A generative adversarial network method for image color cast removal
By building a model suitable for color constancy calculation based on the generative adversarial network, the problem of poor color constancy calculation in the existing technology is solved, and the color offset removal effect is improved in complex light source scenarios, which is enhanced. The robustness and generalization ability of the model.
Patent Information
- Application Number
- CN202111209869.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-18
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-10-18
AI Technical Summary
The lack of appropriate model parameters and network structure in color constancy calculations in the prior art leads to poor color bias removal effects, and the complexity of deep learning networks increases the challenge of model robustness and generalization capabilities.
Using a generative adversarial network-based method, by constructing a color constant data set, using U-Net connection as the generator structure, Markov discriminator as the discriminator structure, introducing Leaky Relu activation function and BN layer, adding Dropout function, and An angle error loss function is introduced to optimize the training model to improve the color shift removal effect.
The model's color shift removal ability under different light sources is improved, and the color shift removal results with good stability and excellent processing performance are obtained, which enhances the model's robustness and generalization ability.
Smart Images

Figure CN114004758B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to computer vision enhancement technology, and in particular to an image enhancement method based on a generative adversarial network for color cast removal. Background Art
[0002] In the research of computer vision, image understanding is one of its important research directions. The basis of image understanding is image features, among which color is one of the most important underlying features of images, and has a strong reference significance in image feature extraction and feature learning. However, the color information of objects in image scenes is unstable, and the color of objects is easily affected by changes in light sources. The purpose of removing image color cast is to correctly perceive the color and obtain the true color information of objects that is independent of light source changes.
[0003] Since the color constancy of color images is inherently limited and complex, the color of an object in an image is the result of the combined effects of the object color, the light source color, and the optical properties of the imaging device. Therefore, traditional color constancy algorithms have added appropriate assumptions or restrictions, which can be roughly divided into two categories: statistical-based methods and learning-based methods. Among them, statistical-based methods assume that images taken under standard light sources follow certain properties, such as the Gray World algorithm and the Von Kries model; learning-based methods require a set of labeled image sets to train the color constancy model. Different statistical-based methods are suitable for different scenarios, such as classifying images into indoor or outdoor scenes, and then determining the most appropriate algorithm for different scenarios. With the development of computer science, deep learning technology has surpassed traditional algorithms in a short period of time and has become the most advanced method for removing color casts in images.
[0004] Convolutional Neural Networks (CNNs) are typical deep learning models with strong model fitting ability used to study color constancy. One of the advantages of CNNs is that the original image can be used as input and feature learning can be integrated into the training process. With a deep structure, CNN can learn complex mappings. Bianco et al. extracted the features of image patches from a pre-trained CNN and provided them to a regression module to estimate the scene illumination. The working process is that in a given color image, non-overlapping patches are first sampled and contrast normalized for each patch by histogram stretching. Then CNN is used to estimate the illumination of each patch and combined with the patch scores to obtain the light source estimation of the image.
[0005] The recent development of Generative Adversarial Networks (GANs) has ushered in a new era of synthetic image generation, which has shown remarkable results in various computer vision tasks. The core principle behind any GAN model is the competition between two modules: the discriminator and the generator. The discriminator learns to distinguish between real and fake samples, while the generator learns to generate fake samples that are indistinguishable from real samples. Conditional image generation based on GANs has also been actively studied. The Pix2Pix algorithm proposed by Isola et al. showed excellent performance in image mapping tasks, and Das et al. applied the Pix2Pix algorithm to the color constancy task and achieved good results.
[0006] The application of deep learning in color constancy research has been recognized by the public. Convolutional neural networks have been fully utilized by many researchers, and generative adversarial networks also have good application prospects. By utilizing the powerful image feature learning ability of deep learning, many artificial assumptions can be discarded, thereby enhancing the robustness and generalization ability of the model. However, deep learning networks are also very complex, involving many parameters such as activation functions, loss functions, and learning rates. Different parameters affect different model effects. Currently, there is a lack of appropriate model parameters and network structures that can achieve better results in color constancy calculations, and there is no data set that can enhance the generalization ability of the model. Summary of the invention
[0007] In view of the problems existing in the prior art, the present invention proposes a method for removing image color cast based on a generative adversarial network, which realizes image color cast removal by enhancing two aspects, namely, constructing a color constancy dataset and improving the quality of the image after color cast removal.
[0008] In view of the problems of existing research methods, the research content of this invention includes the following parts:
[0009] (1) The SFU Grayball dataset used for color constancy calculation is used as the training set. By artificially adding simulated complex light sources, this dataset is expanded to become the experimental dataset required for training color constancy models of uniform and non-uniform light sources.
[0010] (2) A color cast removal model based on generative adversarial network was constructed. The "U-Net" connection was used as the generator structure of the model, the Markov discriminator was used as the discriminator structure, the Leaky Relu activation function was used instead of the Relu activation function, the Batch Normalization (BN) layer and the Dropout() function were added, and a loss function that can better reflect the hue information was introduced to optimize the training model and improve the effect of the model on color cast removal under different light sources.
[0011] The specific implementation steps of the color cast removal method based on the generative adversarial network include:
[0012] Step 1: construct a color-biased image dataset, preprocess the dataset, and form image pairs of images with color bias and images with corrected color bias;
[0013] Step 2, construct a color cast removal model based on a generative adversarial network, use the "U-Net" connection as the generator structure of the color cast removal model, and use the Markov discriminator as the discriminator structure;
[0014] Step 3: Use the Leaky Relu activation function instead of the Relu activation function in the network structure of the generator and the discriminator;
[0015] Step 4, designing a corresponding total loss function, using the image in step 1 to train the color cast removal model constructed in step 2, and obtaining a model file;
[0016] Step 5: Use the model file to remove color cast from the test image.
[0017] Furthermore, the dataset used in step 1 is the SFU Grayball dataset for color constancy calculation. Before training, in order to expand the dataset and simulate color-shifted images under complex light sources, the image data was preprocessed as follows:
[0018] Step 11: Color cast dataset construction under complex light sources
[0019] Since complex light source scenes can be considered as the result of multiple light sources illuminating the same scene together, and each light source has its own chromaticity value, a Gaussian distribution map generated with random parameters is used to simulate sufficiently complex light source situations. This situation can be regarded as a random combination of three light sources of different colors. Then, the standard illumination map of the SFU Grayball dataset is colored using the inverse von Kries transform to obtain a simulated image dataset under complex light source conditions.
[0020] Step 2: Image cropping and stitching
[0021] Since each image in the original data set contains a small gray ball, the ball will affect the network's information extraction and easily cause the network to "cheat", that is, the generator can easily generate specific images directly based on the gray ball. Therefore, the image needs to be cropped to remove the ball before processing the data set;
[0022] In addition, since the model training requires the existence of image pairs of color-biased images and images after color-biased images are corrected, in order to facilitate the import of the data set, the corresponding images need to be spliced to form image pairs;
[0023] Step 3: Randomly group the data set
[0024] The image pair dataset is divided into a training set and a test set, with the ratio controlled to 85% training set and 15% test set, and placed in two different folders;
[0025] The image data is then normalized and converted into a TensorFlow dataset containing training batch channels through the dataset import function. After processing, the dataset can be called for subsequent training.
[0026] Furthermore, the generator includes a downsampling network and an upsampling network, wherein layers of the same dimension in the downsampling network and the upsampling network are connected using a "U-Net";
[0027] The downsampling network includes an input layer and multiple downsampling sequential layers, each of which consists of a convolutional layer + a BatchNorm layer + an activation function. The upsampling network includes multiple upsampling sequential layers and an upsampling network, each of which consists of a deconvolutional layer + a BatchNorm layer + a Dropout layer + an activation function.
[0028] Furthermore, the purpose of the discriminator is to form a game relationship with the generator, so the discriminator has two inputs, one of which is the image generated by the generator, and the other is the image under the corresponding standard light source in the dataset, that is, the ground-truth image; the discriminator first connects the two input images, and then passes through multiple sequential layers composed of convolution layer + BatchNorm layer + activation function, and then passes through two edge filling operations, and finally uses the convolution operation to output the image data as a patch image.
[0029] Furthermore, the BatchNorm layer in the generator is generated by the Batch normalization function.
[0030] Furthermore, the BatchNorm layer in the discriminator is generated by the Batch normalization function
[0031] Furthermore, the total loss function is a linear combination of the adversarial loss, L1 loss, and La loss of the generator and discriminator;
[0032] The optimization objective of the conditional generative adversarial network can be summarized as:
[0033]
[0034] Among them, the generator G attempts to minimize the goal in the confrontation, while the discriminator D attempts to maximize the goal in the confrontation, so the objective function can be written as x is the input image, y is the target image, z is random noise, Indicates the optimization extreme value that needs to be performed under certain input, represents the adversarial loss between the generator and the discriminator;
[0035] The L1 loss function is introduced into the model. The reason why the L1 loss function is introduced instead of the L2 loss function is that the L1 loss function can reduce the blurriness of the generated image and is more conducive to generating clear images. The L1 loss can be expressed as follows:
[0036]
[0037] Where x is the input image, represents the optimal extreme value, represents L1 loss;
[0038] In addition to the two loss functions mentioned above, a new loss term - angle error loss is added. The new loss term optimizes the color deviation in the color constancy calculation. The loss term is recorded as La and the calculation method is as follows:
[0039]
[0040] ε ij is the one-way pixel angle error, and the average value is La=mean(ε ij ), the reason for introducing La is that its calculation formula, namely cosine similarity, is often used to evaluate the effect of color constancy calculation model. Introducing this loss term to optimize the model can make the model have a better effect on color cast removal; e ij is the estimated ratio of the pixel scene light source, e * ij It is the ratio of the pixel light source estimation of the predicted image after training, which is the ratio of the corresponding pixel points of the input image and the predicted image;
[0041] Therefore, the total loss function of this model is a linear combination of the adversarial loss, L1 loss and La loss of the generator and discriminator, recorded as:
[0042]
[0043] Among them, λ1 is the weight coefficient of L1 loss, λ a is the weight coefficient of the new loss term.
[0044] Compared with the prior art, the advantages and beneficial effects of the present invention are: introducing a generative adversarial network into color constancy calculation, constructing an angle error loss function that reflects hue information, optimizing the network structure hierarchy with reference to the Pix2Pix model idea, improving the model's ability to remove color cast in complex light source scenes, and obtaining a color cast removal result with good stability and excellent processing performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 Downsample the network structure for the generator;
[0046] Figure 2 Upsampling network structure for the generator;
[0047] Figure 3 is the discriminator network structure;
[0048] Figure 4 Remove model structure for color cast;
[0049] Figure 5 The irradiance diagram of the simulated complex light source and the irradiance diagram of the standard light source, where (a) is the irradiance diagram of the simulated complex light source, and (b) is the irradiance diagram of the standard light source;
[0050] Figure 6 Schematic diagram of color cast correction using the method of the present invention, wherein (a) is a color cast image, (b) is a true reference image, and (c) is a corrected image. DETAILED DESCRIPTION
[0051] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.
[0052] The present invention can use a computer to train and infer the network, and use the Tensorflow deep learning framework under the Windows operating system. The specific experimental environment configuration is as follows:
[0053] GPU NVIDIA GeForce RTX 2060SUPER CPU AMD Ryzen 5 3600 6-Core Processor programming language Python 3.7 Deep Learning Frameworks TensorFlow-GPU 2.1.0
[0054] The present invention mainly includes two parts: (1) using the SFU Grayball dataset for color constancy calculation as a training set, and artificially adding simulated complex light sources to expand this dataset into an experimental dataset suitable for training uniform light source and non-uniform light source color constancy models.
[0055] (2) A color cast removal model based on generative adversarial network was constructed. The "U-Net" connection was used as the generator structure of the model, the Markov discriminator was used as the discriminator structure, the Leaky Relu activation function was used instead of the Relu activation function, the Batch Normalization (BN) layer and the Dropout() function were added, and a loss function that can better reflect the hue information was introduced to optimize the training model and improve the effect of the model on color cast removal under different light sources.
[0056] The specific implementation is as follows:
[0057] Step 1: Build a dataset.
[0058] The training dataset used in the experiment is the SFU Grayball dataset for color constancy calculation. The dataset contains 11,346 images taken in the real world. These images are taken from many different scenes, including indoors and outdoors, shopping malls, supermarkets, and squares. The images are rich in information, including people and various daily necessities. Before training, in order to expand the dataset and simulate color-shifted images under complex light sources, the image data was preprocessed as follows.
[0059] (1) Construction of color shift dataset under complex light sources
[0060] The calculation of color constancy under complex light sources needs to solve the problem of insufficient number of training set samples. Since the number of existing complex light source data sets is not enough to train deep models, the present invention, based on the SFU Grayball data set, artificially adds simulated complex light sources to modify this data set into a data set suitable for training complex light source color constancy models.
[0061] Since complex light source scenes can be considered as the result of multiple light sources illuminating the same scene together, and each light source has its own chromaticity value, a Gaussian distribution map generated with random parameters can be used to simulate sufficiently complex light source situations. This situation can be regarded as a random combination of three different color light sources. Then, the standard illumination map of the SFU Grayball dataset is colored using the inverse vonKries transform to obtain a simulated image dataset under complex light source conditions. Figure 5 shown.
[0062] (2) Image cropping and stitching
[0063] Since each image in the original data set contains a small gray ball, the ball will affect the network's information extraction and easily cause the network to "cheat", that is, the generator can easily generate a specific image directly based on the gray ball. Therefore, the image needs to be cropped to remove the ball before processing the data set.
[0064] In addition, since network training requires the existence of image pairs with color bias and those after correcting the color bias, in order to facilitate the import of the data set, the corresponding images need to be spliced to form image pairs.
[0065] (3) Random grouping of data sets
[0066] The dataset is divided into training and test sets, with the ratio controlled to 85% training set and 15% test set, and placed in two different folders.
[0067] The image data is then normalized and converted into a TensorFlow dataset containing training batch channels through the dataset import function. After processing, the dataset can be called for the next step of training.
[0068] Step 2: Use “U-Net” connections for layers of the same dimension in the generator.
[0069] The generator consists of two parts: downsampling and upsampling.
[0070] The generator downsampling structure is as follows Figure 1 As shown in the figure, in the downsampling network structure, the first layer receives the input of the image, which is a three-channel color image of 256*256*3, and then passes it to the network structure composed of eight sequential layers in sequence. Each sequential layer is composed of convolution layer + BatchNorm layer + activation function. Each sequential layer receives the image data of the upper layer and then performs a convolution operation with a step size of 2 and passes it to the next layer. When the image data reaches the last sequential layer of the downsampling network, it has been compressed to a size of 1*1*512 and passed to the upsampling network.
[0071] The generator upsampling network structure is as follows Figure 2 As shown in the figure, the upsampling network structure consists of seven sequential layers plus the final deconvolution layer, where each sequential layer consists of a deconvolution layer + BatchNorm layer + Dropout layer + activation function. Each sequential layer performs a deconvolution operation with a step size of 2 to reduce the convolution thickness and increase the image size. In addition, the output of each sequential layer is connected to the corresponding layer in the downsampling network structure according to the convolution thickness ("U-Net" connection). After the connection operation, the convolution thickness of each sequential layer output data becomes twice the original and is passed to the next layer for the same processing operation. When passed through the final deconvolution layer, the image data is restored to an image of the same size as the generator input image (265*256*3) through deconvolution, thus completing the image generation work of the generator.
[0072] The network structure in the "U-Net" connection does not involve any fully connected layers. At the same time, the feature information generated by the downsampling process can be used in the upsampling process. This allows the network to capture shallow image features in deep convolution operations, making the convolution input richer, so that the image information generated by the generator can better reflect the original information of the image.
[0073] Step 3: Use the Markov discriminator as the discriminator network structure, such as Figure 3 shown.
[0074] The purpose of the discriminator is to form a game relationship with the generator. Therefore, the discriminator has two inputs, one of which is the image generated by the generator, and the other is the image under the corresponding standard light source in the data set, that is, the ground-truth image. Both input images are 256*256*3. The discriminator performs convolution downsampling on these two inputs to obtain the features between the two images for subsequent true and false discrimination. The discriminator network first connects the two input images, then passes through three sequential layers consisting of convolution layer + BatchNorm layer + activation function, and then passes through two edge filling operations. The final convolution operation outputs the image data as a patch image.
[0075] Step 4: Use the Leaky Relu activation function in the network structure of the generator and discriminator.
[0076] The output of the leaky relu function has a small slope for negative inputs. Since the derivative is always non-zero, this can reduce the appearance of invalid neurons, allowing gradient-based learning and solving the problem of neurons not learning after the Relu function enters the negative interval.
[0077] Step 5: When building the model, use Batch normalization and Dropout functions to optimize model performance and prevent network overfitting.
[0078] The batch normalization function is used to add the BatchNorm layer. The BatchNorm layers in the generator and discriminator are generated by the batch normalization function. The dropout function is used in the generator upsampling network structure. The dropout function is called for each upsampling operation.
[0079] Batch normalization uses certain normalization methods to force the distribution of input values of any neuron in each layer of the neural network back to the standard normal distribution, so that the activation input value falls in the area where the nonlinear function is more sensitive to the input. In this way, small changes in the input will lead to larger changes in the loss function, making the gradient larger, avoiding gradient vanishing, and speeding up training.
[0080] Dropout solves the overfitting problem by randomly removing feature detectors.
[0081] Step 6: Determine the loss function and the model structure. The color cast removal model structure is as follows: Figure 4 As shown, where I is the input image, I * The prediction graph generated by the generator, I C is the standard illumination image (ground-truth Image) after removing the color bias, and e is the scene light source estimation provided by the dataset author. Both of the above data are provided by the dataset author. * It is the light source estimation of the predicted image after each training, which is used to calculate the angle error and optimize the model. “ / ” indicates pixel-by-pixel calculation.
[0082] In the model, the jump-connected generator is first used to generate a prediction image based on the input image, and the prediction image is passed to the discriminator, which makes a true or false judgment. After this process, the generation and discrimination losses are calculated, and the L1 loss and angle error loss are calculated for the prediction image and the standard illumination image. The loss is multiplied by the weight and then linearly combined as the total loss of the model. The total loss function is a linear combination of the adversarial loss, L1 loss and La loss of the generator and discriminator, as follows;
[0083] The optimization objective of the conditional generative adversarial network can be summarized as:
[0084]
[0085] Among them, the generator G attempts to minimize the goal in the confrontation, while the discriminator D attempts to maximize the goal in the confrontation, so the objective function can be written as x is the input image, y is the target image, z is random noise, Indicates the optimization extreme value that needs to be performed under certain input, represents the adversarial loss between the generator and the discriminator.
[0086] The L1 loss function is introduced in the model of this study. The reason for introducing the L1 loss function instead of the L2 loss function is that the L1 loss function can reduce the blurriness of the generated image and is more conducive to generating clear images. The L1 loss can be expressed as follows:
[0087]
[0088] Where x is the input image, represents the optimal extreme value, represents L1 loss.
[0089] In addition to the two loss functions mentioned above, a new loss term - angle error loss is added. The new loss term optimizes the color deviation in the color constancy calculation. The loss term is recorded as La and the calculation method is as follows:
[0090]
[0091] ε ij is the one-way pixel angle error, and the average value is La=mean(ε ij ). The reason for introducing La is that its calculation formula, namely cosine similarity, is often used to evaluate the effect of color constancy calculation model. Introducing this loss term to optimize the model can make the model have a better effect on color cast removal. e ij is the estimated ratio of the pixel scene light source, which is provided by the dataset author. * ij It is the estimated ratio of the pixel light source of the predicted image after training, which is the ratio of the corresponding pixel points of the input image and the predicted image.
[0092] Therefore, the total loss function of this model is a linear combination of the generator and discriminator losses, L1 loss and La loss, recorded as:
[0093]
[0094] Among them, λ1 is the weight coefficient of L1 loss, λ a is the weight coefficient of the new loss term.
[0095] Step 7: Train the network model and obtain the model file.
[0096] The number of training rounds is set to 200. After the training is completed, the model file of the image enhancement model based on the generative adversarial network for color cast removal proposed by this method is obtained. The correction results are shown in Figure 6 .
[0097] The specific embodiments described herein are merely examples of the present invention. A person skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner without departing from the spirit of the present invention or exceeding the scope defined by the appended claims.
Claims
1. A method for removing image color cast based on generative adversarial network. It is characterized in that The following steps are involved: Step 1: construct a color-biased image dataset, preprocess the dataset, and form image pairs of images with color bias and images with corrected color bias; Step 2: Build a color cast removal model based on a generative adversarial network, use "U-Net" connection as the generator structure of the color cast removal model, and use Markov discriminator as the discriminator structure; Step 3: Use the Leaky Relu activation function instead of the Relu activation function in the network structure of the generator and the discriminator; Step 4, designing a corresponding total loss function, using the image in step 1 to train the color cast removal model constructed in step 2, and obtaining a model file; The total loss function is a linear combination of the adversarial loss, L1 loss, and La loss of the generator and discriminator; The optimization objective of the conditional generative adversarial network can be summarized as: Among them, the generator G attempts to minimize the goal in the confrontation, while the discriminator D attempts to maximize the goal in the confrontation, so the objective function can be written as x is the input image, y is the target image, z is random noise, Indicates the optimization extreme value that needs to be performed under certain input, represents the adversarial loss between the generator and the discriminator; The L1 loss function is introduced into the model. The reason why the L1 loss function is introduced instead of the L2 loss function is that the L1 loss function can reduce the blurriness of the generated image and is more conducive to generating clear images. The L1 loss can be expressed as follows: Where x is the input image, represents the optimal extreme value, represents L1 loss; In addition to the two loss functions mentioned above, a new loss term - angle error loss is added. The new loss term optimizes the color deviation in the color constancy calculation. The loss term is recorded as La and the calculation method is as follows: ε ij is the one-way pixel angle error, and the average value is La=mean(ε ij ), the reason for introducing La is that its calculation formula, namely cosine similarity, is often used to evaluate the effect of color constancy calculation model. Introducing this loss term to optimize the model can make the model have a better effect on color cast removal; e ij is the estimated ratio of the pixel scene light source, e * ij It is the ratio of the pixel light source estimation of the predicted image after training, which is the ratio of the corresponding pixel points of the input image and the predicted image; Therefore, the total loss function of the color cast removal model is a linear combination of the adversarial loss, L1 loss, and La loss of the generator and discriminator, recorded as: Among them, λ 1 is the weight coefficient of L1 loss, λ a is the weight coefficient of the new loss term; Step 5: Use the model file to remove color cast from the test image.
2. According to the method for removing color cast of an image based on a generative adversarial network as described in claim 1, Features: The dataset used in step 1 is the SFU Grayball dataset for color constancy calculation. Before training, in order to expand the dataset and simulate color-shifted images under complex light sources, the image data was preprocessed as follows: Step 11: Color cast dataset construction under complex light sources Since complex light source scenes can be considered as the result of multiple light sources illuminating the same scene together, and each light source has its own chromaticity value, a Gaussian distribution map generated with random parameters is used to simulate sufficiently complex light source situations. This situation can be regarded as a random combination of three light sources of different colors. Then, the standard illumination map of the SFU Grayball dataset is colored using the inverse von Kries transform to obtain a simulated image dataset under complex light source conditions. Step 2: Image cropping and stitching Since each image in the original data set contains a small gray ball, the ball will affect the network's information extraction and easily cause the network to "cheat", that is, the generator can easily generate specific images directly based on the gray ball. Therefore, the image needs to be cropped to remove the ball before processing the data set. In addition, since the model training requires the existence of image pairs of color-biased images and images after color-biased images are corrected, in order to facilitate the import of the data set, the corresponding images need to be spliced to form image pairs; Step 3: Randomly group the data set The image pair dataset is divided into a training set and a test set, with the ratio controlled to 85% training set and 15% test set, and placed in two different folders; The image data is then normalized and converted into a TensorFlow dataset containing training batch channels through the dataset import function. After processing, the dataset can be called for subsequent training.
3. According to the method for removing color cast of an image based on a generative adversarial network as described in claim 1, Features: In step 2, the generator includes a downsampling network and an upsampling network, where the layers with the same dimensions in the downsampling network and the upsampling network are connected using "U-Net"; The downsampling network includes an input layer and multiple downsampling sequential layers, each of which consists of a convolutional layer + a BatchNorm layer + an activation function. The upsampling network includes multiple upsampling sequential layers and an upsampling network, each of which consists of a deconvolutional layer + a BatchNorm layer + a Dropout layer + an activation function.
4. According to the image color cast removal model based on generative adversarial network described in claim 1, Features: The purpose of the discriminator in step 2 is to form a game relationship with the generator. Therefore, the discriminator has two inputs, one of which is the image generated by the generator, and the other is the image under the corresponding standard light source in the dataset, that is, the ground-truth image; the discriminator first connects the two input images, and then passes through multiple sequential layers consisting of convolution layer + BatchNorm layer + activation function, and then passes through two edge filling operations, and finally uses the convolution operation to output the image data as a patch image.
5. According to the method for removing color cast of an image based on a generative adversarial network as described in claim 3, Features: The BatchNorm layer in the generator is generated by the Batch normalization function.
6. According to the method for removing color cast of an image based on a generative adversarial network as described in claim 4, Features: The BatchNorm layer in the discriminator is generated by the Batch normalization function.
