An image enhancement method based on global and channel attention multi-scale aggregation network

Abstract

The application discloses an image enhancement method based on a global and channel attention multi-scale aggregation network, and comprises the following steps: acquiring an image data set to be processed; constructing a low-light image enhancement model; inputting the image data set to be processed into the low-light image enhancement model for processing to obtain an enhanced image; wherein the low-light image enhancement model is obtained through training of a training set, and the training set is a low-light image data set and a corresponding reference image data set. The application improves the utilization efficiency of low-light image information, thereby better restoring image details, adjusting image saturation, reducing noise and improving visual effects.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an image enhancement method based on a global and channel attention multi-scale aggregation network. Background Technology

[0002] Images captured in low-light environments typically suffer from insufficient brightness, poor contrast, loss of detail, and increased noise. These problems severely impact the visual quality of the images and the effectiveness of subsequent image processing and analysis tasks. Traditional low-light enhancement methods, such as histogram equalization and gamma correction, can improve image brightness to some extent, but often lead to over-enhancement, resulting in side effects such as oversaturation in bright areas, loss of detail in dark areas, and noise diffusion.

[0003] With the advancement of deep learning technology, low-light image enhancement methods based on convolutional neural networks (CNNs) have gradually become mainstream. These methods typically employ an end-to-end learning approach, enabling them to learn complex mapping functions and achieve brightness adjustment and detail enhancement. However, existing deep learning methods still have shortcomings when processing low-light images: on the one hand, local convolutional structures struggle to capture overall image brightness information, resulting in uneven brightness in the enhanced image; on the other hand, single-scale feature extraction methods fail to retain rich detail information and lack the ability to fuse features at different scales. Furthermore, existing attention mechanisms often focus on local features or single channels, failing to fully utilize global statistical information. These problems often lead to enhancement results with blurred details, poor contrast, and residual noise. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention proposes an image enhancement method based on a global and channel attention multi-scale aggregation network. This method integrates three core components: a global brightness-aware pixel calibration module, an adaptive global-channel attention unit, and a multi-scale encoder feature aggregation module, thereby effectively improving the performance of low-light image enhancement models.

[0005] To achieve the above objectives, this invention provides an image enhancement method based on a global and channel attention multi-scale aggregation network, comprising:

[0006] Obtain the image dataset to be processed;

[0007] Construct a low-light image enhancement model;

[0008] The image dataset to be processed is input into the low-light image enhancement model for processing to obtain the enhanced image;

[0009] The low-light image enhancement model is obtained by training a training set, which is a low-light image dataset and its corresponding reference image dataset. The low-light image enhancement model is based on global brightness-aware pixel calibration, adaptive global-channel attention mechanism and multi-scale encoder for feature aggregation.

[0010] Preferably, the low-light image enhancement model includes:

[0011] Global brightness-aware pixel calibration module: used to perform initial brightness enhancement on input low-light images;

[0012] Encoder module: Used to obtain the features of the encoded image;

[0013] Multi-scale encoder feature aggregation module: Used to obtain aggregated features of multi-scale encoder features of encoded image features;

[0014] Downsampling module: used to obtain downsampled image features with a spatial size of half that of the encoded image features and a channel dimension of twice that of the encoded image features;

[0015] Upsampling module: used to obtain upsampled image features with a spatial size twice that of the downsampled image features and a channel dimension half that of the downsampled image features;

[0016] Decoder module: Used to obtain the features of the decoded image.

[0017] Preferably, the processing procedure of the global brightness sensing pixel calibration module includes:

[0018] Calculate the mean information of the low-light image to be processed in the channel dimension;

[0019] The mean information is stitched together with the low-light image to be processed in the channel dimension;

[0020] Pixel-level initial adjustment weights are obtained through adaptive learning via convolutional layers;

[0021] The pixel-level initial adjustment weights are multiplied by the original image and connected through residuals to obtain a preliminary enhanced image.

[0022] Preferably, the encoder module and the decoder module have the same structure, including a first residual unit for global feature enhancement and a second residual unit for inter-channel information exchange.

[0023] Preferably, the first residual unit is used to normalize and map the input features, and retain the efficient information flow through a gating mechanism. Then, it is used for global feature enhancement through an adaptive global-channel attention unit to obtain the globally enhanced image features.

[0024] The second residual unit is used to expand the channel dimension of the globally enhanced image features, selectively retain important features through a gating mechanism, and finally restore the image features to the original channel dimension, resulting in image features with full inter-channel information exchange.

[0025] Preferably, the adaptive global-channel attention unit includes parallel global attention branches and channel attention branches;

[0026] The global attention branch is used to extract global spatial attention weights for the input features;

[0027] The channel attention branch is used to generate channel attention weights for the input features;

[0028] The global spatial attention weights and the channel attention weights are adaptively fused to obtain a comprehensive attention weight; the comprehensive attention weight is used to enhance the input features to obtain globally enhanced image features.

[0029] Preferably, the multi-scale encoder feature aggregation module includes:

[0030] The feature alignment unit is used to perform spatial resolution alignment on multi-scale features to obtain multi-scale features with uniform size.

[0031] A multi-branch attention weighting unit is used to perform channel attention weighting on features at various scales;

[0032] The feature fusion unit is used to splice and fuse weighted multi-scale features to obtain multi-scale aggregated features.

[0033] The modulation output unit is used to perform channel modulation on the multi-scale aggregated features and add them to the decoder input features to output the enhanced decoder input features.

[0034] Preferably, the low-light image enhancement model is an end-to-end network model structure, wherein the output of the global brightness-aware pixel calibration module is used as the input of the encoder module;

[0035] The multi-scale features extracted by the encoder are fused by the multi-scale encoder feature aggregation module and then input together with the features output by the upsampling module into the decoder module for reconstruction.

[0036] The transformation between the feature map space and the channel dimension is achieved through the downsampling module and the upsampling module.

[0037] Preferably, the training loss function of the low-light image enhancement model includes a reconstruction loss function, a perceptual loss function, a structural consistency loss function, and an edge loss function.

[0038] Compared with the prior art, the present invention has the following advantages and technical effects:

[0039] This invention deeply mines information from low-light images through global and channel attention and multi-scale encoder feature aggregation, achieving excellent performance in processing low-light image datasets. Compared with existing technologies, it can better restore image details, adjust image saturation, reduce noise, and improve visual effects. Attached Figure Description

[0040] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0041] Figure 1 This is a flowchart of an image enhancement method based on a global and channel attention multi-scale aggregation network according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the global brightness sensing pixel calibration module according to an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the encoder and decoder modules according to an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the multi-scale encoder feature aggregation module structure according to an embodiment of the present invention. Detailed Implementation

[0045] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0046] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0047] This embodiment proposes an image enhancement method based on a global and channel attention multi-scale aggregation network, such as... Figure 1 ,include:

[0048] Obtain the image dataset to be processed;

[0049] Construct a low-light image enhancement model;

[0050] The image dataset to be processed is input into the low-light image enhancement model for processing to obtain the enhanced image;

[0051] The low-light image enhancement model is obtained by training a training set, which is a low-light image dataset and its corresponding reference image dataset. The low-light image enhancement model is based on global brightness-aware pixel calibration, adaptive global-channel attention mechanism and multi-scale encoder for feature aggregation.

[0052] Furthermore, the low-light image enhancement model includes:

[0053] Global brightness-aware pixel calibration module: used to perform initial brightness enhancement on input low-light images;

[0054] Encoder module: Used to obtain the features of the encoded image;

[0055] Multi-scale encoder feature aggregation module: Used to obtain aggregated features of multi-scale encoder features of encoded image features;

[0056] Downsampling module: used to obtain downsampled image features with a spatial size of half that of the encoded image features and a channel dimension of twice that of the encoded image features;

[0057] Upsampling module: used to obtain upsampled image features with a spatial size twice that of the downsampled image features and a channel dimension half that of the downsampled image features;

[0058] Decoder module: Used to obtain the features of the decoded image.

[0059] Specifically, based on low-light images, a global brightness-aware pixel calibration module is constructed to obtain images with preliminary brightness enhancement;

[0060] Based on the image after initial brightness enhancement, an encoder module is constructed to obtain the encoded image features;

[0061] Based on the encoded image features, a multi-scale encoder feature aggregation module is constructed to obtain the aggregated features of the multi-scale encoder features of the encoded image features;

[0062] Based on the encoded image features, a downsampling module is constructed to obtain downsampled image features with a spatial size of half that of the encoded image features and a channel dimension of twice that of the encoded image features;

[0063] Based on the downsampled image features, an upsampling module is constructed to obtain upsampled image features with a spatial size twice that of the downsampled image features and a channel dimension half that of the downsampled image features;

[0064] A decoder module is constructed based on the aggregated features of the multi-scale encoder features and the upsampled image features to obtain the decoded image features;

[0065] A low-light image enhancement model is constructed based on a global brightness-aware pixel calibration module, a downsampling module, an encoder module, an upsampling module, a decoder module, and a multi-scale encoder feature aggregation module.

[0066] By combining the low-light image dataset and its corresponding reference image dataset, the low-light image enhancement model is trained to obtain the trained low-light image enhancement model.

[0067] Acquire a low-light image, input the low-light image into a trained low-light image enhancement model, and obtain the enhanced image.

[0068] Furthermore, such as Figure 2 The processing steps of the global brightness-sensing pixel calibration module include:

[0069] Calculate the mean information of the low-light image to be processed in the channel dimension;

[0070] The mean information is stitched together with the low-light image to be processed in the channel dimension;

[0071] Pixel-level initial adjustment weights are obtained through adaptive learning via convolutional layers;

[0072] The pixel-level initial adjustment weights are multiplied by the original image and connected through residuals to obtain a preliminary enhanced image.

[0073] Low-light image I low The prior information features of brightness are I lum For a single-channel image, the average is obtained by adding the pixel values ​​of the R, G, and B channels pixel by pixel according to their positions; I lum The image is concatenated with the original image along the channel dimension, and the number of channels is expanded by a 1×1 convolutional layer. Then, it is passed through a 5×5 depthwise separable convolutional layer to extract illumination features. Finally, another 1×1 convolutional layer is used to generate an illumination enhancement map M. The input low-illumination image is multiplied pixel by pixel with M and added to the original image to obtain the calibrated enhanced image I. calibrated , that is I calibrated =I low +I low *M, thereby enabling perceptual calibration and enhancement of global brightness.

[0074] Furthermore, the encoder module and the decoder module have the same structure, including a first residual unit for global feature enhancement and a second residual unit for inter-channel information interaction.

[0075] The first residual unit is used to normalize and map the input features, and retains the efficient information flow through a gating mechanism. Then, it is used to perform global feature enhancement through an adaptive global-channel attention unit to obtain the globally enhanced image features.

[0076] The second residual unit is used to expand the channel dimension of the globally enhanced image features. It selectively retains important features through a gating mechanism and finally restores the image features to the original channel dimension, resulting in image features with full inter-channel information exchange.

[0077] Specifically, such as Figure 3 The encoder and decoder modules have the same structure, consisting of two residual blocks used to enhance local and global features. Input feature X enters the first residual block (i.e., the first residual unit), then passes through a layer of normalization, a 1×1 convolution, and a 3×3 depthwise separable convolution to obtain the feature map. The SimpleGate mechanism ensures efficient information flow, followed by an adaptive global-channel attention unit to obtain a larger receptive field, enabling global feature enhancement and inter-channel information interaction. Finally, a 1×1 convolution restores the feature dimension.

[0078] The second residual block (i.e., the second residual unit) emphasizes the correlation between channels and feature enhancement. The input features are first normalized by a layer, followed by a 1×1 convolution that doubles the channel dimension for channel expansion. The SimpleGate mechanism selectively preserves important features, and another 1×1 convolution transforms the features back to the original channel size.

[0079] Furthermore, such as Figure 3 The global-channel attention unit includes parallel global attention branches and channel attention branches;

[0080] The global attention branch is used to extract global spatial attention weights for the input features;

[0081] The channel attention branch is used to generate channel attention weights for the input features;

[0082] The global spatial attention weights and the channel attention weights are adaptively fused to obtain a comprehensive attention weight; the comprehensive attention weight is used to enhance the input features to obtain globally enhanced image features.

[0083] Specifically, the structure of the adaptive global-channel attention unit is as follows: the unit contains two parallel branches.

[0084] The first branch is the global attention branch: the input feature X is first preprocessed by a 3×3 depthwise separable convolution and a 1×1 pointwise convolution; then it is processed through two paths. The first path sequentially performs global max pooling, linear mapping, and the Sigmoid activation function on the feature to obtain the attention score. max .

[0085] The second path sequentially performs global average pooling, linear mapping, and Sigmoid activation on the features to obtain the attention score. avg Score max With Score avg The sum is the global attention weight W. global .

[0086] The second branch is the channel attention branch: global average pooling is performed on the input feature X, and channel attention weights W are generated through 1×1 convolution. channel .

[0087] Finally, the outputs of the two branches are weighted and fused using a learnable parameter α, and then multiplied with the input feature X to obtain the output X. ′ .

[0088] The calculation formula is as follows:

[0089] Score avg =σ(Fc1(GAP(X)));

[0090] Score max =σ(Fc2(GMP(X)));

[0091] W global =Score avg +Score max ;

[0092] X ′ =X·(α·W) global +(1-α)·W channel +1);

[0093] In the formula, GAP and GMP represent global average pooling and global max pooling, respectively, Fc is a linear layer, and σ is the Sigmoid function.

[0094] After each layer of the encoder, a downsampling module is included. This module uses a convolution operation with a kernel size of 2×2 and a stride of 2. After downsampling, the spatial size of the features is reduced to half of its original size, while the channel dimension is doubled. For example, the feature dimension after the first encoder layer and downsampling process changes from the initial C×H×W to...

[0095] Before each decoder module, a lightweight upsampling module is provided, consisting of 1×1 convolutions and a PixelShuffle operation. First, the number of channels in the input features is increased from C to C×2 using 1×1 convolutions. Then, the PixelShuffle operation reassembles the channel data, converting it into spatial information, thus achieving a 2x spatial upsampling while halving the number of channels. Finally, the spatial size of the output features is doubled, and the number of channels is restored to C, meaning the feature dimension changes from C×H×W to C×2H×2W.

[0096] Furthermore, the multi-scale encoder feature aggregation module includes:

[0097] The feature alignment unit is used to perform spatial resolution alignment on multi-scale features to obtain multi-scale features with uniform size.

[0098] A multi-branch attention weighting unit is used to perform channel attention weighting on features at various scales;

[0099] The feature fusion unit is used to splice and fuse weighted multi-scale features to obtain multi-scale aggregated features.

[0100] The modulation output unit is used to perform channel modulation on the multi-scale aggregated features and add them to the decoder input features to output the enhanced decoder input features.

[0101] Specifically, a multi-scale encoder feature aggregation module is constructed, as shown in the schematic diagram below. Figure 4 As shown, taking the first decoder module as an example: this module is used to fuse features from different scales of the encoder to achieve multi-scale information interaction and enhancement. Specifically, firstly, the features E1, E2, E3, and E4 from each level of the encoder are uniformly sampled to the same spatial resolution as E4 using bilinear interpolation. For each scale of the sampled features, they are processed by the same lightweight channel attention module. Within this module, the sampled features are reduced to half their original channel dimension using a 1×1 convolution, activated by the LeakyReLU activation function, and then processed by a 3×3 depthwise separable convolution. Following this, global average pooling is performed along the channel dimension, and finally, a 1×1 convolution is used to convert the features to the input channel dimension, generating attention weights. These weights are then adaptively weighted by multiplying the branch features after Sigmoid activation. After each branch receives its weighted features, they are concatenated along the channel dimension, and then fused using a 1×1 convolution layer. This is followed by LeakyReLU activation and a 3×3 depthwise separable convolution layer for further refinement. Finally, channel attention modulation (e.g., 1×1 convolution after global average pooling) is applied to the fused result to obtain the final multi-scale feature output F. MS ;FMS The input D1 of the first decoder is added pixel by pixel to obtain the final input D1 of the first encoder. ′ Through the multi-scale encoder feature aggregation module, details and structural information in low-light scenes can be effectively integrated across different scales, improving the expressive power of features.

[0102] Furthermore, the low-light image enhancement model is an end-to-end network model structure, wherein the output of the global brightness-aware pixel calibration module is used as the input of the encoder module;

[0103] The multi-scale features extracted by the encoder are fused by the multi-scale encoder feature aggregation module and then input together with the features output by the upsampling module into the decoder module for reconstruction.

[0104] The transformation between the feature map space and the channel dimension is achieved through the downsampling module and the upsampling module.

[0105] Specifically, an end-to-end image enhancement network is constructed by cascading a global brightness-aware pixel calibration module, a downsampling module, an encoder module, an upsampling module, a decoder module, and a multi-scale encoder feature aggregation module. The output of the global brightness-aware pixel calibration module is used as the input of the encoder module. The multi-scale features extracted by the encoder module are input to the multi-scale encoder feature aggregation module for fusion. The fused features are then reconstructed in the decoder module along with the features output by the upsampling module. The feature map spatial resolution and channel dimension are transformed through the downsampling and upsampling modules. Finally, the enhanced image is output through the decoder module, thus completing the construction of the low-light image enhancement model.

[0106] Furthermore, the training loss function of the low-light image enhancement model includes a reconstruction loss function, a perceptual loss function, a structural consistency loss function, and an edge loss function.

[0107] Specifically, reconstruction loss L rec The calculation formula is:

[0108]

[0109] In the formula, This represents the value of the enhanced image at the i-th pixel. This represents the value of the reference image at the i-th pixel corresponding to the enhanced image, where N is the total number of pixels being compared, and i is the pixel index, with a value range of i = 1, 2, ..., N.

[0110] Perceptual loss function L per The formula for calculation is:

[0111]

[0112] In the formula, C j H j and W j Let L and L represent the perceptual loss function respectively. per The number of channels, height, and width of the output feature map of the j-th convolutional layer in a convolutional neural network. L represents the perceptual loss function per The features of the real image extracted by the j-th convolutional layer of the convolutional neural network. L represents the perceptual loss function per The features of the enhanced image extracted by the j-th convolutional layer of the convolutional neural network.

[0113] Structural consistency loss L ssim The calculation formula is:

[0114] L ssim =1-SSIM(I pre ,I gt );

[0115] In the formula, SSIM(I pre ,I gt ) Function to calculate image I pre and Image I gt The similarity between them ranges from [0,1].

[0116] Edge loss L edge The calculation method is as follows:

[0117]

[0118] In the formula, L mse The mean squared error loss F(*) represents the edge extraction operation based on the Laplacian pyramid, and I pre For the enhanced image output by the model, I gt To enhance the reference image corresponding to the image, To enhance the value of the image at the i-th pixel position, To enhance the value of the reference image corresponding to the image at the i-th pixel position, N is the total number of pixels being compared, and i is the pixel index, with a value range of i = 1, 2, ..., N.

[0119] Low-light image I low Input the image enhancement model and train iteratively. When the number of training iterations reaches the preset value of 1500, the trained image enhancement model is obtained.

[0120] Will I low Input to the global brightness perception calibration module to obtain the calibrated image I en That is, enhanced image Ien ;

[0121] Enhance image I en The input to the encoder undergoes various convolutional operations and adaptive global-channel attention units, and is downsampled to obtain encoder image features E at different scales. i ,,i represents the number of encoder blocks.

[0122] Encoder features at multiple scales are input into a multi-scale encoder feature aggregation module to obtain feature maps at multiple scales. These feature maps are then used as input to the decoder to obtain the decoded image features F. de .

[0123] Decoded feature F de The image is fused with the original input image via residual connections to obtain the enhanced result I. pre .

[0124] The low-light image is input into the trained image enhancement model, and the enhanced image is obtained through the processing of the aforementioned modules.

[0125] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An image enhancement method based on a global and channel attention multi-scale aggregation network, characterized in that, include: Obtain the image dataset to be processed; Construct a low-light image enhancement model; The image dataset to be processed is input into the low-light image enhancement model for processing to obtain the enhanced image; The low-light image enhancement model is obtained by training a training set, which is a low-light image dataset and its corresponding reference image dataset. The low-light image enhancement model is based on global brightness-aware pixel calibration, adaptive global-channel attention mechanism and multi-scale encoder for feature aggregation. Constructing the low-light image enhancement model includes: Based on low-light images, a global brightness-aware pixel calibration module is constructed to obtain images with preliminary brightness enhancement. Based on the image after initial brightness enhancement, an encoder module is constructed to obtain the encoded image features; Based on the encoded image features, a multi-scale encoder feature aggregation module is constructed to obtain the aggregated features of the multi-scale encoder features of the encoded image features; Based on the encoded image features, a downsampling module is constructed to obtain downsampled image features with a spatial size of half that of the encoded image features and a channel dimension of twice that of the encoded image features; Based on the downsampled image features, an upsampling module is constructed to obtain upsampled image features with a spatial size twice that of the downsampled image features and a channel dimension half that of the downsampled image features; A decoder module is constructed based on the aggregated features of the multi-scale encoder features and the upsampled image features to obtain the decoded image features; The low-light image enhancement model is constructed based on a global brightness-aware pixel calibration module, a downsampling module, an encoder module, an upsampling module, a decoder module, and a multi-scale encoder feature aggregation module. The encoder module and the decoder module have the same structure, including a first residual unit for global feature enhancement and a second residual unit for inter-channel information interaction; The first residual unit is used to normalize and map the input features, and retains the efficient information flow through a gating mechanism. Then, it is used to perform global feature enhancement through an adaptive global-channel attention unit to obtain the globally enhanced image features. The second residual unit is used to expand the channel dimension of the globally enhanced image features, selectively retain important features through a gating mechanism, and finally restore the image features to the original channel dimension, thereby obtaining image features with full communication between channels. The adaptive global-channel attention unit includes parallel global attention branches and channel attention branches; The global attention branch is used to extract global spatial attention weights for the input features; The channel attention branch is used to generate channel attention weights for the input features; The global spatial attention weights and the channel attention weights are adaptively fused to obtain a comprehensive attention weight; the comprehensive attention weight is used to enhance the input features to obtain globally enhanced image features; The multi-scale encoder feature aggregation module includes: The feature alignment unit is used to perform spatial resolution alignment on multi-scale features to obtain multi-scale features with uniform size. A multi-branch attention weighting unit is used to perform channel attention weighting on features at various scales; The feature fusion unit is used to splice and fuse weighted multi-scale features to obtain multi-scale aggregated features. The modulation output unit is used to perform channel modulation on the multi-scale aggregated features and add them to the decoder input features to output the enhanced decoder input features.

2. The image enhancement method based on a global and channel attention multi-scale aggregation network according to claim 1, characterized in that, The processing procedure of the global brightness sensing pixel calibration module includes: Calculate the mean information of the low-light image to be processed in the channel dimension; The mean information is stitched together with the low-light image to be processed in the channel dimension; Pixel-level initial adjustment weights are obtained through adaptive learning via convolutional layers; The pixel-level initial adjustment weights are multiplied by the original image and connected through residuals to obtain a preliminary enhanced image.

3. The image enhancement method based on a global and channel attention multi-scale aggregation network according to claim 1, characterized in that, The low-light image enhancement model is an end-to-end network model structure, wherein the output of the global brightness-aware pixel calibration module is used as the input of the encoder module; The multi-scale features extracted by the encoder are fused by the multi-scale encoder feature aggregation module and then input together with the features output by the upsampling module into the decoder module for reconstruction. The transformation between the feature map space and the channel dimension is achieved through the downsampling module and the upsampling module.

4. The image enhancement method based on a global and channel attention multi-scale aggregation network according to claim 1, characterized in that, The training loss function of the low-light image enhancement model includes a reconstruction loss function, a perceptual loss function, a structural consistency loss function, and an edge loss function.

Images