Efficient progressive low-illumination image enhancement training method

Abstract

The invention provides a high-efficiency progressive low-illumination image enhancement training method. The method comprises the following steps: carrying out rapid enhancement illumination processing on a low-illumination input image IL by adopting a rapid illumination adjustment module to obtain an enhanced image IEM; performing color correction and noise reduction on the enhanced image IEM by using a global rapid enhancement module to obtain an enhanced image IEI; and extracting large-scale context information, strip-shaped context information, global features and frequency domain information of the low-illumination input image IL and the enhanced image IEM by using a multi-scale image recovery network, and obtaining a recovered image through Laplace transformation and inverse Laplace transformation.

Description

Technical Field

[0001] The present invention relates to an image training method, in particular to an efficient progressive low-light image enhancement training method. Background Art

[0002] Images taken under low-light conditions usually exhibit a darker and blurrier visual effect, accompanied by significant detail loss, which has a serious impact on subsequent image processing tasks. In actual detection scenarios, low-light conditions are often inevitable. Therefore, how to effectively enhance low-light images has become a key challenge in image processing research. Low-light images usually have two main problems: one is that insufficient light causes the image to become darker, and the other is the significant noise problem. Although methods such as histogram equalization or gamma correction can increase the image brightness, due to inaccurate adjustment, color distortion is often introduced, and the noise problem remains unsolved. The latest methods by Jiang (2023), Hou (2024), Zamir (2022), Xu (2022), Liang (2023), etc. have achieved good results in light adjustment. However, these light enhancement networks are usually very complex, lack the ability of fast light enhancement, and perform poorly in denoising. Summary of the Invention

[0003] The purpose of the present invention is to provide an efficient progressive low-light image enhancement training method, including:

[0004] Step S100, using a fast light adjustment module to perform fast enhanced light processing on the low-light input image I L to obtain an enhanced image I EM ;

[0005] Step S200, using a global fast enhancement module to perform color correction and noise reduction on the enhanced image I EM to obtain an enhanced image I EI ;

[0006] Step S300, using a multi-scale image restoration network to extract the large-scale context information, strip-shaped context information, global features, and frequency domain information of the low-light input image I L and the enhanced image I EM and obtaining a restored image through Laplace transform and inverse Laplace transform.

[0007] Furthermore, an objective function is established to optimize the network used in training. The objective function of the fast light adjustment module is a color loss function and a brightness loss function; the objective function of the global fast enhancement module is an L1 loss function, an L SSIM loss function, and a structure loss function; the loss function of the multi-scale image restoration network is a reconstruction loss function and a low-frequency consistency supervision function.

[0008] Furthermore, the specific process of step S100 includes:

[0009] Step S101, convert the low-light input image I L into an HSV image, and use histogram equalization technology on the V channel of the HSV image to extract the V-Map of the image;

[0010] Step S102, adjust the V-Map through the V-Map adjustment module to obtain the V-Map adjustment map;

[0011] Step S103, through the connection module, connect the V-Map and I L for low-light adjustment to obtain the image I L-a ;

[0012] Step S104, fuse the image I L , the V-Map adjustment map, and the image I L-a to obtain the enhanced image I EM .

[0013] Furthermore, in step S102, the V-Map adjustment module sets two multi-layer perceptrons

[0014] Furthermore, the specific process of step S200 includes:

[0015] Step S201, use the convolutional layer to obtain the local information and high-frequency details of the enhanced image I EM ;

[0016] Step S202, use the fast single-head attention module to perform color correction and noise reduction on the enhanced image I EM through learning and training to obtain the enhanced image I EI .

[0017] Furthermore, the specific process of step S300 includes:

[0018] Step S301, adopt the Laplace decomposition method to separate the different frequency features of the image I L and the enhanced image I EI , and stack these features along the channel dimension to obtain the stacked map M S ;

[0019] Step S302, perform downsampling on the stacked map M S to reduce the resolution;

[0020] Step S303, extract the deep representation through the multi-scale image restorer, and the deep representation includes large-scale context information, strip-shaped context information, global features, and frequency domain information;

[0021] Step S304, restore it to a high-resolution image representation through the transposed convolution of the encoder network's upsampling;

[0022] Step S305, the high-resolution image uses convolution to generate a learned residual image, and adds it to the stacked graph M S to obtain the finally restored enhanced Laplacian graph M d ;

[0023] Step S306, perform an inverse Laplace transform on M d to obtain the restored image m r .

[0024] Furthermore, the multi-scale image restorer in step S303 includes a large-scale branch, a local branch, a global branch, and a frequency domain branch; the specific process of step S303 includes:

[0025] Step S3031, capture the local detail information of the downsampled stacked graph M S through the large convolution kernel in the large-scale branch, and extract the large-scale context information;

[0026] Step S3032, capture the strip-shaped context information of the downsampled stacked graph M S through the 1×K and K×1 depth convolutions in the local branch;

[0027] Step S3033, obtain the global features of the downsampled stacked graph M S through the fast single-head self-attention module in the global branch;

[0028] Step S3034, capture the high-frequency and low-frequency frequency domain information of the downsampled stacked graph M S through the frequency domain branch;

[0029] Step S3035, downsample the large-scale context information, strip-shaped context information, global features, and frequency domain information and splice them with the stacked graph M S to obtain the deep representation.

[0030] Compared with the prior art, the present invention has the following advantages: (1) The present invention designs a lightweight fast low-light enhancement network, including a fast illumination adjustment module and a global fast enhancement module, which can quickly enhance the image illumination and remove noise with only 25K parameters; (2) The present invention designs a multi-scale image restoration network, which combines a large convolution kernel, a strip attention mechanism, a frequency domain attention mechanism, and a dual-domain channel attention mechanism. This network can efficiently capture multi-scale features from global to local and from spatial domain to frequency domain, effectively remove noise and further optimize illumination restoration; (3) The present invention combines a fast low-light enhancement network and a multi-scale restoration network, which can efficiently restore illumination and details while removing noise.

[0031] The present invention will be further described below in conjunction with the accompanying drawings of the specification. Description of the Drawings

[0032] Figure 1 It is a schematic flow diagram of the method of the present invention.

[0033] Figure 2 It is a schematic diagram of the visual comparison of the enhancement results of different methods on LOLv1 and LOLv2.

[0034] Figure 3 It is a schematic diagram of the visual comparison of the DICM, LIME, MEF, NPE and VV data sets.

[0035] Figure 4 It is a schematic diagram of the visual comparison of the input UIEB underwater image and the EPIENet enhanced image. Detailed Implementation Manner

[0036] Combined with Figure 1 , an efficient progressive low-light image enhancement training method includes the following steps:

[0037] Step S100, using a fast illumination adjustment module (Fast Illumination Adjustment module, FIA in the figure) to perform fast enhanced illumination processing on the low-light input image I L to obtain an enhanced image I EM ;

[0038] Step S200, using a global fast enhancement module (Global Fast Enhancement module, GFE in the figure) to perform color correction and noise reduction on the enhanced image I EM to obtain an enhanced image I EI ;

[0039] Step S300, using a multi-scale image restoration network to extract the large-scale context information, strip-shaped context information, global features, and frequency domain information of the low-light input image I L and the enhanced image I EM , and obtaining a restored image through Laplace transform and inverse Laplace transform.

[0040] In step S100, the enhanced image I EM = FIA(I L ), and the specific process includes:

[0041] Step S101, converting the low-light input image I L into an HSV image, and using histogram equalization technology for the V channel of the HSV image to extract the V-Map of the image;

[0042] Step S102: Input the V-Map into the V-Map adjustment module (V-Map adjustment in the figure) and the connection module (in the figure ). Among them, in the V-Map adjustment module, adjust the V-Map through two multi-layer perceptrons (MLP in the figure) to obtain the V-Map adjustment map (V-adjustment-Map in the figure), and in the connection module, perform low-light adjustment with I L to obtain the image I L-a (Low-light adjustment image in the figure);

[0043] Step S103: Fuse the image I L , the V-Map adjustment map, and the image I L-a to obtain the enhanced map I EM .

[0044] In Step S102, in the V-Map adjustment module, calculate the difference and obtain the overshoot parameter through two multi-layer perceptrons (MLP) to prevent image overexposure. In Step S103, the fusion goal of the three images is to improve the overall image quality through complex transformations and feature fusions, especially to restore details and color balance under low-light conditions.

[0045] In Step S200, obtain the enhanced image I EI = GFE(I EM ), and the specific process includes:

[0046] Step S201: Use the convolutional layer (the first DConv3 convolutional layer in GFE in the figure) to obtain the local information and high-frequency details of the enhanced map I EM ;

[0047] Step S202: Use the fast single-head attention module (FSSA module in GFE in the figure) to perform color correction and noise reduction on the enhanced map I EM during learning and training to obtain the enhanced image I EI .

[0048] The input feature F in (the enhanced map I EM ) after convolution to the FSSA module described in Step S202 has a shape of H×W×C, where C represents the number of channels, and H×W specifies the spatial position of the pixels. The Transformer architecture of the FSSA module includes an unbiased fully connected layer (Reshape layer in the figure) and a fully connected layer (the layer represented by C×C in the figure). Project the input feature F in to the query (Q), key (K), and value (V) through the Reshape layer and the linear layer (Linear layer in the figure): Q = XW Q , K = XWK , V = XW V , where the shapes of Q, K, and V are HW×C; W Q , W K , W V are the projection matrices for query, key, and value respectively; X is a matrix obtained by transforming the input feature F in with the shape of H×W×C; the self-attention mechanism is defined as Attention(Q, K, V) = V·softmax(K T Q / α), where softmax() transforms the output of the fully connected layer into a probability distribution, and α is a learnable parameter; adding the learnable positional information encoding P to obtain F out .

[0049] In step S300, a multi-scale image restoration network (MIRNet in the figure) is established to restore the image. In order to enable the multi-scale image restoration network to capture more image details, Laplacian decomposition is used to separate the different frequency features of the image I L and the enhanced image I EI , and stack these features along the channel dimension:

[0050] M L = Lap(I L )

[0051] M EI = Lap(I EI )

[0052] Use the stacked graph M S = Stack[M L , M EI as the input of the multi-scale restoration network.

[0053] The specific process of step S300 includes:

[0054] Step S301, perform downsampling to reduce the resolution on the stacked graph M S , increase the number of channels and reduce the spatial dimension;

[0055] Step S302, extract the deep representation through the multi-scale image restorer (MIR) module;

[0056] Step S303, restore it to a high-resolution image representation through the transposed convolution of the upsampling of the encoder network to expand the spatial dimension and halve the number of channels;

[0057] Step S304, the high-resolution image uses convolution to generate the learned residual image, and adds it to the stacked graph M S to obtain the finally restored enhanced Laplacian graph M d ;

[0058] Step S305, perform the inverse Laplace transform on M d to obtain the restored image m r (Restored Image in the figure).

[0059] In step S301, downsampling reduces the resolution, increases the number of channels, and decreases the spatial size, enabling the network to learn multi-scale features in a more compact representation while reducing the computational cost.

[0060] In step S302, the MIR module includes a large-scale branch (Large branch in the figure), a local branch (Local branch in the figure), a global branch (Global branch in the figure), and a frequency domain branch (Frequency domain branch in the figure). The specific process of step S302 includes:

[0061] Step S3021, capture the local detail information of the downsampled stacked graph M S through the large convolutional kernel in the large-scale branch, and extract the large-scale contextual information;

[0062] Step S3022, capture the strip-shaped contextual information of the downsampled stacked graph M S through the depthwise convolutions of 1×K and K×1 in the local branch;

[0063] Step S3023, obtain the global features of the downsampled stacked graph M S through the fast single-head self-attention module (FSSA) in the global branch;

[0064] Step S3024, capture the high-frequency and low-frequency frequency domain information of the downsampled stacked graph M S through the frequency domain branch;

[0065] Step S3025, downsample the large-scale contextual information, strip-shaped contextual information, global features, and frequency domain information and splice them with the stacked graph M S to obtain the deep representation.

[0066] In step S3021, the contextual information refers to the relationship between pixels, including local and global relationships, which can help the network understand the overall structure of the image rather than just the value of a single pixel.

[0067] In step S3024, the frequency domain branch includes a dual-domain channel attention module (DCAM in the figure) and a frequency-based attention module (FAM in the figure).

[0068] For the given input downsampled stacked graph M in DCAM S, first apply frequency channel attention (FCA) to it

[0069] X F = W 1×1 (GAP(M S ))

[0070]

[0071] where, W 1×1 represents a 1×1 convolution, GAP represents global average pooling, X FCA represents the output of the frequency channel attention, F1, F2 respectively represent the two-way fast Fourier transform after GAP, X F represents the intermediate result, represents the element-wise multiplication operation. Through Fourier processing, the global features can be effectively adjusted according to the spectral convolution theorem. Secondly, the modulated X FCA in the frequency domain is input into the spatial channel attention module (SCA) to obtain X DCAM

[0072]

[0073] where, IF represents the inverse Fourier transform, and DCAM only enhances the dual-domain features at the channel level in a coarse-grained manner.

[0074] In FAM, a frequency-based attention module is applied to refine the spectrum in a fine-grained manner to obtain X FAM

[0075]

[0076] FAM helps the network better focus on the frequency components related to reconstructing high-quality images and capture important information in the frequency domain.

[0077] In this embodiment, a 63×63 deep convolution kernel is used in the large branch. This large convolution kernel can capture local details and extract large-scale context information, enhance the contrast between bright and dark regions in the low-light image, so as to achieve a more natural brightening effect. It also expands the receptive field of the network, enabling it to capture local and global information, which is crucial for restoring the contours and structures of underexposed regions.

[0078] In the local branch, considering the key role of local information in image restoration and inspired by strip-based self-attention, 1×K and K×1 deep convolutions are introduced to capture strip-shaped context information and extract features in different directions. This design enables this branch to more comprehensively integrate multi-scale information.

[0079] In the global branch, the multi-scale image restorer (MIR) is mainly trained on cropped image patches of size 3×256×256, and the resulting bottleneck features have a spatial size of 64×64. Therefore, the largest odd convolutional kernel is used in the large branch. However, in the inference stage, the input image is much larger than the training patches, resulting in a 63×63 convolutional kernel being insufficient to cover the global context. To address this issue, the global modeling ability is enhanced by introducing a fast single-head self-attention module, which can quickly and effectively process the structure and content of the image while controlling the computational cost.

[0080] In the frequency domain branch, frequency domain processing can capture high-frequency and low-frequency information, helping the network understand the structure and texture of the image. It is more robust to noise and blur than spatial domain methods. In low-light images, frequency domain processing reveals brightness variations and edge features, enhancing the restoration effect.

[0081] For the training of the fast illumination adjustment module, the global fast enhancement module, and the multi-scale image restoration network, an objective function (loss) is established to optimize the parameters in the module / network. In the fast illumination adjustment module, L color and L brightness loss functions are introduced. The L color loss function is used to address the color deviation caused by inaccurate illumination, ensuring consistency with the reference image and thus improving the visual quality; the L brightness loss function makes the brightness order of the enhanced image match that of the reference image, thereby enhancing the brightness perception. In the global fast enhancement module, L structure 、L1 and L SSIM loss functions are introduced. The L structure loss function performs structure matching by comparing gradient information, improving details and reducing noise; the L1 loss function focuses on pixel-level differences; the L SSIM loss function considers structural similarity, providing a comprehensive framework for image quality improvement.

[0082] L FLIE = L1 + L SSIM + L color + L brightness + L structure

[0083] In the multi-scale image restoration network, the reconstruction loss function L r and the low-frequency consistency supervision function L i are introduced. The L r is used to guide the multi-scale prediction M d to be closer to the ground truth decomposition map M gt , which are obtained through the same Laplacian decoupling pipeline; the L i is used to maintain the illumination restoration result of the first stage

[0084]

[0085] L MIR = L r + β·L i

[0086] where K represents the Laplacian hierarchy, and L MIR represents the total loss of the multi-scale image restoration network, and m r , m gt , m l represent the restored image, the ground truth image, and the low-light input result after Laplacian decomposition respectively, and β represents the balanced scaling factor.

[0087] Comparative Example

[0088] 1. Experimental Setup

[0089] Datasets: Experiments were conducted on multiple low-light image enhancement (LLIE) and underwater image enhancement (UIE) datasets. In the LLIE task, the method of the present invention was applied to paired and unpaired datasets. For paired data, the LOLv1 and LOLv2 datasets were used, which are benchmarks for the low-light image enhancement task. LOLv1 contains 485 images for training and testing, while LOLv2 is divided into two subsets: real capture and synthetic. The real capture subset (LOLv2Real) contains 689 training images and 100 test images, and the synthetic subset (LOLv2 Synthetic) includes 900 training images and 100 test images. For unpaired data, five datasets were used: LIME, MEF, NPE, VV, and DICM. For the UIE task, the UIEB dataset was used, and the UIEB dataset was further divided into a training set, a validation set (R90), and a test set (C60).

[0090] Evaluation Metrics: For paired datasets, PSNR and SSIM were used to evaluate pixel-level accuracy, and LPIPS was used to evaluate perceptual quality. For real-world datasets, NIQE was used as a reference-free metric. In the UIE task, UIQM and UCIQE were also used to evaluate image quality.

[0091] Implementation Details: The framework was implemented using PyTorch and run on a single NVIDIA 2080Ti or 3080 GPU. In the fast low-light enhancement stage, the learning rate was fixed at 1×10 -4 , the batch size was 40, and the images were randomly cropped to 200×200. In the multi-scale image restoration stage, the Adam optimizer and the cosine annealing strategy were used to dynamically adjust the learning rate, and the initial learning rate was 5×10 -4, the batch size is 4, and the images are randomly cropped to 256×256. In the test phase, the input images are padded using reflection padding to make their size a multiple of 8×8. After inference, the padded images are cropped back to the original size.

[0092] 2. Evaluation Results

[0093] The method involved in the present invention is implemented based on the Efficient Progressive Low-Light Image Enhancement Network (EPIENet). EPIENet includes a lightweight fast low-light enhancement network (FLIENet) and a multi-scale image restoration network (MIRNet). FLIENet contains a fast illumination adjustment module (FIA) and a global fast enhancement module (GFE).

[0094] Full-reference evaluation: The present invention is quantitatively compared with the current state-of-the-art methods on the LOL-v1 and LOL-v2 datasets. The detailed results are shown in Table 1 (the gray model adjusts its output color using the ground truth mean (GT-Mean) value. For each part, the best result is shown in bold, and the second-best result is underlined). The table is divided into two parts: the first part compares the models that use the ground truth image mean for color adjustment, and the second part lists the models that do not rely on this adjustment. On the LOL-v1 and LOL-v2 datasets, the method of the present invention performs well in most metrics. Especially on the LOL-v2-syn dataset, EPIENet achieved an impressive PSNR value of 30.88 dB, which is 4.84 dB higher than the second-place MRQ. It can be observed that existing methods usually have difficulty balancing high perceptual quality (measured by LPIPS) and pixel-level accuracy. In contrast, EPIRNet performs excellently in both, achieving the highest SSIM (0.960) and the lowest LPIPS (0.029). In addition, as shown in Table 2, FLIENet also performs strongly. Compared with the equally efficient FLIGHTNet, the PSNR of FLIENet is increased by 5.71 dB, the SSIM is increased by 4.6 points, and it has a very fast running time. Table 4 further demonstrates the excellent SSIM performance of our model in the UIE task.

[0095] No-reference evaluation: For no-reference low-light images, on the unannotated datasets of DICM, LIME, MEF, NPE, and VV, the effect of FLIENet trained on LOLv1 or LOLv2 is evaluated, and its performance is measured by the NIQE metric. The comparison with the state-of-the-art methods is shown in Table 3, where the method of the present invention has a significant improvement in the NIQE metric. For no-reference underwater image enhancement, as shown in Table 4, the method of the present invention is also evaluated on the C60 test set.

[0096] Qualitative comparison: In the low-light task,Figure 2 and Figure 3 shows a visual comparison of EPIENet with state-of-the-art algorithms. Existing methods often result in color distortion, contain overexposed or underexposed regions, fail to effectively suppress noise, or generate blurry images with black dots and unnatural artifacts. In contrast, EPIENet effectively enhances low visibility, low contrast, or low light regions, reliably removes noise without introducing black dots or artifacts, and robustly preserves colors. It can also be observed that FLIENet performs well in restoring illumination, although there is still some noise. Note that Figure 3 the five datasets in Figure 3 make the visual results in Figure 3 more convincing and fair, providing stronger evidence of the effectiveness of the method of the present invention. As Figure 4 shown, EPIENet significantly improves image brightness and perceptual quality while ensuring reasonable color accuracy. In the underwater image enhancement task, as Figure 4 shown, our EPIENet performs excellently.

[0097] Table 1 Quantitative comparison of LOL-v1 and v2 datasets using PSNR, SSIM, and LPIPS

[0098]

[0099] Table 2 Performance table of Fast Low Light Enhancement Network

[0100]

[0101] Table 3 No-reference image evaluation table of Fast Low Light Enhancement Network

[0102]

[0103] ↓ in terms of NIQE. The best results are shown in bold.

[0104] Table 4 Quantitative comparison of UIEB-R90 and UIEB-C60 datasets in terms of PSNR, SSIM, UIQM, and UCIQE

[0105]

[0106] 3. Ablation experiments

[0107] To verify the effectiveness of the network and its sub-modules of the present invention, eight ablation experiments were conducted on the LOLv2-syn dataset, and the quantitative results are shown in Table 5. In the first stage of the ablation experiment, the V-Map adjustment module (VA) and the fast single-head self-attention module (FSSA) were tested. The VA module improved the lighting effect, but also highlighted the necessity of the FSSA module. Without FSSA, there were still problems of spatial noise and color inconsistency in the output image, resulting in lower PSNR and SSIM metrics. The FSSA module effectively solved these problems. In the second stage, the FLIE enhancement results were input into MIRNet for training, and the local, large-scale, global, and frequency-domain branches were gradually added to the MIR model. The local branch improved the detail recovery by enhancing local information and reducing noise, and improved PSNR and SSIM. The large-scale branch used depthwise separable convolutions, had a larger receptive field, captured more global information, maintained efficiency, and improved the network's adaptability to lighting changes. In addition, the global and frequency-domain branches enhanced by FSSA and the dual-domain channel attention mechanism significantly improved the performance.

[0108] Table 5 Results of the ablation study on the LOL-v2-syn dataset

[0109]

[0110] 4. Conclusion

[0111] FIENet enhances the lighting by calculating and adjusting the V channel of low-light images, reducing the high computational cost brought by complex models. It also uses a global fast enhancement module to effectively denoise and optimize lighting and color. The experimental results show that the lighting enhancement network is both fast and lightweight, and the effect is significant. To further improve denoising and lighting correction, a progressive multi-scale recovery module was added to MIRNet, which includes four branches: a global branch, a frequency-domain branch, a large-scale branch, and a local branch. These branches work together to learn feature representations from global to local and from spatial to frequency domain, effectively removing noise and optimizing lighting recovery.

Claims

1. An efficient progressive low-light image enhancement training method, characterized in that: include: Step S100: using a fast illumination adjustment module to adjust the low-light input image I L Perform rapid enhancement lighting processing to obtain enhanced image I EM ; Step S200: using a global fast enhancement module to enhance the image I EM Perform color correction and noise reduction to obtain enhanced image I EI ; Step S300, using a multi-scale image restoration network to extract the low-light input image I L and enhanced image I EM The large-scale context information, strip shape context information, global features, and frequency domain information are used to obtain the restored image through Laplace transform and inverse Laplace transform.

2. The method according to claim 1, characterized in that Establish an objective function to optimize the network used for training. The objective function of the fast illumination adjustment module is the color loss function and the brightness loss function; the objective function of the global fast enhancement module is the L1 loss function, L SSIM Loss function and structural loss function; the loss function of the multi-scale image restoration network is the reconstruction loss function and the low-frequency consistency supervision function.

3. The method according to claim 1 or 2, characterized in that: The specific process of step S100 includes: Step S101: input low-light image I L Convert it into an HSV image, and use histogram equalization technology to extract the V-Map of the image from the V channel of the HSV image; Step S102, adjusting the V-Map by a V-Map adjustment module to obtain a V-Map adjustment map; Step S103, connect V-Map and I L Perform low light adjustment to obtain image I L-a ; Step S104: image I L 、V-Map adjustment map、Image I L-a Fusion enhanced Figure I EM .

4. The method according to claim 3, characterized in that In step S102, the V-Map adjustment module sets two multi-layer perceptrons.

5. The method according to claim 1 or 2, characterized in that: The specific process of step S200 includes: Step S201, using the convolutional layer to obtain the enhanced image I EM Local information and high-frequency details; Step S202: Use a fast single-head attention module to enhance image I EM Learning training to perform color correction and noise reduction to obtain enhanced image I EI .

6. The method according to claim 1 or 2, characterized in that: The specific process of step S300 includes: Step S301: Use Laplace decomposition method to transform image I L and enhanced image I EI The different frequency features of are separated and these features are stacked along the channel dimension to obtain the stacked graph M S ; Step S302: stacking graph M S Perform downsampling to reduce resolution; Step S303, extracting a deep representation through a multi-scale image restorer, where the deep representation includes large-scale context information, strip shape context information, global features, and frequency domain information; Step S304, restoring the high-resolution image representation through up-sampled transposed convolution of the encoder network; Step S305: The high-resolution image is generated by convolution to learn the residual image and combined with the stacked image M S Add together to get the final restored enhanced Laplacian graph M d ; Step S306: d Perform inverse Laplace transform to get the restored image m r .

7. The method according to claim 6, characterized in that The multi-scale image restorer in step S303 includes a large-scale branch, a local branch, a global branch, and a frequency domain branch; the specific process of step S303 includes: Step S3031, capture the downsampled stacked image M through the large convolution kernel in the large-scale branch S local detail information and extract large-scale context information; Step S3032, capture the downsampled stacked image M through 1×K and K×1 depth convolution in the local branch S Contextual information of the strip shape; Step S3033, obtain the downsampled stacked image M through the fast single-head self-attention module in the global branch S The global characteristics of Step S3034, capturing the down-sampled stacked graph M through the frequency domain branch S The high-frequency and low-frequency frequency domain information; Step S3035: downsample the large-scale context information, the stripe-shaped context information, the global features, and the frequency domain information and combine them with the stacked image M. S Concatenate to get deep representation.

Images