CN115601644A - Power transmission line image enhancement method under low illumination based on generation countermeasure network - Google Patents

Abstract

The invention discloses a method for enhancing an image of a power transmission line under low illumination based on a generated countermeasure network, which is characterized by comprising the following steps: the invention designs a residual error module based on a mixed attention mechanism, and designs a generation network capable of extracting more effective characteristic information by combining a parallel cavity convolution module and the like so as to enhance the image of the power transmission line under low illumination. Secondly, the invention also designs a countermeasure network of a double-discrimination network based on a global discrimination network and a local discrimination network, and improves the discrimination capability of the countermeasure network on the input image. Finally, the invention also designs a loss function of the power transmission line image enhancement network under low illuminance based on the generated countermeasure network. The invention can effectively improve the image brightness of the power transmission line under low illuminance, simultaneously avoid the occurrence of overexposure or underexposure of the enhanced image and the occurrence of artifact phenomenon, retain more image detail information and improve the image quality of the enhanced power transmission line.

Description

A Method of Image Enhancement of Transmission Lines under Low Illumination Based on Generative Adversarial Network

technical field

The invention belongs to the technical field of image enhancement under low light illumination, and relates to a method for image enhancement of transmission lines under low light illumination based on a generative confrontation network.

Background technique

The important transmission lines above the 220kV voltage level in my country have basically deployed online monitoring equipment based on image processing. The cameras installed on the high-voltage transmission towers collect images of the transmission lines and surrounding areas, and then detect broken strands, hanging foreign objects, and overturned transmission lines. Analyze ice, external force damage, and the status of power transmission equipment, discover hidden dangers and deal with them in a timely manner. However, when encountering cloudy days with low illuminance, the images collected on site are not clear, which cannot effectively reflect the actual situation of the transmission line site, which affects the online monitoring of the transmission line. For this reason, it is necessary to improve the clarity of online monitoring images of transmission lines under low light illumination, which is conducive to improving the accuracy of online monitoring of transmission lines under low light illumination.

The image enhancement methods of transmission lines under low-light illumination can be divided into two categories. One category is to use traditional methods to enhance transmission line images under low-light illumination, and the other category is to enhance low-light images based on learning methods. Traditional methods can be divided into two directions, one of which is the histogram equalization method, and the other is the image enhancement method based on Retinex theory. The low-light image enhancement method of transmission line based on histogram equalization not only enhances the image, but also increases the background noise, and also easily causes local oversaturation of the image and serious loss of local information. The low-light image enhancement method of transmission lines based on Retinex theory will produce halo phenomenon in the enhanced image in areas with large brightness differences, which will affect the visual effect of the image. In addition, there is insufficient edge sharpening, abrupt shadow borders, some colors are distorted, and textures are not clear. The traditional image enhancement method is relatively simple and fast, but it does not take into account the context information in the image, so the effect of the enhanced image is not ideal.

With the development of deep learning technology, people have proposed low-light image enhancement methods for transmission lines based on deep learning. These methods use pairs (low-light images and original high-definition images) or unpaired transmission line images to train the network and obtain A model that can be used for low-light image enhancement of transmission lines. Although, compared with the traditional image enhancement method, the image quality of the transmission line image enhanced by the deep learning-based low-light image enhancement method of the transmission line is relatively better, but to a certain extent, there are still serious artifacts and loss of some details. It affects the image quality of the enhanced transmission line, and then affects the accuracy of online monitoring of transmission lines based on artificial intelligence in low-light environments.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for image enhancement of transmission lines under low-light illumination based on generative adversarial networks, which can effectively enhance a single image of transmission lines under low-light illumination, and avoid artifacts and artifacts in the enhanced image. Preserve more image detail information and improve the image quality of the enhanced transmission line.

The solution of the present invention to solve the technical problem is to provide a method for image enhancement of transmission lines under low light illumination based on generative confrontation network, which is characterized in that the specific steps are as follows:

1) Build a dataset

By selecting low-light images and normal-light images with different backgrounds from the transmission line video surveillance system and processing them, an unpaired sample training set and a paired sample test set are constructed;

2) Build a generative network

Constructing a network for augmenting images of transmission lines in low-light conditions;

3) Build a confrontation network

Construct a network for judging whether an input transmission line image is a real image or a fake image;

Define the transmission line image under the original normal light illumination as the true image;

Define the output image of the generative network as a fake image;

4) Construct the loss function of the generated confrontation network

Used to measure the performance of the generated network and the performance of the confrontation network during the network training process;

5) Network model training

The optimal generation network and confrontation network are obtained through network model training;

6) Network model performance evaluation

Input the low-light images in the paired sample test set constructed in step 1 to the trained generation network obtained in step 5 to obtain enhanced transmission line images to measure the enhancement of the network model to transmission line images under low-light illumination ability;

7) Network model application

Deploy the trained network model on the server, and enhance the low-light transmission line images sent back from the site to obtain enhanced images.

Further, said step 1) constructing a data set includes the following steps:

(1) Select low-light images and normal-light images with different backgrounds from the transmission line video monitoring system, name the low-light images as the original low-light group, and name the normal-light images as the normal group;

(2) Divide the images of the normal group into a normal group and a normal group, and use the original low-light group images and the normal group to construct an unpaired sample training set;

(3) Process the images of the normal two groups to obtain the corresponding low-light images, which are named as processing low-light groups;

(4) Construct a paired sample test set with the images of the normal group and the processed low-light group, so as to construct the unpaired sample training set and the paired sample test set.

Further, the images of the normal group are divided into the normal group 1 and the normal group 2, the images of the normal group 1 are 70% of the images of the normal group, and the images of the normal group 2 are 30% of the images of the normal group.

Described step 2) in constructing generating network, generating network is made up of A network, B network and C network, specifically as follows:

(1) The A network is used to preprocess the input low-light transmission line image, which is composed of brightness attention mapping to reduce the occurrence of overexposure or underexposure of the generated image. The brightness attention map obtained after preprocessing and the input transmission line The low-light illumination images of the lines are added and used as the input image of the B network;

(2) The B network is used to extract the low-frequency information of the input transmission line low-light illumination image, which consists of a convolution module, a LeakyRelu activation function, the first combination module, the second combination module and a residual module based on a mixed attention mechanism, where :

①The first combination module and the second combination module are composed of the first branch and the second branch, and the output feature maps of the first branch and the second branch of each combination module are obtained by splicing operations to obtain the output feature map of the combination module ;

②The residual module based on the mixed attention mechanism consists of a convolution module, a LeakyRelu activation function, a convolution module, a parallel attention module, a convolution module, a LeakyRelu activation function, a convolution module, Another parallel attention module and a convolution module;

③The input feature map of the residual module based on the mixed attention mechanism described in step ② is added element-wise to the output feature map of the last convolution module in the residual module based on the mixed attention mechanism, and the result is obtained based on The final output feature map of the residual module of the hybrid attention mechanism;

The CDC network is mainly used to extract the high-frequency information of the input image, which consists of the first mixing module, a residual module based on the mixed attention mechanism, the second mixing module, a residual module based on the mixed attention mechanism, and a convolution module and a Tanh activation function, where:

Both the first mixing module and the second mixing module are composed of an upsampling, a convolution module and a LeakyRelu activation function;

(4) The output of the first mixing module of the C network and the output of the first combination module of the B network perform feature fusion through a splicing operation, and the output of the second mixing module of the C network and the input of the first combination module of the B network perform feature fusion through a splicing operation Fusion, so as to realize the fusion of high-frequency features and low-frequency features;

(5) The output of the C network and the brightness attention map image obtained by the A network are multiplied elementwise to obtain an image;

(6) Add the image obtained in step (5) to the low-light image input by the generation network to obtain the enhanced transmission line image.

Further, the first combination module and the second combination module of the B network are composed of a first branch and a second branch, wherein:

The first branch consists of parallel atrous convolution modules;

The second branch consists of a residual module based on a hybrid attention mechanism and a downsampling.

Further, the two parallel attention modules in the residual module based on the hybrid attention mechanism are both composed of a channel attention module and a pixel attention module.

Further, the construction of the confrontation network in the step 3) is expressed as:

The confrontation network consists of a global discriminative network and a local discriminative network, where:

The global discriminative network is sequentially composed of 2 combination modules, a residual dilated convolution module and 2 combination modules, and its input images are fake images and real images;

The local discriminative network consists of 6 combination modules whose input images are randomly cropped images of fake and real images.

Further, the combination module and the residual hole convolution module of the global discriminant network are:

The combination module of the global discriminant network is composed of convolution+LeakyReLU activation function in turn;

The main branch of the residual atrous convolution module of the global discriminant network consists of three atrous convolutions with atrous rates of 2, 3 and 5, respectively.

Further, the combination module of the local discriminant network is a network module based on convolution+LeakyReLU activation function.

Further, the loss function of constructing generation confrontation network in said step 4) is expressed as:

和

分别是对抗网络的全局判别网络的损失函数和局部判别网络的损失函数，L_Per和L_Pix分别是生成网络的感知损失函数和像素损失函数，α，β，γ和ω分别是上述对应损失函数的权重。In the formula,

and

are the loss function of the global discriminant network and the loss function of the local discriminant network of the confrontation network, respectively, L _Per and L _Pix are the perceptual loss function and pixel loss function of the generative network, respectively, α, β, γ and ω are the above-mentioned corresponding loss functions the weight of.

Further, the loss function of the global discrimination network, the loss function of the local discrimination network, the perception loss function and the pixel loss function of the generation network of the confrontation network are:

The loss function of the global discriminant network is expressed as:

In the formula, D ^G is the global discriminant network, G is the generation network, z and x are the input image of the generation network and the input image of the confrontation network, respectively;

The loss function of the local discriminant network is expressed as:

In the formula, D ^L is the local discriminant network;

The perceptual loss function of the generative network is expressed as:

是在VGG-19预训练网络；In the formula, x is the input image of the generator network, W and H are the width and height of the image respectively,

is the VGG-19 pre-trained network;

The pixel loss function of the generative network is expressed as:

Further, the step 5) model training is specifically as follows:

(1) Input the low-light transmission line image in the training set into the generation network to obtain the enhanced image;

(2) Input the enhanced image and the normal illumination image in the training set into the confrontation network, and judge whether the input image is the enhanced image or the original normal illumination image;

(3) By using the gradient descent method to optimize the loss function, continuously update the generation network and loss network parameters, and finally complete the training of the generation network and the confrontation network, and obtain the optimal generation network and confrontation network.

Further, the step 6) network model performance evaluation is specifically as follows:

(1) Input the low-light images in the paired sample test set constructed in step 1 to the trained generation network obtained in step 5 to obtain enhanced transmission line images to measure the performance of the network model on transmission line images under low light gain power;

(2) Calculate the structural similarity SSIM and peak signal-to-noise ratio PSNR between the enhanced image and the corresponding normal illumination image in the test set, respectively, to measure the ability of the network model to enhance the transmission line image under low light;

(3) If the average SSIM or average PSNR value is low, adjust the network model parameters and continue training. When the average SSIM and average PSNR value reaches above the ideal value, keep the network model weight, which is used to enhance the transmission line image under low light.

The present invention provides a method for image enhancement of power transmission lines under low light illumination based on generative confrontation network, which can effectively improve the brightness of power transmission line images under low light illumination while avoiding overexposure or underexposure and artifacts in the enhanced image appear, retain more image detail information, and improve the image quality of the enhanced transmission line.

Description of drawings

Fig. 1 is a kind of flow chart of the transmission line image enhancement method under the low-light illumination based on generation confrontation network of the present invention;

Fig. 2 is a structural diagram of the generation network and the parallel dilated convolution module of the present invention;

Fig. 3 is a structural diagram of the residual module based on the mixed attention mechanism of the present invention;

Fig. 4 is a structural diagram of the confrontation network of the present invention.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Referring to Fig. 1-Fig. 4, embodiment 1, this embodiment provides a method for image enhancement of transmission lines under low light illumination based on generative confrontation network, and the specific steps are as follows:

1) Build a dataset

By selecting low-light images and normal-light images of different backgrounds from the transmission line video surveillance system and processing them, an unpaired sample training set and a paired sample test set are constructed. The steps are as follows:

(1) Select low-light images and normal-light images with different backgrounds from the transmission line video monitoring system, name the low-light images as the original low-light group, and name the normal-light images as the normal group;

(2) Divide the images of the normal group into a normal group and a normal group two. The images of the normal group are 70% of the images of the normal group, and the images of the second group of normal are 30% of the images of the normal group. image and a normal set of images to construct an unpaired sample training set;

(3) Process the images of the normal two groups to obtain the corresponding low-light images, which are named as processing low-light groups;

(4) Use the images of the normal two groups and the images of the processed low-light group to construct a paired sample test set, thereby constructing an unpaired sample training set and a paired sample test set.

Low-light images and normal-light images with different backgrounds are selected from the transmission line video surveillance system, and the image pixel size is adjusted to 600×400.

2) Build a generative network

Constructing a network for augmenting images of transmission lines in low-light conditions;

3) Build a confrontation network

Construct a network for judging whether an input transmission line image is a real image or a fake image;

Define the transmission line image under the original normal light illumination as the true image;

Define the output image of the generative network as a fake image;

4) Construct the loss function of the generated confrontation network

Used to measure the performance of the generated network and the performance of the confrontation network during the network training process;

5) Network model training

The details of model training are as follows:

(2) Input the low-light transmission line image in the training set into the generation network to obtain the enhanced image;

(2) Input the enhanced image and the normal illumination image in the training set into the confrontation network, and judge whether the input image is the enhanced image or the original normal illumination image;

(3) By using the gradient descent method to optimize the loss function, continuously update the generation network and loss network parameters, and finally complete the training of the generation network and the confrontation network, and obtain the optimal generation network and confrontation network.

6) Network model performance evaluation

(1) Input the low-light images in the paired sample test set constructed in step 1 to the trained generation network obtained in step 5 to obtain enhanced transmission line images to measure the performance of the network model on transmission line images under low light gain power;

(2) Calculate the structural similarity SSIM and peak signal-to-noise ratio PSNR between the enhanced image and the corresponding normal illumination image in the test set, respectively, to measure the ability of the network model to enhance the transmission line image under low light;

(3) If the average SSIM or average PSNR value is low, adjust the network model parameters and continue training. When the average SSIM and average PSNR value reaches above the ideal value, keep the network model weight, which is used to enhance the transmission line image under low light.

7) Network model application

Deploy the trained network model on the server, and enhance the low-light transmission line images sent back from the site to obtain enhanced images.

Described step 2) in constructing generating network, generating network is made up of A network, B network and C network, specifically as follows:

(1) The A network is used to preprocess the input low-light transmission line image, which is composed of a brightness attention map to reduce the occurrence of overexposure or underexposure of the generated image. The brightness attention map obtained after preprocessing and the input transmission line The low-light illumination images of the lines are added and used as the input image of the B network;

(2) The B network is used to extract the low-frequency information of the input transmission line low-light illumination image, which consists of a convolution module, a LeakyRelu activation function, a first combination module, a second combination module and a residual module based on a mixed attention mechanism, where :

① The first combination module and the second combination module are composed of the first branch and the second branch, and the output feature maps of the first branch and the second branch of each combination module are obtained by splicing operations to obtain the output feature map of the combination module ;

②The residual module based on the mixed attention mechanism consists of a convolution module, a LeakyRelu activation function, a convolution module, a parallel attention module, a convolution module, a LeakyRelu activation function, a convolution module, Another parallel attention module and a convolution module;

③The input feature map of the residual module based on the mixed attention mechanism described in step ② is added element-wise to the output feature map of the last convolution module in the residual module based on the mixed attention mechanism, and the result is obtained based on The final output feature map of the residual module of the hybrid attention mechanism;

The CDC network is mainly used to extract the high-frequency information of the input image, which consists of the first mixing module, a residual module based on the mixed attention mechanism, the second mixing module, a residual module based on the mixed attention mechanism, and a convolution module and a Tanh activation function, where:

Both the first mixing module and the second mixing module are composed of an upsampling, a convolution module and a LeakyRelu activation function;

(4) The output of the first mixing module of the C network and the output of the first combination module of the B network perform feature fusion through a splicing operation, and the output of the second mixing module of the C network and the input of the first combination module of the B network perform feature fusion through a splicing operation Fusion, so as to realize the fusion of high-frequency features and low-frequency features;

(5) The output of the C network and the brightness attention map image obtained by the A network are multiplied elementwise to obtain an image;

(6) Add the image obtained in step (5) to the low-light image input by the generation network to obtain the enhanced transmission line image.

Both the first combination module and the second combination module of the B network are composed of a first branch and a second branch, wherein:

The first branch consists of parallel atrous convolution modules;

The second branch consists of a residual module based on a hybrid attention mechanism and a downsampling.

The two parallel attention modules in the residual module based on the hybrid attention mechanism are both composed of a channel attention module and a pixel attention module.

Said step 3) constructing the confrontation network is expressed as:

The confrontation network consists of a global discriminative network and a local discriminative network, where:

The global discriminative network is sequentially composed of 2 combination modules, a residual dilated convolution module and 2 combination modules, and its input images are fake images and real images;

The local discriminative network consists of 6 combination modules whose inputs are randomly cropped images of fake and real images.

The combination module and residual hole convolution module of the global discriminant network are:

The combination module of the global discriminant network is composed of convolution+LeakyReLU activation function in turn;

The main branch of the residual atrous convolution module of the global discriminant network consists of three atrous convolutions with atrous rates of 2, 3 and 5, respectively.

The combination module of the local discriminant network is a network module based on convolution+LeakyReLU activation function.

The loss function of constructing generation confrontation network in described step 4) is expressed as:

和

分别是对抗网络的全局判别网络的损失函数和局部判别网络的损失函数，L_Per和L_Pix分别是生成网络的感知损失函数和像素损失函数，α，β，γ和ω分别是上述对应损失函数的权重。In the formula,

and

are the loss function of the global discriminant network and the loss function of the local discriminant network of the confrontation network, respectively, L _Per and L _Pix are the perceptual loss function and pixel loss function of the generative network, respectively, α, β, γ and ω are the above-mentioned corresponding loss functions the weight of.

The loss function of the global discrimination network of the confrontation network, the loss function of the local discrimination network, the perception loss function and the pixel loss function of the generation network are:

The loss function of the global discriminant network is expressed as:

In the formula, D ^G is the global discriminant network, G is the generation network, z and x are the input image of the generation network and the input image of the confrontation network, respectively;

The loss function of the local discriminant network is expressed as:

In the formula, D ^L is the local discriminant network;

The perceptual loss function of the generative network is expressed as:

是在VGG-19预训练网络；In the formula, x is the input image of the generator network, W and H are the width and height of the image respectively,

is the VGG-19 pre-trained network;

The pixel loss function of the generative network is expressed as:

The convolution modules described in this embodiment all have the same structure.

The residual modules based on the hybrid attention mechanism described in this embodiment are composed of the same structure.

The parallel attention modules described in this embodiment all have the same structure.

The LeakyRelu activation functions described in this embodiment all have the same structure.

The combined modules described in this embodiment have the same structure.

This embodiment is realized by using the prior art.

Claims (9)

1. A transmission line image enhancement method under low light illumination based on generation confrontation network, it is characterized in that, concrete steps are as follows: 1) Build a dataset By selecting low-light images and normal-light images with different backgrounds from the transmission line video surveillance system and processing them, an unpaired sample training set and a paired sample test set are constructed; 2) Build a generative network Constructing a network for augmenting images of transmission lines in low-light conditions; 3) Build a confrontation network Construct a network for judging whether an input transmission line image is a real image or a fake image; Define the transmission line image under the original normal light illumination as the true image; Define the output image of the generative network as a fake image; 4) Construct the loss function of the generated confrontation network Used to measure the performance of the generated network and the performance of the confrontation network during the network training process; 5) Network model training The optimal generation network and confrontation network are obtained through network model training; 6) Network model performance evaluation Input the low-light images in the paired sample test set constructed in step 1 to the trained generation network obtained in step 5 to obtain enhanced transmission line images to measure the enhancement of the network model to transmission line images under low-light illumination ability; 7) Network model application Deploy the trained network model on the server, and enhance the low-light transmission line images sent back from the site to obtain enhanced images. 2. the power transmission line image enhancement method under low light illumination based on generation confrontation network as claimed in claim 1, is characterized in that, described step 2) in constructing generation network, generation network is made up of A network, B network and C network, details as follows: (1) The A network is used to preprocess the input low-light transmission line image, which is composed of brightness attention mapping to reduce the occurrence of overexposure or underexposure of the generated image. The brightness attention map obtained after preprocessing and the input transmission line The low-light illumination images of the lines are added and used as the input image of the B network; (2) The B network is used to extract the low-frequency information of the input transmission line low-light illumination image, which consists of a convolution module, a LeakyRelu activation function, the first combination module, the second combination module and a residual module based on a mixed attention mechanism, where : ①The first combination module and the second combination module are composed of the first branch and the second branch, and the output feature maps of the first branch and the second branch of each combination module are obtained by splicing operations to obtain the output feature map of the combination module ; ②The residual module based on the mixed attention mechanism consists of a convolution module, a LeakyRelu activation function, a convolution module, a parallel attention module, a convolution module, a LeakyRelu activation function, a convolution module, Another parallel attention module and a convolution module; ③ Add the input feature map of the residual module based on the mixed attention mechanism described in step ② to the output feature map of the last convolution module in the residual module based on the mixed attention mechanism, and obtain a mixture based on The final output feature map of the residual module of the attention mechanism; The CDC network is mainly used to extract the high-frequency information of the input image, which consists of the first mixing module, a residual module based on the mixed attention mechanism, the second mixing module, a residual module based on the mixed attention mechanism, and a convolution module and a Tanh activation function, where: Both the first mixing module and the second mixing module are composed of an upsampling, a convolution module and a LeakyRelu activation function; (4) The output of the first mixing module of the C network and the output of the first combination module of the B network perform feature fusion through a splicing operation, and the output of the second mixing module of the C network and the input of the first combination module of the B network perform feature fusion through a splicing operation Fusion, so as to realize the fusion of high-frequency features and low-frequency features; (5) The output of the C network and the brightness attention map image obtained by the A network are multiplied elementwise to obtain an image; (6) Add the image obtained in step (5) to the low-light image input by the generation network to obtain the enhanced transmission line image. 3. the power transmission line image enhancement method under low light illumination based on generation confrontation network as claimed in claim 2, is characterized in that, the first combination module and the second combination module of described B network are all formed by the first branch and the second branch composed of: The first branch consists of parallel atrous convolution modules; The second branch consists of a residual module based on a hybrid attention mechanism and a downsampling. 4. the power transmission line image enhancement method under low light illumination based on generation confrontation network as claimed in claim 2, is characterized in that, The two parallel attention modules in the residual module based on the hybrid attention mechanism are both composed of a channel attention module and a pixel attention module. 5. the power transmission line image enhancement method under the low light illumination based on generation confrontation network as claimed in claim 1, is characterized in that, described step 3) constructs confrontation network and expresses as: The confrontation network consists of a global discriminative network and a local discriminative network, where: The global discriminative network is sequentially composed of 2 combination modules, a residual dilated convolution module and 2 combination modules, and its input images are fake images and real images; The local discriminative network consists of 6 combination modules whose inputs are randomly cropped images of fake and real images. 6. the power transmission line image enhancement method under low light illumination based on generation confrontation network as claimed in claim 5, it is characterized in that, the combination module and the residual hole convolution module of the global discriminant network are: The combined module of the global discriminant network and the combined module of the local discriminant network are all sequentially composed of convolution+LeakyReLU activation functions; The main branch of the residual atrous convolution module consists of three atrous convolutions with atrous rates of 2, 3 and 5, respectively. 7. The image enhancement method of transmission lines under low light illumination based on generative confrontation network as claimed in claim 5, wherein the combination module of the local discriminant network is a network module based on convolution+LeakyReLU activation function. 8. the power transmission line image enhancement method under low light illumination based on generation confrontation network as claimed in claim 1, is characterized in that, described step 4) in constructing the loss function of generation confrontation network is expressed as:

In the formula,

and

are the loss function of the global discriminant network and the loss function of the local discriminant network of the confrontation network, respectively, L _Per and L _Pix are the perceptual loss function and pixel loss function of the generative network, respectively, α, β, γ and ω are the above-mentioned corresponding loss functions the weight of. 9. The transmission line image enhancement method under low light illumination based on generation confrontation network as claimed in claim 8, characterized in that, the loss function of the global discriminant network of the confrontation network, the loss function of the local discriminant network, the perception of the generation network The loss function and the pixel loss function are: The loss function of the global discriminant network is expressed as:

In the formula, D ^G is the global discriminant network, G is the generation network, z and x are the input image of the generation network and the input image of the confrontation network, respectively; The loss function of the local discriminant network is expressed as:

In the formula, D ^L is the local discriminant network; The perceptual loss function of the generative network is expressed as:

In the formula, x is the input image of the generator network, W and H are the width and height of the image respectively,

is the VGG-19 pre-trained network; The pixel loss function of the generative network is expressed as:

Images