A power transmission line image enhancement method under low light intensity based on a generative adversarial network

Abstract

The application discloses a kind of low light intensity under power transmission line image enhancement method based on generative adversarial network, its characteristics are: the application designs residual module based on hybrid attention mechanism, and can extract more effective feature information by designing generation network such as parallel hole convolution module, to enhance the power transmission line image under low light intensity.Secondly, the application also designs the adversarial network of double discriminative network based on global discriminative network and local discriminative network, improves the discriminative ability of input image of adversarial network.Finally, the application also designs the loss function of low light intensity under power transmission line image enhancement network based on the above generative adversarial network.The application can effectively improve the brightness of power transmission line image under low light intensity, while avoiding the appearance of overexposure or underexposure, and the appearance of artifact phenomenon of enhanced image, retaining more image detail information, improving the quality of enhanced power transmission line image.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement technology under low illumination, and relates to a method for enhancing images of power transmission lines under low illumination based on generative adversarial networks. Background Technology

[0002] In my country, most major power transmission lines with voltage levels above 220kV have been equipped with image-processing-based online monitoring equipment. Cameras mounted on high-voltage transmission towers collect images of the transmission lines and their surroundings, allowing for the analysis of issues such as broken strands, hanging foreign objects, icing, external damage, and the condition of transmission equipment. This enables timely detection and handling of potential hazards. However, on cloudy days with low light levels, the images captured are unclear and cannot effectively reflect the actual conditions of the transmission lines, affecting online monitoring. Therefore, it is necessary to improve the clarity of online monitoring images of transmission lines under low light conditions, thereby improving the accuracy of online monitoring in such environments.

[0003] Image enhancement methods for transmission lines under low illumination can be divided into two categories: traditional methods and learning-based methods. Traditional methods can be further divided into histogram equalization (HEM) and Retinex-based image enhancement methods. Histogram equalization-based methods, while enhancing the image, also introduce background noise and are prone to local oversaturation and significant loss of local information. Retinex-based methods produce halos in areas with large brightness differences, affecting the visual quality. Furthermore, they suffer from insufficient edge sharpening, abrupt shadow boundaries, color distortion, and unclear textures. While traditional image enhancement methods are relatively simple and fast, they do not consider contextual information within the image, resulting in less than ideal enhanced image quality.

[0004] With the development of deep learning technology, deep learning-based methods for enhancing low-light images of power transmission lines have been proposed. These methods use paired (low-light image and original high-resolution image) or unpaired power transmission line images to train networks and obtain models that can be used for low-light image enhancement of power transmission lines. Although deep learning-based methods for enhancing low-light images of power transmission lines generally produce higher-quality enhanced images compared to traditional image enhancement methods, artifacts and significant loss of detail still exist to some extent, affecting the quality of the enhanced images and consequently the accuracy of AI-based online monitoring of power transmission lines in low-light environments. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for enhancing transmission line images under low illumination based on generative adversarial networks, which can effectively enhance a single transmission line image under low illumination, avoid artifacts in the enhanced image, retain more image detail information, and improve the quality of the enhanced transmission line image.

[0006] The solution to the technical problem of this invention is: to provide an image enhancement method for transmission lines under low illumination based on generative adversarial networks, characterized by the following specific steps:

[0007] 1) Constructing the dataset

[0008] By selecting low-light and normal-light images with different backgrounds from the power transmission line video monitoring system and processing them, an unpaired sample training set and a paired sample test set were constructed.

[0009] 2) Constructing a generative network

[0010] Construct a network for enhancing images of power transmission lines under low illumination;

[0011] 3) Constructing adversarial networks

[0012] Construct a network to determine whether an input image of a power transmission line is a real image or a fake image;

[0013] The image of the transmission line under normal illumination is defined as the true image;

[0014] The output image of the generator network is defined as a fake image;

[0015] 4) Construct the loss function for generative adversarial networks.

[0016] Used to measure the performance of generative networks and adversarial networks during the network training process;

[0017] 5) Network model training

[0018] The optimal generative network and adversarial network are obtained through network model training;

[0019] 6) Network model performance evaluation

[0020] The low-light images from the paired sample test set constructed in step 1 are input into the trained generative network obtained in step 5 to obtain the enhanced transmission line images, so as to measure the network model's ability to enhance transmission line images under low light conditions.

[0021] 7) Application of network models

[0022] The trained network model is deployed on a server to enhance low-light transmission line images transmitted from the field, resulting in enhanced images.

[0023] Furthermore, step 1) of constructing the dataset includes the following steps:

[0024] (1) Select low-light and normal-light images with different backgrounds from the power 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.

[0025] (2) Divide the images of the normal group into normal group 1 and normal group 2, and construct an unpaired sample training set using the images of the original low-light group and the images of normal group 1;

[0026] (3) The images of the normal two groups are processed to obtain the corresponding low-light images, which are named the processed low-light group;

[0027] (4) Construct a paired sample test set using images from the normal group and images from the low-light group, thereby constructing an unpaired sample training set and a paired sample test set.

[0028] Furthermore, the images in the normal group were divided into Normal Group 1 and Normal Group 2. The images in Normal Group 1 consisted of 70% of the images in the normal group, and the images in Normal Group 2 consisted of 30% of the images in the normal group.

[0029] In step 2), the generator network is constructed, which consists of network A, network B, and network C, as follows:

[0030] (1) Network A is used to preprocess the input low-light transmission line image. It consists of a brightness attention map to reduce the overexposure or underexposure of the generated image. The brightness attention map obtained after preprocessing is added to the input low-light transmission line image and used as the input image of Network B.

[0031] (2) The B network is used to extract low-frequency information from the input low-light image of the transmission line. It consists of a convolutional module, a LeakyReLU activation function, a first combination module, a second combination module, and a residual module based on a hybrid attention mechanism, wherein:

[0032] ① Both the first and second combination modules consist of a first branch and a second branch. The output feature maps of the first and second branches of each combination module are spliced ​​together to obtain the output feature map of that combination module.

[0033] ②The residual module based on the hybrid attention mechanism consists of a convolutional module, a LeakyReLU activation function, a convolutional module, a parallel attention module, a convolutional module, a LeakyReLU activation function, a convolutional module, another parallel attention module, and a convolutional module in sequence.

[0034] ③ Add the input feature map of the residual module based on the hybrid attention mechanism described in step ② to the output feature map of the last convolutional module in the residual module based on the hybrid attention mechanism element by element to obtain the final output feature map of the residual module based on the hybrid attention mechanism.

[0035] (3) The C network is mainly used to extract high-frequency information from the input image. It consists of a first hybrid module, a residual module based on a hybrid attention mechanism, a second hybrid module, another residual module based on a hybrid attention mechanism, a convolutional module, and a Tanh activation function, in sequence.

[0036] Both the first and second mixing modules consist of an upsampling module, a convolutional module, and a LeakyReLU activation function.

[0037] (4) The output of the first mixing module of the C network and the output of the first combination module of the B network are fused 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 are fused through a splicing operation, thereby realizing the fusion of high-frequency features and low-frequency features;

[0038] (5) The output of network C is element-wise multiplied with the brightness attention map image obtained from network A to obtain the image;

[0039] (6) Add the image obtained in step (5) to the low-light image input to the generator network to obtain the enhanced transmission line image.

[0040] Furthermore, both the first and second combination modules of the B network are composed of a first branch and a second branch, wherein:

[0041] The first branch consists of parallel dilated convolution modules;

[0042] The second branch consists of a residual module based on a hybrid attention mechanism and a downsampling.

[0043] Furthermore, 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.

[0044] Furthermore, the construction of the adversarial network in step 3) is represented as follows:

[0045] Adversarial networks consist of a global discriminant network and a local discriminant network, wherein:

[0046] The global discriminant network consists of two combined modules, one residual dilated convolution module, and two combined modules. Its input images are fake and real images.

[0047] The local discriminant network consists of 6 combined modules, and its input image is a randomly cropped image of a fake image and a real image.

[0048] Furthermore, the combined module and residual dilated convolution module of the global discriminant network are as follows:

[0049] The combined module of the global discriminant network consists of convolution + LeakyReLU activation function in sequence;

[0050] The main branch of the residual dilated convolution module of the global discriminant network consists of three dilated convolutions with dilation rates of 2, 3 and 5, respectively.

[0051] Furthermore, the combined module of the local discriminant network is a network module based on convolution + LeakyReLU activation function.

[0052] Furthermore, the loss function for constructing the generative adversarial network in step 4) is expressed as:

[0053]

[0054] In the formula, and These are the loss functions of the global discriminant network and the local discriminant network in an adversarial network, respectively. Per and L Pix These are the perceptual loss function and pixel loss function of the generative network, respectively, and α, β, γ and ω are the weights of the corresponding loss functions.

[0055] Furthermore, the loss functions of the global discriminant network, the local discriminant network, the perceptual loss function of the generative network, and the pixel loss function of the adversarial network are as follows:

[0056] The loss function of the global discriminant network is expressed as:

[0057]

[0058] In the formula, D G G is the global discriminator network, G is the generator network, and z and x are the input images of the generator network and the adversarial network, respectively.

[0059] The loss function of the local discriminant network is expressed as:

[0060]

[0061] In the formula, D L For local discrimination networks;

[0062] The perceptual loss function of the generative network is expressed as:

[0063]

[0064] In the formula, x is the input image of the generator network, and W and H are the width and height of the image, respectively. It is a VGG-19 pre-trained network;

[0065] The pixel loss function of the generator network is expressed as:

[0066]

[0067] Furthermore, step 5) of model training is as follows:

[0068] (1) Input the low-light transmission line images from the training set into the generator network to obtain the enhanced images;

[0069] (2) Input the enhanced image and the normal illumination image in the training set into the adversarial network and determine whether the input image is the enhanced image or the original normal illumination image;

[0070] (3) By using the gradient descent method to optimize the loss function, the parameters of the generator network and the loss network are continuously updated, and finally the training of the generator network and the adversarial network is completed, resulting in the optimal generator network and the adversarial network.

[0071] Furthermore, the network model performance evaluation in step 6) is as follows:

[0072] (1) Input the low-light images in the paired sample test set constructed in step 1 into the trained generative network obtained in step 5 to obtain the enhanced transmission line images, so as to measure the network model's ability to enhance transmission line images under low light conditions.

[0073] (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 to measure the network model’s ability to enhance transmission line images under low illumination.

[0074] (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 values ​​reach or exceed the ideal values, maintain the network model weights for enhancing transmission line images under low illumination.

[0075] This invention provides a method for enhancing transmission line images under low illumination based on generative adversarial networks. It can effectively improve the brightness of transmission line images under low illumination while avoiding overexposure or underexposure and artifacts in the enhanced image, preserving more image detail information and improving the quality of the enhanced transmission line image. Attached Figure Description

[0076] Figure 1 This is a flowchart of a low-light image enhancement method for transmission lines based on generative adversarial networks according to the present invention.

[0077] Figure 2 This is a structural diagram of the generative network and parallel dilated convolution module of the present invention;

[0078] Figure 3 This is a structural diagram of the residual module based on the hybrid attention mechanism of this invention;

[0079] Figure 4 This is the adversarial network structure diagram of the present invention. Detailed Implementation

[0080] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0081] See Figures 1-4 Example 1: This example provides a method for enhancing images of transmission lines under low illumination based on generative adversarial networks. The specific steps are as follows:

[0082] 1) Constructing the dataset

[0083] By selecting low-light and normal-light images with different backgrounds from the power 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:

[0084] (1) Select low-light and normal-light images with different backgrounds from the power 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.

[0085] (2) Divide the images of the normal group into normal group 1 and normal group 2. The images of normal group 1 are 70% of the images of the normal group, and the images of normal group 2 are 30% of the images of the normal group. Construct an unpaired sample training set using the original low-light group images and normal group 1 images.

[0086] (3) The images of the normal two groups are processed to obtain the corresponding low-light images, which are named the processed low-light group;

[0087] (4) Construct a paired sample test set using images from the normal two groups and images from the low-light processed group, thereby constructing an unpaired sample training set and a paired sample test set.

[0088] Low-light and normal-light images with different backgrounds were selected from the power transmission line video monitoring system, and their pixel size was adjusted to 600×400.

[0089] 2) Constructing a generative network

[0090] Construct a network for enhancing images of power transmission lines under low illumination;

[0091] 3) Constructing adversarial networks

[0092] Construct a network to determine whether an input image of a power transmission line is a real image or a fake image;

[0093] The image of the transmission line under normal illumination is defined as the true image;

[0094] The output image of the generator network is defined as a fake image;

[0095] 4) Construct the loss function for generative adversarial networks.

[0096] Used to measure the performance of generative networks and adversarial networks during the network training process;

[0097] 5) Network model training

[0098] The model training process is as follows:

[0099] (2) Input the low-light transmission line images from the training set into the generator network to obtain the enhanced images;

[0100] (2) Input the enhanced image and the normal illumination image in the training set into the adversarial network and determine whether the input image is the enhanced image or the original normal illumination image;

[0101] (3) By using the gradient descent method to optimize the loss function, the parameters of the generator network and the loss network are continuously updated, and finally the training of the generator network and the adversarial network is completed, resulting in the optimal generator network and the adversarial network.

[0102] 6) Network model performance evaluation

[0103] (1) Input the low-light images in the paired sample test set constructed in step 1 into the trained generative network obtained in step 5 to obtain the enhanced transmission line images, so as to measure the network model's ability to enhance transmission line images under low light conditions.

[0104] (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 to measure the network model’s ability to enhance transmission line images under low illumination.

[0105] (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 values ​​reach or exceed the ideal values, maintain the network model weights for enhancing transmission line images under low illumination.

[0106] 7) Application of network models

[0107] The trained network model is deployed on a server to enhance low-light transmission line images transmitted from the field, resulting in enhanced images.

[0108] In step 2), the generator network is constructed, which consists of network A, network B, and network C, as follows:

[0109] (1) Network A is used to preprocess the input low-light transmission line image. It consists of a brightness attention map to reduce the overexposure or underexposure of the generated image. The brightness attention map obtained after preprocessing is added to the input low-light transmission line image and used as the input image of Network B.

[0110] (2) The B network is used to extract low-frequency information from the input low-light image of the transmission line. It consists of a convolutional module, a LeakyReLU activation function, a first combination module, a second combination module, and a residual module based on a hybrid attention mechanism, wherein:

[0111] ① Both the first and second combination modules consist of a first branch and a second branch. The output feature maps of the first and second branches of each combination module are spliced ​​together to obtain the output feature map of that combination module.

[0112] ②The residual module based on the hybrid attention mechanism consists of a convolutional module, a LeakyReLU activation function, a convolutional module, a parallel attention module, a convolutional module, a LeakyReLU activation function, a convolutional module, another parallel attention module, and a convolutional module in sequence.

[0113] ③ Add the input feature map of the residual module based on the hybrid attention mechanism described in step ② to the output feature map of the last convolutional module in the residual module based on the hybrid attention mechanism element by element to obtain the final output feature map of the residual module based on the hybrid attention mechanism.

[0114] (3) The C network is mainly used to extract high-frequency information from the input image. It consists of a first hybrid module, a residual module based on a hybrid attention mechanism, a second hybrid module, another residual module based on a hybrid attention mechanism, a convolutional module, and a Tanh activation function, in sequence.

[0115] Both the first and second mixing modules consist of an upsampling module, a convolutional module, and a LeakyReLU activation function.

[0116] (4) The output of the first mixing module of the C network and the output of the first combination module of the B network are fused 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 are fused through a splicing operation, thereby realizing the fusion of high-frequency features and low-frequency features;

[0117] (5) The output of network C is element-wise multiplied with the brightness attention map image obtained from network A to obtain the image;

[0118] (6) Add the image obtained in step (5) to the low-light image input to the generator network to obtain the enhanced transmission line image.

[0119] The first and second combination modules of the B network are both composed of a first branch and a second branch, wherein:

[0120] The first branch consists of parallel dilated convolution modules;

[0121] The second branch consists of a residual module based on a hybrid attention mechanism and a downsampling.

[0122] 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.

[0123] The construction of the adversarial network in step 3) is represented as follows:

[0124] Adversarial networks consist of a global discriminant network and a local discriminant network, wherein:

[0125] The global discriminant network consists of two combined modules, one residual dilated convolution module, and two combined modules. Its input images are fake and real images.

[0126] The local discriminant network consists of 6 combined modules, and its input is a randomly cropped image of a fake image and a real image.

[0127] The combined module and residual dilated convolution module of the global discriminant network are:

[0128] The combined module of the global discriminant network consists of convolution + LeakyReLU activation function in sequence;

[0129] The main branch of the residual dilated convolution module of the global discriminant network consists of three dilated convolutions with dilation rates of 2, 3 and 5, respectively.

[0130] The combined module of the local discriminant network is a network module based on convolution + LeakyReLU activation function.

[0131] The loss function for constructing the generative adversarial network in step 4) is expressed as follows:

[0132]

[0133] In the formula, and These are the loss functions of the global discriminant network and the local discriminant network in an adversarial network, respectively. Per and L Pix These are the perceptual loss function and pixel loss function of the generative network, respectively, and α, β, γ and ω are the weights of the corresponding loss functions.

[0134] The loss functions of the global discriminant network, the local discriminant network, the perceptual loss function of the generative network, and the pixel loss function of the adversarial network are as follows:

[0135] The loss function of the global discriminant network is expressed as:

[0136]

[0137] In the formula, D G G is the global discriminator network, G is the generator network, and z and x are the input images of the generator network and the adversarial network, respectively.

[0138] The loss function of the local discriminant network is expressed as:

[0139]

[0140] In the formula, D L For local discrimination networks;

[0141] The perceptual loss function of the generative network is expressed as:

[0142]

[0143] In the formula, x is the input image of the generator network, and W and H are the width and height of the image, respectively. It is a VGG-19 pre-trained network;

[0144] The pixel loss function of the generator network is expressed as:

[0145]

[0146] All convolutional modules described in this embodiment have the same structure.

[0147] The residual modules based on the hybrid attention mechanism described in this embodiment all have the same structure.

[0148] The parallel attention modules described in this embodiment all have the same structure.

[0149] The LeakyRelu activation functions described in this embodiment all have the same structure.

[0150] The combined modules described in this embodiment have the same structure.

[0151] This embodiment is implemented using existing technology.

Claims

1. A method for enhancing images of transmission lines under low illumination based on generative adversarial networks, characterized in that, The specific steps are as follows: 1) Constructing the dataset By selecting low-light and normal-light images with different backgrounds from the power transmission line video monitoring system and processing them, an unpaired sample training set and a paired sample test set were constructed. 2) Constructing a generative network A network is constructed to enhance images of transmission lines under low illumination. The generator network consists of networks A, B, and C, as detailed below: (1) Network A is used to preprocess the input low-light transmission line image. It consists of a brightness attention map to reduce the overexposure or underexposure of the generated image. The brightness attention map obtained after preprocessing is added to the input low-light transmission line image and used as the input image of Network B. (2) The B network is used to extract low-frequency information from the input low-light image of the transmission line. It consists of a convolutional module, a LeakyReLU activation function, a first combination module, a second combination module, and a residual module based on a hybrid attention mechanism, wherein: ① Both the first and second combination modules consist of a first branch and a second branch. The output feature maps of the first and second branches of each combination module are spliced ​​together to obtain the output feature map of that combination module. ②The residual module based on the hybrid attention mechanism consists of a convolutional module, a LeakyReLU activation function, a convolutional module, a parallel attention module, a convolutional module, a LeakyReLU activation function, a convolutional module, another parallel attention module, and a convolutional module in sequence. ③ Add the input feature map of the residual module based on the hybrid attention mechanism described in step ② to the output feature map of the last convolutional module in the residual module based on the hybrid attention mechanism element by element to obtain the final output feature map of the residual module based on the hybrid attention mechanism. (3) The C network is mainly used to extract high-frequency information from the input image. It consists of a first hybrid module, a residual module based on a hybrid attention mechanism, a second hybrid module, another residual module based on a hybrid attention mechanism, a convolutional module, and a Tanh activation function, in sequence. Both the first and second mixing modules consist of an upsampling module, a convolutional 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 are fused 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 are fused through a splicing operation, thereby realizing the fusion of high-frequency features and low-frequency features; (5) The output of network C is element-wise multiplied with the brightness attention map image obtained from network A to obtain the image; (6) Add the image obtained in step (5) to the low-light image input to the generator network to obtain the enhanced transmission line image; 3) Constructing adversarial networks Construct a network to determine whether an input image of a power transmission line is a real image or a fake image; The image of the transmission line under normal illumination is defined as the true image; The output image of the generator network is defined as a fake image; 4) Construct the loss function for generative adversarial networks Used to measure the performance of generative networks and adversarial networks during the network training process; 5) Network model training The optimal generative network and adversarial network are obtained through network model training; 6) Network model performance evaluation The low-light images from the paired sample test set constructed in step 1 are input into the trained generative network obtained in step 5 to obtain the enhanced transmission line images, so as to measure the network model's ability to enhance transmission line images under low light conditions. 7) Application of network models The trained network model is deployed on a server to enhance low-light transmission line images transmitted from the field, resulting in enhanced images.

2. The method for enhancing images of transmission lines under low illumination based on generative adversarial networks as described in claim 1, characterized in that, The first and second combination modules of the B network are both composed of a first branch and a second branch, wherein: The first branch consists of parallel dilated convolution modules; The second branch consists of a residual module based on a hybrid attention mechanism and a downsampling.

3. The method for enhancing images of transmission lines under low illumination based on generative adversarial networks as described in claim 1, 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.

4. The method for enhancing images of transmission lines under low illumination based on generative adversarial networks as described in claim 1, characterized in that, The construction of the adversarial network in step 3) is represented as follows: Adversarial networks consist of a global discriminant network and a local discriminant network, wherein: The global discriminant network consists of two combined modules, one residual dilated convolution module, and two combined modules. Its input images are fake and real images. The local discriminant network consists of 6 combined modules, and its input is a randomly cropped image of a fake image and a real image.

5. The method for enhancing images of transmission lines under low illumination based on generative adversarial networks as described in claim 4, characterized in that, The combined module and residual dilated convolution module of the global discriminant network are: Both the global discriminant network's combined module and the local discriminant network's combined module are composed of convolution + LeakyReLU activation function; The main branch of the residual dilated convolution module consists of three dilated convolutions with dilation rates of 2, 3 and 5, respectively.

6. The method for enhancing images of transmission lines under low illumination based on generative adversarial networks as described in claim 4, characterized in that, The combined module of the local discriminant network is a network module based on convolution + LeakyReLU activation function.

7. The method for enhancing images of transmission lines under low illumination based on generative adversarial networks as described in claim 1, characterized in that, The loss function for constructing the generative adversarial network in step 4) is expressed as: In the formula, and These are the loss functions of the global discriminant network and the local discriminant network in an adversarial network, respectively. Per and L Pix These are the perceptual loss function and pixel loss function of the generative network, respectively, and α, β, γ and ω are the weights of the corresponding loss functions mentioned above.

8. The method for enhancing images of transmission lines under low illumination based on generative adversarial networks as described in claim 7, characterized in that, The loss functions of the global discriminant network, the local discriminant network, the perceptual loss function of the generative network, and the pixel loss function of the adversarial network are as follows: The loss function of the global discriminant network is expressed as: In the formula, D G G is the global discriminator network, G is the generator network, and z and x are the input images of the generator network and the adversarial network, respectively. The loss function of the local discriminant network is expressed as: In the formula, D L For local discrimination networks; The perceptual loss function of the generative network is expressed as: In the formula, x is the input image of the generator network, and W and H are the width and height of the image, respectively. It is a VGG-19 pre-trained network; The pixel loss function of the generator network is expressed as:

Images