Low-illumination image enhancement method combining multi-channel parallel attention and cross fusion

Abstract

The invention discloses a low-illumination image enhancement method combining multi-channel parallel attention and cross fusion, and relates to the field of photoelectric display. In order to solve the problems of high model complexity, brightness and color gamut information coupling, artifact fusion and the like in low-illumination image enhancement, the invention provides a lightweight low-illumination image enhancement network, brightness and chrominance information are separated in a YCbCr color space, brightness structure features are extracted through a lightweight denoising module and a multi-head attention mechanism, and a low-illumination image is obtained. And a multi-channel parallel attention module (MCPA) is designed to enhance the chromaticity context modeling capability. According to the invention, a brightness-guided cross fusion module (LCA) is introduced, and cooperative enhancement of the structure and the color is realized through joint channel-space attention optimization. According to the method, a relatively high image quality index is obtained on multi-class public low-illumination data, meanwhile, the parameter quantity and the calculation quantity are remarkably reduced, and the method is suitable for edge device deployment. According to the method, color cast and fusion artifacts can be effectively inhibited, and the consistency of structural details and colors is improved.

Description

Technical Field

[0001] This invention relates to the field of low-light image enhancement technology, specifically to a low-light image enhancement method that combines multi-channel parallel attention and cross-fusion. Background Technology

[0002] In recent years, with the widespread application of image acquisition equipment in complex lighting environments, low-light image enhancement technology has become increasingly important in the field of computer vision. Images acquired in low-light environments often suffer from insufficient brightness, low contrast, noise interference, and loss of detail. These problems significantly reduce the visual quality of the images and affect the performance of subsequent image processing, object detection, and semantic segmentation tasks. Researching efficient low-light image enhancement methods is crucial for improving the robustness and application potential of computer vision systems.

[0003] Existing low-light image enhancement methods are mainly divided into two categories: those based on traditional image processing and those based on deep learning. Traditional methods (such as decomposition methods based on Retinex theory) achieve enhancement by separating illumination and reflection components, but they are poorly adaptable to complex scenes and prone to artifacts or color distortion. Deep learning-based models can significantly improve enhancement results by learning complex illumination-noise relationships, but they generally suffer from high model complexity, large computational costs, and high deployment costs, making them difficult to apply on resource-constrained edge devices. In addition, existing methods often suffer from information coupling and fusion artifacts when processing brightness and color gamut information, resulting in insufficient structural fidelity and color consistency of the enhanced image. Some lightweight models reduce computational costs but sacrifice enhancement results and have poor robustness, making them difficult to meet the needs of practical applications. Therefore, developing a low-light image enhancement method that balances lightweight design and high performance is of great significance. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide a low-light image enhancement method that combines multi-channel parallel attention and cross-fusion, which aims to effectively reduce model complexity, improve the structural fidelity and color consistency of the enhanced image, and is suitable for deployment on edge devices.

[0005] To achieve the above objectives, this invention provides a low-light image enhancement method combining multi-channel parallel attention and cross-fusion, the method comprising the following steps:

[0006] S1: Input the low-light image into the YCbCr color space, separate the luminance (Y) channel and chrominance (Cb, Cr) channels to obtain the corresponding luminance and chrominance features.

[0007] S2: Enhance the features based on the trained lightweight model: extract structural features in the Y channel through a lightweight denoising module and multi-head self-attention (MHSA), extract chromaticity context features in the Cb and Cr channels through a multi-channel parallel attention module (MCPA), and perform feature fusion through a luminance-guided cross-fusion module (LCA) to output an enhanced image;

[0008] S3: Used to obtain the training phase of the trained model: optimize the model with a multidimensional loss function, and jointly constrain it from dimensions such as structural consistency, perceptual quality, pixel distribution and color fidelity;

[0009] S4: Output the enhanced image.

[0010] In one specific embodiment, the color space conversion in step S1 satisfies the following linear relationship:

[0011] Y(i,j)=0.299R(i,j)+0.587G(i,j)+0.114B(i,j);

[0012] Cb(i,j)=-0.168736R(i,j)-0.331264G(i,j)+0.5B(i,j)+0.5;

[0013] Cr(i,j)=0.5R(i,j)-0.418688G(i,j)-0.081312B(i,j)+0.5.

[0014] It is worth mentioning that, in other specific implementations, full-range coefficients that are linearly equivalent to the above relationship are allowed to be used.

[0015] In one specific embodiment, the luminance (Y) channel processing in step S2 includes:

[0016] a) A lightweight denoising module is used to suppress noise and enhance the structure of Y features. The denoising module includes linear transformations of depthwise separable convolutions and / or 1×1 convolutions.

[0017] b) Before multi-head self-attention (MHSA) computation, downsampling or pooling is performed on the feature map to compress the number of tokens, followed by linear mapping of queries, keys, and values ​​and completion of attention computation;

[0018] c) The output of the MHSA adopts residual connection and normalization; the number of attention heads h can take any value from 2 to 8.

[0019] In one specific implementation, the multi-channel parallel attention module (MCPA) in step S2 extracts chroma context and performs weighted fusion through multi-scale pooling branches, wherein:

[0020] a) Feed the input feature X into s scale branches to obtain {F1,...,F...} s The scale set can be one or more of {1, 2, 3, 6};

[0021] b) Obtain non-negative dynamic weights {a1,...,a1} by Softmax normalization along the channel dimension. s The weights are generated by an attention map produced by a 1×1 convolution of X;

[0022] c) The fusion form is: and

[0023] d) Each scale branch is adaptively averaged pooled to the target size and then upsampled back to the original size before participating in the fusion.

[0024] In one specific embodiment, the Luminance Guided Cross-Merging (LCA) module in step S2 includes:

[0025] a) Cascaded integration: F col =Conv(Concat(F S ,F C ), where Conv is a 1×1 or 3×3 convolution, F S For brightness branching features, F C This is a chromaticity branch feature;

[0026] b) Brightness guidance: G S =Conv 1×1 (F S ), and weighted by channel attention (SE), F add =SE(F col +βG S ), where β is the learnable guidance coefficient;

[0027] c) Global dependency modeling and residuals: F out =MHSA(F add )+F add , for subsequent reconstruction.

[0028] In one specific implementation, to obtain the trained model described in step S2, a composite loss is used for training optimization, wherein the composite loss is defined as:

[0029] L total

[0030] =α1L spatial +α2L perceptual +α3L distribution +α4L consistency+α5

[0031] L PSNR

[0032] +α6L color

[0033] Where α1 to α6 are non-negative weights; α i ≥0, i=1,…,6. L spatial For structural consistency loss based on structural similarity, L perceptual For perceptual loss based on pre-trained features, L distribution Pixel distribution / histogram Figure 1 Loss of sexual function, L consistency For spatial or multi-scale consistency loss, L PSNR To compensate for the signal-to-noise ratio improvement loss by using mean square error as a proxy, L color This is a constraint on color deviation in the chroma channel.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] (1) Brightness-color decoupling + brightness-guided fusion: Brightness and chromaticity semantics are explicitly separated in YCbCr space. Combined with LCA, the color response is guided by brightness characteristics, which can effectively suppress color bias and fusion artifacts, and improve structural fidelity and color consistency.

[0036] (2) Lightweight and efficient attention: The luminance branch uses MHSA with lightweight denoising and token compression, and the chrominance branch adopts MCPA multi-scale adaptive fusion; while maintaining enhanced quality, the number of parameters and FLOPs are significantly reduced, making it suitable for real-time deployment on mobile / edge devices.

[0037] (3) Robust generalization and easy deployment: It performs stably in both paired and unpaired scenarios, and has low engineering implementation costs. The above experimental performance and complexity data are used to illustrate the technical effects and deployment feasibility, and do not constitute a limitation on the scope of protection of this invention. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the method implementation of an embodiment of the present invention;

[0039] Figure 2 This is a structural diagram of the model in an embodiment of the present invention;

[0040] Figure 3 This is a structural diagram of the MCPA module in an embodiment of the present invention;

[0041] Figure 4 This is a structural diagram of the LCA module in an embodiment of the present invention;

[0042] Figure 5This is a schematic diagram of YCbCr color space conversion in an embodiment of the present invention;

[0043] Figure 6 This is a visual schematic diagram of the low-light image enhancement effect in an embodiment of the present invention. Detailed Implementation

[0044] The embodiments of this patent are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this patent, and should not be construed as limiting this patent.

[0045] This invention provides a low-light image enhancement method combining multi-channel parallel attention and cross-fusion, comprising the following steps:

[0046] S1: Color space decoupling preprocessing. The input low-light image is converted from RGB space to YCbCr space, separating luminance (Y) and chrominance (Cb, Cr) information to generate corresponding luminance and chrominance feature data. Their linear relationship can be expressed as (where R, G, B, Y, Cb, Cr are the normalized intensities of the three channels / components of pixels at the same location):

[0047] Y(i,j)=0.299R(i,j)+0.587G(i,j)+0.114B(i,j)

[0048] Cb(i,j)=-0.168736R(i,j)-0.331264G(i,j)+0.5B(i,j)+0.5

[0049] Cr(i,j)=0.5R(i,j)-0.418688G(i,j)-0.081312B(i,j)+0.5

[0050] S2: Construct and train a lightweight model and perform enhanced inference. A lightweight network is constructed, and the trained model is used to enhance the input image: structural features are extracted in the luminance (Y) channel using lightweight denoising and multi-head self-attention (MHSA), and chrominance (Cb / Cr) context modeling is enhanced in the chrominance (Cb / Cr) channel using a multi-channel parallel attention module (MCPA); subsequently, joint structure-color modeling and reconstruction are completed in the luminance-guided cross-fusion module (LCA), outputting the enhanced image.

[0051] S3: Multi-dimensional optimization based on composite loss. A composite loss function is constructed to jointly optimize the model from dimensions such as structural consistency, perceptual quality, pixel distribution, and color fidelity, so as to ensure the enhancement of image structural fidelity and color consistency.

[0052] S4: Visualize and optionally evaluate the enhancement results generated by S2 and S3. The enhancement results are visualized, and performance evaluation and comparative analysis can be performed using common metrics such as PSNR, SSIM, NIQE, and BRISQUE.

[0053] In a preferred embodiment, in step S2, the lightweight model features include denoising and global modeling of the luminance (Y) branch. In the luminance channel processing, a lightweight denoising module is introduced to suppress noise and highlight structural edges; this module can consist of a linear transformation of depthwise separable convolutions and 1×1 convolutions, thereby generating structure-sensitive feature maps at extremely low cost, significantly reducing the number of parameters and FLOPs. To improve global modeling capabilities, a multi-head self-attention mechanism (MHSA) is incorporated; downsampling or pooling can be performed before MHSA to compress the number of tokens and reduce the computational complexity of attention. The output yields the luminance branch structural features F. S (Size and number of channels can be set according to the implementation).

[0054] In a preferred embodiment, in step S2, the lightweight model is characterized by a multi-channel parallel attention (MCPA) module for the chroma (Cb / Cr) branch. To enhance the chroma context, a multi-channel parallel attention module (MCPA) is designed. Let the chroma input feature be X, and the output of the i-th scale branch be F. i The dynamic weight is a i With the residual weights a0 (all non-negative, derived from the attention map generated by the 1×1 convolution of X and normalized using Softmax in the channel dimension), the multi-scale adaptive fusion form is:

[0055]

[0056]

[0057] The scale set can be one or more of {1,2,3,6}; each branch adopts adaptive average pooling to the target size, normalization and linear mapping, and then upsampling back to the original space to participate in the fusion, thereby significantly improving the color context modeling capability and alleviating color drift in low-light scenes while maintaining lightweight design.

[0058] In a preferred embodiment, in step S2, the lightweight model is characterized by a luminance-guided cross-fusion module (LCA).

[0059] To achieve synergistic enhancement of structure and color, a luminance-guided cross-fusion module (LCA) is introduced. This module dynamically adjusts the chromaticity response based on luminance characteristics and combines channel and spatial attention for joint optimization. Let F... S For brightness branching features, F C For chroma branch fusion features (X can be taken)' (or its linear mapping). The cascading fusion, luminosity guidance, and global dependency modeling of LCA can be expressed as:

[0060] F col =Conv(Concat(F S ,F C ))

[0061] G S =Conv 1×1 (F S )

[0062] F add =SE(F col +βG S )

[0063] F out =MHSA(F add )+F add

[0064] Where Conv can be a 1×1 or 3×3 convolution, SE(·) represents Squeeze-Excitation channel attention, MHSA(·) represents multi-head self-attention, and β is a learnable guidance coefficient (which can be initialized to 0.2). Finally, F... out The image is fed into the reconstruction head and inversely transformed back to RGB space to obtain the enhanced image.

[0065] In a preferred embodiment, in step S3, a composite loss function is constructed for multi-dimensional optimization. To obtain the trained model described in step S2, a composite loss function is used for end-to-end optimization, integrating factors such as structure, perception, distribution, consistency, PSNR proxy, and color constraints. The composite loss is defined as:

[0066] L total =α1L spatial +α2L perceptual +α3L distribution +α4L consistency +α5L PSNR +α6L color ,α i ≥0

[0067] Among them, L spatial For structural consistency loss based on structural similarity, L perceptual For the perceptual loss in the pre-trained feature space, L distribution pixel / luminance histogram Figure 1 Consistency or statistical distance constraints, L consistency For spatial / multi-scale consistency constraints, L PSNR The signal-to-noise ratio improvement term is substituted by mean square error (MSE), L colorThis is a color deviation constraint in the Cb / Cr space. The weights α1 to α6 can be adjusted according to the validation set performance to balance structural fidelity, perceived quality, and color consistency.

[0068] In a preferred embodiment, in step S4, optional performance evaluation and visualization are performed using general metrics.

[0069] Enhanced images are visualized and compared, and can be evaluated using common metrics such as PSNR, SSIM, NIQE, and BRISQUE (evaluation is not a necessary step in the method). Typical formulas for calculating PSNR and SSIM are as follows:

[0070]

[0071] In the formula, R is the maximum pixel value (R = 255 for an 8-bit image), MSE is the mean square error; μ, σ, σ xy These represent the mean, variance, and covariance, respectively; C1 and C2 are stable terms. NIQE and BRISQUE are no-reference image quality metrics used to quantify naturalness and spatial quality.

[0072] like Figures 1-6 As shown, in the first embodiment of the present invention, a low-light image enhancement method combining multi-channel parallel attention and cross-fusion is provided, including the following steps:

[0073] S1: YCbCr color space decoupling preprocessing. The low-light image is converted from RGB space to YCbCr space, separating the luminance (Y) and chrominance (Cb, Cr) components to generate corresponding luminance and chrominance feature data. A typical linear relationship is:

[0074] Y(i,j)=0.299R(i,j)+0.587G(i,j)+0.114B(i,j)

[0075] Cb(i,j)=-0.168736R(i,j)-0.331264G(i,j)+0.5B(i,j)+0.5

[0076] Cr(i,j)=0.5R(i,j)-0.418688G(i,j)-0.081312B(i,j)+0.5

[0077] Where R, G, B, Y, Cb, and Cr are the channel / component values ​​at the same pixel (i,j), and their values ​​can be normalized to [0,1]. The above coefficients can be adjusted within equivalent linear coefficients such as full-range / TV-range without departing from the spirit of the invention.

[0078] Furthermore, it is worth mentioning that the present invention is not limited to the above coefficients. In practical applications, the above coefficients can be adjusted according to the actual situation.

[0079] In step S1, the training / validation data can be selected from publicly available low-light image datasets, such as paired supervised datasets LOLv1, LOLv2, LSRW, and unpaired real low-light datasets LIME, NPE, MEF, DICM, etc. Since the annotations and quantities of each dataset are different, the annotation format can be converted, randomly cropped, normalized, and divided (e.g., divided into training / validation / test sets in a 7:2:1 ratio) according to project needs. The above selections do not constitute a limitation of the present invention.

[0080] S2: Construct and train a lightweight model, and use the trained model for augmented inference. A lightweight network is constructed, and the trained model is used to augment the input image: Structural features are extracted from the luminance (Y) channel using lightweight denoising and multi-head self-attention (MHSA); chrominance (Cb / Cr) channel chrominance context modeling is enhanced using a multi-channel parallel attention module (MCPA); subsequently, joint structure-color modeling and reconstruction are completed in the luminance-guided cross-fusion module (LCA), outputting the augmented image. Details are as follows.

[0081] Network Implementation and Training Configuration: Specifically, a lightweight model is built using the PyTorch deep learning framework; the number of training epochs can be set to 200 (or adjusted as needed), the input image size can be set to 256×256, and the batch size can be set to 32; the weights are saved every few epochs, and PSNR and SSIM are calculated using the validation set to evaluate the model performance. The weights with the best overall performance are selected as the final model. The above hyperparameters can be adjusted according to computing power and task requirements and are not considered limitations. The constructed network model structure is as follows: Figure 2 As shown.

[0082] Luminance Branch (Y): Lightweight Denoising + MHSA Global Modeling: In luminance channel processing, a lightweight denoising module (such as a linear transformation of depthwise separable convolution DWConv and / or 1×1 convolution) is introduced to suppress noise, highlight edge structures, and significantly reduce the number of parameters and FLOPs. Subsequently, multi-head self-attention (MHSA) is combined, and when necessary, the number of tokens is compressed by pooling / downsampling before constructing query / key / value vectors to improve global context modeling capabilities, outputting luminance structural features F. S .

[0083] Chromaticity branch (Cb / Cr), Multi-channel parallel attention (MCPA): In the chromaticity channel, MCPA modules are designed to enhance the fusion of chromaticity context and multi-scale information. For example... Figure 3 As shown, let the chromaticity input feature be X, and the scale set be X. (This can be adjusted according to implementation). The i-th scale branch is upsampled back to its original size after adaptive average pooling, normalization, and 1×1 convolution to obtain F. i Applying a 1×1 convolution to X yields... For each channel attention map, perform Softmax along the channel dimension to obtain non-negative weights {a0, a1, ..., a...} S}, satisfying constraints The fusion form is as follows:

[0084]

[0085] Obtain the chromaticity fusion feature F C =X ' (or its linear mapping). This module improves color context modeling capabilities and alleviates color drift issues in low-light scenes while maintaining a lightweight design.

[0086] Luminance-Guided Cross-Blending (LCA): To achieve synergistic enhancement of structure and color, a luminance-guided LCA module (such as...) is introduced. Figure 4 The system dynamically adjusts the chromaticity response based on luminance characteristics and performs joint optimization by combining channel and spatial attention.

[0087] F col =Conv(Concat(F S ,F C ))

[0088] G S =Conv 1×1 (F S )

[0089] F add =SE(F col +βG S )

[0090] F out =MHSA(F add )+F add

[0091] Where Conv can be a 1×1 or 3×3 convolution, SE(·) represents Squeeze-Excitation channel attention, MHSA(·) represents multi-head self-attention, and β is a learnable guidance coefficient (which can be initialized to 0.2). Finally, F... out The image is fed into the reconstruction head and inversely transformed back to RGB space to obtain the enhanced image.

[0092] S3: Construct a composite loss for training and optimization to ensure structural fidelity and color consistency. To obtain the trained model described in step S2, end-to-end optimization is performed using a composite loss, jointly constraining aspects such as structural consistency, perceptual quality, pixel distribution, consistency, PSNR proxy, and color deviation. The composite loss is defined as:

[0093] L total

[0094] =α1L spatial +α2L perceptual +α3L distribution +α4L consistency +α5

[0095] L PSNR

[0096] +α6L color ,α i ≥0

[0097] Among them, L spatial For structural consistency loss based on structural similarity, L perceptual For the perceptual loss in the pre-trained feature space, L distribution pixel / luminance histogram Figure 1 Consistency or statistical distance constraints, L consistency For spatial / multi-scale consistency constraints, L PSNR The signal-to-noise ratio improvement term is substituted by mean square error (MSE), L color This refers to color deviation constraints in the Cb / Cr color space. The weights α1 to α6 can be adjusted based on validation set performance. The optimizer, learning rate strategy, batch size, and number of training epochs are engineering parameters that can be set according to resource and task requirements and do not constitute limitations.

[0098] S4: Visualize the enhanced images generated by S2 and S3, and perform optional performance evaluation. To test the model's generalization ability and effectiveness in low-light scenes, common metrics such as PSNR, SSIM, NIQE, and BRISQUE can be used for evaluation (evaluation is not a necessary step in the method). Among them:

[0099]

[0100] In the formula, R is the maximum pixel value (R = 255 for an 8-bit image), MSE is the mean square error; μ, σ, σ xy These represent the mean, variance, and covariance, respectively; C1 and C2 are stable terms. NIQE and BRISQUE are no-reference image quality metrics used to quantify naturalness and spatial quality. Figure 5 This is a schematic diagram of YCbCr conversion. Figure 6Examples of enhanced comparisons and visualizations of different methods in typical scenarios.

[0101] In summary, the low-light image enhancement method combining multi-channel parallel attention and cross-fusion described in this invention can significantly reduce model complexity while maintaining high enhancement quality, effectively suppress color cast and fusion artifacts, and improve structural details and color consistency. It is suitable for real-time or near-real-time deployment on mobile / edge devices.

[0102] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A low-light image enhancement method combining multi-channel parallel attention and cross fusion, characterized in that, The method comprises the following steps: S1: inputting a low-illumination image into a YCbCr color space, separating a luminance (Y) channel and a chrominance (Cb, Cr) channel, and obtaining corresponding luminance and chrominance features. S2: performing enhancement processing on the features based on a trained lightweight model: in the Y channel, extracting structural features through a lightweight denoising module and a multi-head self-attention (MHSA) extraction structure, in the Cb and Cr channels, extracting chrominance context features through a multi-channel parallel attention module (MCPA), and performing feature fusion through a luminance-guided cross-fusion module (LCA) to output an enhanced image; S3: a training stage for obtaining the trained model: optimizing the model with a multi-dimensional loss function, and jointly constraining from the dimensions of structural consistency, perceptual quality, pixel distribution, and color fidelity; S4: outputting the enhanced image.

2. The method of claim 1, wherein, The color space conversion of step S1 satisfies the following linear relationship: Y(i,j)=0.299R(i,j)+0.587G(i,j)+0.114B(i,j); Cb(i,j)=-0.168736R(i,j)-0.331264G(i,j)+0.5B(i,j)+0.5; Cr(i,j)=0.5R(i,j)-0.418688G(i,j)-0.081312B(i,j)+0.

5.

3. The method according to claim 1 or 2, characterized in that, The luminance (Y) channel processing of step S2 comprises: a) using a lightweight denoising module to suppress noise and enhance structure of the Y feature, wherein the denoising module comprises a depth separable convolution and / or a linear transformation of a 1×1 convolution; b) before multi-head self-attention (MHSA) calculation, performing down-sampling or pooling on the feature map to compress the token number, then performing linear mapping of query, key, and value, and completing attention calculation; c) the output end of the MHSA adopts residual connection and normalization; the number of attention heads h can take any value between 2 and 8.

4. The method according to claim 1 or 2, characterized in that, The multi-channel parallel attention module (MCPA) of step S2 extracts chrominance context through multi-scale pooling branches and performs weighted fusion, wherein: a) input features X are sent into s scale branches to get {F1,...,F s}, the scale set can take one or more of {1,2,3,6}; b) non-negative dynamic weights {a1,...,a s} are obtained by channel-wise Softmax normalization, which are generated from the attention map produced by the 1x1 convolution on X; c) the fusion form is: and d) each scale branch adopts adaptive average pooling to the target size and then up-samples back to the original size before participating in fusion.

5. The method according to any one of claims 1 - 4, characterized in that, The luminance-guided cross-fusion module (LCA) of step S2 comprises: a) Cascade fusion: F col = Conv(Concat(F S , F C )), where Conv is a 1x1 or 3x3 convolution, F S is the luminance branch feature, and F C is the chrominance branch feature; b) brightness guide: G S = Conv 1×1 (F S ), and weighted by channel attention (SE), F add = SE(F col + βG S ), where β is a learnable guide coefficient; c) Global dependency modeling with residuals: F out = MHSA(F add ) + F add , for subsequent reconstruction.

6. The method according to any one of claims 1 - 5, characterized in that, To obtain the trained model of step S2, a composite loss is used for training optimization, and the composite loss is defined as: L total = a1L spatial + a2L perceptual + a3L distribution + a4L consistency + a5L PSNR + a6L color where a1~a6are non-negative weights; a i ≥ 0, i = 1, …, 6. L spatial is the structure consistency loss based on structural similarity, L perceptual is the perceptual loss based on pre-trained features, L distribution is the pixel distribution / histogram consistency loss, L consistency is the spatial or multi-scale consistency loss, L PSNR is the signal-to-noise ratio improvement loss in proxy of mean square error, L color is the color bias constraint on the chroma channel.

Images