calculated, and the parameters of the generative model are optimized according to the gap.

Using a generative model with Wasserstein distance added, the loss function is as follows:

表示满足1-Lipschitz限制的函数。In the formula, E represents the expectation of distance; z represents the number of samples; P _g represents the sample distribution generated by the generator; D represents the probability of whether the real input and the label are fused and input into the discriminator to obtain whether it is real data;

represents a function satisfying the 1-Lipschitz limit.

The input of the generative model is a color-distorted image, and its activation function is the same as that of the discriminant model, and SELU is used as the activation function. Model normalization adopts pixel feature vector normalization. It is located after the convolutional layer, so that each normalized feature vector has a unit length, which can be used to constrain the unhealthy competition between the generative model and the discriminant model, which leads to problems such as signal range out of bounds. The normalization announcement is specifically expressed as:

In the formula, b _{x, y} are the normalized feature vectors, a _{x, y} are the original feature vectors, and N is the number of feature maps; e is a tiny positive number used to avoid dividing by 0.

In the formula, the left side of the equation is the normalized feature vector, the numerator on the right side of the equation is the original feature vector, and N is the number of feature maps. The optimizer of the model can use RMSProp (root mean square prop) to update the parameters in the generated model. This optimizer needs to set the global learning rate, initial parameters, numerical stabilization and decay rate, and the optimizer can automatically adjust the learning rate.

203. The generation model accepts a batch of X, generates m correction samples according to the data distribution of the image samples in the batch, and inputs the corrected samples into the discriminant model for learning the discriminant model.

204. Generate a discriminant model. When the Wasserstein distance is used as the loss function in the generative adversarial network, the gradient penalty factor is added to the loss function, and when calculating the loss, it is necessary to sample from the corrected image area and the high-quality color image area, and also from the corrected image area. and sampling in the middle of the high-quality color image area. The sampling method is as follows: adding a random number α between 0 and 1, and randomly interpolating sampling on the corrected image area and the high-quality color image area. The formula is expressed as:

In the formula, α represents a random number between 0-1, x _r represents the generated sample area; x _g represents the real sample area.

205. In the specific training process, first fix the parameters in the generation model unchanged, input the corrected image and the high-quality color image together into the discriminant model, and the discriminant model calculates the difference between the corrected image and the high-quality color image, and calculates the discriminant. loss. The discriminative loss is back-propagated from the output layer to the hidden layer until the input layer. This process uses the RMSProp optimizer to update the parameters of the discriminant model, so that the discriminant model after optimizing the parameters can learn the corrected image and high-quality color image. features, thereby enabling the discriminative model to distinguish between corrected images and high-quality color images.

In the present application, an iterative method is adopted, and the discriminant model discriminates the corrected image and the high-quality color image until the discriminant model can correctly distinguish the corrected image and the high-quality color image.

206. Fix the parameters of the discriminant model, and train the generative model. The generative model receives color-distorted image samples, generates a corrected image, and transmits the corrected image and high-quality color image to the trained discriminant model to calculate the loss. The generative model expects to generate images that are close to high-quality color images. discriminant model. The parameters in the generative model are updated by backpropagation, and the generative model parameters are updated using the RMSProp optimizer. After the parameters are adjusted, the generation model generates the corrected image again, and input it into the discriminant model to see if the discriminant model can correctly distinguish the corrected image from the high-quality color image, and iterate continuously until the judgment model cannot distinguish between the corrected image and the high-quality color image. image.

207. The generative adversarial network can be gradually increased. The principle is to first train the initially corrected images, and then gradually transition to generating higher-resolution pictures. The work that needs to be done in the transition phase is to bring the rectified image generated by the generative adversarial network and the high-quality color image closer. When the previous stage of training is completed, Tensorflow saves the weights of the generative adversarial network in a folder, and then builds the weights of the generative adversarial network in the next stage. The new generative adversarial network will use the weight parameters of the previous stage, and the network layers of the generative model and the discriminant model will be deepened, and then the transition stage will be performed. In this process, the generative model performs upsampling and convolution, the results of upsampling and convolution are weighted and added to obtain the final result, and the discriminant model performs downsampling operation.

208. After the transition stage is completed, the model enters the stabilization stage. In this stage, the model needs to continue to update the weight parameters of the generative adversarial network, so that the generated generative adversarial network is closer to high-quality color images. 206, 207 are repeated until the model can stably generate images with realistic colors. Adjust the model hyperparameters and repeat the above process, select the optimal parameters, and the training is complete.

The above are the method embodiments of the present application, and the present application also includes a system schematic diagram in an embodiment of an image color correction system, as shown in FIG. 2 , including:

The data acquisition module 301 is configured to divide the collected image data into two data sets, including a color-distorted image set and a high-quality color image set corresponding to the color-distorted image set.

The image correction module 302 is used for inputting the color-distorted image into the constructed generative model to obtain a corrected image.

The discriminant model training module 303 is used to fix the parameters of the generated model, and input the corrected image and the high-quality color image into the discriminant model, so that the discriminant model judges the corrected image and the high-quality color image, and optimizes according to the judgment result The parameters of the discriminant model are repeated, and the operations in the discriminant model training module are repeated until the discriminant model distinguishes the corrected image and the high-quality color image.

The generation model training module 304 is used for substituting the corrected image and the high-quality color image into the loss function, calculating the loss between the corrected image and the high-quality color image, optimizing the parameters of the generating model according to the loss, and converting the corrected image and the high-quality color image. The images and high-quality color images are substituted into the loss function and input into the trained discriminant model, so that the discriminant model can distinguish the corrected images and high-quality color images, and repeat the operations in the training module of the generation model until the discriminant model cannot distinguish them. The corrected image as well as the high quality color image.

In a specific embodiment, it also includes:

The preprocessing module is used to preprocess the color distorted image set and the high-quality color image set. The preprocessing includes dividing the color-distorted image set and the high-quality color image set into a training set and a test set; randomly cropping the training images in the training set, Obtain multiple image blocks, flip the image blocks up and down, left and right, and obtain a variety of training images.

The sampling module is used to perform regional sampling on the corrected image and high-quality color image. The regional sampling is as follows:

Random difference sampling on the line connecting the corrected image to the high-quality color image samples:

In the formula, α represents a random number between 0-1, x _r represents the generated sample area; x _g represents the real sample area.

The application also provides an image color correction device, including a processor and a memory: the memory is used for storing program codes, and transmitting the program codes to the processor; Examples of calibration methods.

The present application also provides a computer-readable storage medium for storing program codes, and the program codes are used to execute the embodiments of the image color correction method of the present application.

Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above may refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

In this application, the terms "comprising" and "having" and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to the explicit lists Those steps or units listed may instead include other steps or units not expressly listed or inherent to these processes, methods, products or devices.

It should be understood that, in this application, "at least one (item)" refers to one or more, and "a plurality" refers to two or more. "And/or" is used to describe the relationship between related objects, indicating that there can be three kinds of relationships, for example, "A and/or B" can mean: only A, only B, and both A and B exist , where A and B can be singular or plural. The character "/" generally indicates that the associated objects are an "or" relationship. "At least one item(s) below" or similar expressions thereof refer to any combination of these items, including any combination of single item(s) or plural items(s). For example, at least one (a) of a, b or c, can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c" ", where a, b, c can be single or multiple.

In the several embodiments provided in this application, it should be understood that the disclosed system and method may be implemented in other manners. For example, the apparatus embodiments described above are only illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules or components may be combined or Can be integrated into another system, or some features can be ignored, or not implemented. On the other hand, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in electrical, mechanical or other forms.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

The integrated unit, if implemented in the form of a software functional unit and sold or used as an independent product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of software products in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, and the computer software products are stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (full English name: Read-Only Memory, English abbreviation: ROM), random access memory (English full name: RandomAccess Memory, English abbreviation: RAM), magnetic disk or Various media that can store program codes, such as optical discs.

As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand: The technical solutions described in the embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions in the embodiments of the present application.

Claims (10)

1. an image color correction method, is characterized in that, comprises: 101. Divide the collected image data into two data sets, including a color-distorted image set and a high-quality color image set corresponding to the color-distorted image set; 102. Input the color distortion image into the constructed generative model to obtain a corrected image; 103. Train a discriminant model: fix the parameters of the generation model, and input the corrected image and the high-quality color image into the discriminant model, so that the discriminant model is responsible for the corrected image and the high-quality color image. The image is judged, the parameters of the discriminant model are optimized according to the judgement result, and step 103 is repeated until the discriminant model distinguishes the corrected image and the high-quality color image, that is, the discriminant model training is completed; 104. Train a generative model: substitute the corrected image and the high-quality color image into a loss function, calculate the loss between the corrected image and the high-quality color image, and calculate the loss between the corrected image and the high-quality color image. The parameters of the generation model are optimized, and the corrected image and the high-quality color image are input into the trained discriminant model, so that the discriminant model can determine the corrected image and the high-quality color image. The images are distinguished, and step 104 is repeated until the discriminant model cannot distinguish the corrected image and the high-quality color image, that is, the training of the generative model is completed. 2 . The image color correction method according to claim 1 , wherein the collected image data is divided into two data sets, including a color-distorted image set and a high-quality color image set corresponding to the color-distorted image set. 3 . After that it also includes: preprocessing the set of color-distorted images and the set of high-quality color images; The preprocessing includes: dividing the color-distorted image set and the high-quality color image set into a training set and a test set; randomly cropping the training images in the training set to obtain a plurality of image blocks, and performing the processing on the image blocks. Flip up and down, left and right to get a variety of training images. 3. The image color correction method according to claim 1, wherein the loss function is specifically:

In the formula, E represents the expectation of distance; z represents the number of samples; P _g represents the sample distribution generated by the generator; D represents the probability of whether the real input and the label are fused and input into the discriminator to obtain whether it is real data;

represents a function satisfying the 1-Lipschitz limit. 4. The image color correction method according to claim 1, wherein the normalization function is specifically:

In the formula, b _{x, y} are the normalized feature vectors, a _{x, y} are the original feature vectors, and N is the number of feature maps; e is a tiny positive number used to avoid dividing by 0. 5 . The image color correction method according to claim 1 , further comprising: before inputting the corrected image and the high-quality color image corresponding to the corrected image into the loss function. 6 . Including, performing regional sampling on the corrected image and the high-quality color image, and the regional sampling is specifically: Random difference sampling on the line connecting the corrected image and the high-quality color image samples:

In the formula, α represents a random number between 0 and 1; x _r represents the generated sample area; x _g represents the real sample area. 6. An image color correction system, characterized in that, comprising: a data acquisition module, which is configured to divide the collected image data into two data sets, including a color-distorted image set and a high-quality color image set corresponding to the color-distorted image set; an image correction module, which is used for inputting the color-distorted image into the constructed generative model to obtain a corrected image; A discriminant model training module, which is used to fix the parameters of the generated model, and input the corrected image and the high-quality color image into the discriminant model, so that the discriminant model can correct the corrected image. The image and the high-quality color image are judged, the parameters of the discriminant model are optimized according to the judgement result, and the operations in the discriminant model training module are repeated until the discriminant model distinguishes the corrected image and the high-quality color image; a generative model training module, which is used to substitute the corrected image and the high-quality color image into a loss function, and calculate the loss between the corrected image and the high-quality color image, The parameters of the generation model are optimized according to the loss, and the corrected image and the high-quality color image are substituted into the loss function and input into the trained discriminant model, so that the discriminant model can affect the The corrected image and the high-quality color image are distinguished, and the operations in the generation model training module are repeated until the discrimination model cannot distinguish the corrected image from the high-quality color image. 7. The image color correction system according to claim 6, further comprising: A preprocessing module, the preprocessing module is configured to perform preprocessing on the set of color-distorted images and the set of high-quality color images, and the preprocessing includes Divide into a training set and a test set; randomly crop the training images in the training set to obtain a plurality of image blocks, and flip the image blocks up and down and left and right to obtain a variety of training images. 8. The image color correction system according to claim 6, further comprising: A sampling module, the sampling module is used to perform regional sampling on the corrected image and the high-quality color image, and the regional sampling is specifically: Random difference sampling on the line connecting the corrected image and the high-quality color image samples:

In the formula, α represents a random number between 0-1, x _r represents the generated sample area; x _g represents the real sample area. 9. An image color correction device, characterized in that the device comprises a processor and a memory: the memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the image color correction method according to any one of claims 1-5 according to the instructions in the program code. 10. A computer-readable storage medium, wherein the computer-readable storage medium is used to store program codes, and the program codes are used to execute the image color correction method according to any one of claims 1-5.

Images