A multi-attention low-illumination image enhancement method based on M-LBP guidance

Abstract

The application discloses a kind of based on M-LBP guide's multi-attention low-illumination image enhancement method, belong to image processing technical field.It includes the following steps: first, to the input low-illumination image, using improved M-LBP feature extraction method generates the feature map containing texture information;Second, using cyclic decomposition module, the input low-illumination image is decomposed into relatively independent illumination component and reflection component;Using illumination adjustment module, the illumination component is enhanced, and the enhanced illumination component is obtained;Using double-branch denoising module, the final reflection component is denoised, and the denoised reflection component is obtained;Finally, the illumination component after enhancement processing is fused with the reflection component after denoising processing, and the enhanced image is obtained.The application improves the brightness of low-illumination image, and through texture guide, double-branch denoising retains details and suppresses noise, improves subjective visual clarity and color authenticity.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a multi-attention low-light image enhancement method based on M-LBP (Mean-Local Binary Pattern). Background Technology

[0002] Low-light images commonly suffer from low contrast, unclear details, severe noise, and low visibility, affecting not only visual quality but also interfering with information transmission in advanced tasks such as object detection and image segmentation. Low-light image enhancement technology, by improving brightness, correcting contrast, and suppressing noise, has become a crucial aspect of image processing. Current low-light enhancement algorithms largely originate from Retinex theory, deep learning, and histogram equalization. However, traditional Retinex algorithms are prone to color casts and artifacts due to inaccurate illumination estimation; while histogram equalization can brighten images, it can over-enhance noise and lose local details. Furthermore, because noise and detail are highly coupled in low-light images, the multi-objective optimization of "enhancement-denoising-detail preservation" presents a contradiction. Therefore, a more precise feature guidance mechanism and a multi-module collaborative architecture are needed to balance the constraints of each objective.

[0003] Currently, although some progress has been made in the research of low-light image enhancement methods, most methods only focus on the optimization of a single target, and problems such as color distortion and noise amplification will occur during the enhancement process. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a multi-attention low-light image enhancement method based on M-LBP guidance, which can effectively enhance the image while protecting the image's edge texture and other details.

[0005] Technical solution: A multi-attention low-light image enhancement method based on M-LBP guidance, comprising the following steps:

[0006] S1, an improved M-LBP feature extraction method is used to generate M-LBP feature maps containing texture information for low-light images;

[0007] S2, through the cyclic decomposition module, the low-light image is decomposed into illumination and reflection components for the second time, first obtaining the initial illumination component. With the initial reflection component Then, the initial reflection component The final illumination component is obtained by secondary decomposition. With the final reflection component The M-LBP feature map containing texture information is superimposed with the initial illumination component. With the final illumination component As input to the illumination adjustment module, the M-LBP feature map containing texture information and the final reflection component are used. As input to the dual-branch denoising module;

[0008] S3, uses a lighting adjustment module to adjust the superimposed initial lighting components. Enhancement is performed to obtain the enhanced illumination component; a dual-branch denoising module is used to denoise the final reflection component. Denoising is performed to obtain the denoised reflection component;

[0009] S4 multiplies the enhanced illumination component and the denoised reflection component element by element to output the final enhanced image.

[0010] Furthermore, the implementation steps of the improved M-LBP feature extraction method are as follows:

[0011] S11: Arrange the pixels within the selected window area radially in eight directions, and calculate the average value of the center pixel together with the neighboring pixels in each of the eight directions. Then with 8 The mean value is used as the center pixel. ;

[0012] S12, the 8 in this area With the center pixel respectively Comparison: If If the value is 1, the corresponding neighborhood is assigned a value of 1; otherwise, it is assigned a value of 0.

[0013] S13. Take the result obtained in step S12 and form an eight-bit binary value in a clockwise direction. Then convert the eight-bit binary value into a decimal number from 0 to 255 as the M-LBP value.

[0014] S14, repeat steps S11-S13 for each pixel of the image, and finally form an M-LBP feature map from all M-LBP values.

[0015] Furthermore, the decomposition network DecomNet in the recurrent decomposition module consists of U-Net and residual blocks. The decomposition network DecomNet adopts a hybrid long-short hop connection method. During the downsampling of the input image, Conv 4×4 with stride=2 is used to replace the max pooling layer, and residual blocks are applied to each layer of the encoder-decoder, where Conv represents convolution.

[0016] Furthermore, the illumination adjustment module consists of a multi-level feature attention module and a U-Net, where the U-Net is an encoder-decoder structure. The processing procedure is as follows: using superimposed initial illumination components... and final illumination component As input, basic features are first extracted through a single-layer convolutional layer and channel attention blocks. The channel attention blocks learn channel weights and highlight high-information feature channels through a squeezing-activation-attention process. Subsequently, the encoder downsamples the features to a small scale, and the decoder reconstructs the local illumination distribution using large-scale illumination information. At the same time, the M-LBP feature map scaled up to the same scale is fed together with the encoder / decoder features into a multi-level feature attention module. Texture details are extracted and enhanced through three multi-level feature attention modules in sequence. Finally, the features processed by the multi-level feature attention modules are fused through residual dense blocks. The fused features are added to the residual of the original input image to generate the enhanced illumination component.

[0017] Furthermore, the dual-branch denoising module uses the M-LBP feature map and the final reflection component The input is a multi-feature perception module that extracts positive features in the upper branch and a multi-modal attention fusion module that inverts the M-LBP feature map in the lower branch to mine low contrast information. The extracted positive features and low contrast information are superimposed and then processed by the noise integration module and the feature refinement module in turn to remove the noise of the reflection component and obtain the denoised reflection component.

[0018] Furthermore, the multi-feature perception module uses parallel dilated convolutions (1x, 2x, and 3x) to extract multi-scale features, followed by channel attention and residual processing. The multimodal attention fusion module includes channel attention, pixel attention, and local residuals, generating channel weights through global average pooling and pixel weights through convolutional layers. The noise integration module consists of convolution, ReLU, and Tanh hyperbolic functions, while the feature refinement module consists of multiple sets of sequentially connected ReLU, BN, and convolutions. The noise integration module deeply fuses the noise-related features extracted from the upper and lower branches to generate a noise feature representation.

[0019] Compared with the prior art, the significant advantages of this invention are as follows:

[0020] 1. This invention relies on the synergistic effect of M-LBP feature extraction and cyclic decomposition modules. The cyclic decomposition module decomposes low-light images into illumination components and reflection components. Through the feature learning capabilities of the illumination adjustment module combined with multi-level channel attention, and the positive feature extraction and low contrast information mining of the dual-branch denoising module, a model structure of feature guidance-precise decomposition-component optimization-fusion output is formed. This not only solves the problem of brightness improvement of low-light images, but also preserves details and suppresses noise through texture guidance and dual-branch denoising. While removing noise from the reflection component, it preserves texture details and provides higher quality input data for subsequent advanced image processing.

[0021] 2. This invention generates feature maps containing texture information through an improved M-LBP feature extraction method. A unique dual-branch denoising module is designed for the reflection component. The upper branch extracts positive features through a multi-feature perception module, while the lower branch inverts the image and mines low-contrast information through a multi-modal attention fusion module. Denoising is then completed by combining a noise integration module and a feature refinement module. This results in a 1.4–9.8 dB improvement in peak signal-to-noise ratio (PSNR) and a 0.03–0.62 improvement in structural similarity (SSIM) compared to existing models. It effectively reduces the perceived image quality index (LPIPS) by 0.05–0.35 and the no-reference image quality assessment index (NIQE) by 0.5–2.4. Subjective visual sharpness and color fidelity are significantly improved, and objective performance is superior.

[0022] 3. Thanks to the sensitivity of M-LBP to texture and lighting features and the versatility of its network structure, this invention can directly adapt to RGB image enhancement in different low-light scenes such as low light and backlight without adjusting parameters for specific scenes. It has a wider range of applications and is more in line with practical application needs. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention;

[0024] Figure 2 This is a schematic diagram of the low-light image enhancement process of the present invention;

[0025] Figure 3 This is a schematic diagram of the calculation process of the improved M-LBP feature extraction method of this invention;

[0026] Figure 4 This is a schematic diagram of the decomposition process of the cyclic decomposition module of the present invention;

[0027] Figure 5 This is a structural diagram of the DecomNet decomposition network of this invention;

[0028] Figure 6 This is a structural diagram of the illumination adjustment module of the present invention;

[0029] Figure 7 This is a schematic diagram of the multi-level feature attention module of the present invention;

[0030] Figure 8 This is a structural diagram of the dual-branch noise reduction module of the present invention;

[0031] Figure 9 This is a schematic diagram of the multi-feature perception module of the present invention;

[0032] Figure 10 This is a schematic diagram of the multimodal attention fusion module of the present invention;

[0033] Figure 11 These are the original low-light images used in the experiments of this invention, where (a) is image A in the LOL dataset, (b) is image B in the LIME dataset, (c) is image C in the DICM dataset, and (d) is image D in the MEF dataset.

[0034] Figure 12 yes Figure 11 Image A in the figure uses a comparison of the enhancement effects of six models, where (a) is the enhancement effect of the LIME model, (b) is the enhancement effect of the Retinex-Net model, (c) is the enhancement effect of the KinD model, (d) is the enhancement effect of the RRDNet model, (e) is the enhancement effect of the zero-DCE model, and (f) is the enhancement effect of the model of the present invention.

[0035] Figure 13 yes Figure 11 Image B in the figure uses a comparison of the enhancement effects of six models, where (a) is the enhancement effect of the LIME model, (b) is the enhancement effect of the Retinex-Net model, (c) is the enhancement effect of the KinD model, (d) is the enhancement effect of the RRDNet model, (e) is the enhancement effect of the zero-DCE model, and (f) is the enhancement effect of the model of the present invention.

[0036] Figure 14 yes Figure 11 Image C in the figure uses a comparison of the enhancement effects of six models, where (a) is the enhancement effect of the LIME model, (b) is the enhancement effect of the Retinex-Net model, (c) is the enhancement effect of the KinD model, (d) is the enhancement effect of the RRDNet model, (e) is the enhancement effect of the zero-DCE model, and (f) is the enhancement effect of the model of the present invention.

[0037] Figure 15 yes Figure 11 Image D in the figure uses six model enhancement effect comparison diagrams, where (a) is the enhancement effect diagram of LIME model, (b) is the enhancement effect diagram of Retinex-Net model, (c) is the enhancement effect diagram of KinD model, (d) is the enhancement effect diagram of RRDNet model, (e) is the enhancement effect diagram of zero-DCE model, and (f) is the enhancement effect diagram of the model of the present invention. Detailed Implementation

[0038] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 , Figure 2As shown, this invention provides a multi-attention low-light image enhancement method based on M-LBP guidance, comprising the following steps:

[0040] Step 1: Use the improved M-LBP feature extraction method to generate feature maps containing texture information for low-light images;

[0041] Preferably, the M-LBP operator has a wider sampling range than the traditional 3×3 LBP operator, and can simultaneously preserve both local and global texture information of the image, avoiding the problem of insufficient texture extraction. In the improved M-LBP feature extraction method (i.e., the M-LBP algorithm) of this invention, a 5×5 window is used, and the calculation process is as follows: Figure 3 As shown, the specific implementation steps are as follows:

[0042] Step 11, arrange the pixels within the selected window area according to... Figure 3 The diagram shows a radial arrangement in eight directions, with the center pixel being averaged together with its neighboring pixels in each of the eight directions. Then with 8 The mean value is used as the center pixel. .

[0043] Step 12, select the 8 elements within the selected window area. With the center pixel If a comparison is made, If the value is 1, the corresponding neighborhood is assigned a value of 1; otherwise, it is assigned a value of 0.

[0044] Step 13: Take the result obtained in step 12 and arrange it into an eight-bit binary value in a clockwise direction. Then convert the eight-bit binary value into a decimal number from 0 to 255, which is the M-LBP value.

[0045] Step 14: Repeat steps 11-13 for each pixel of the image. All M-LBP values ​​will eventually form an M-LBP feature map.

[0046] Step 2: The low-light image is decomposed a second time into an illumination component and a reflection component containing very little light using the Recurrent Decomposition Module (RDec-Net), thus obtaining the initial illumination component. With the initial reflection component Then, the initial reflection component The final illumination component is obtained by secondary decomposition. With the final reflection component Superimposed initial illumination component With the final illumination component As input to the illumination adjustment module, the final reflection component As input to the dual-branch denoising module. The decomposition process is as follows: Figure 4 As shown.

[0047] Preferably, sequential decomposition reduces illumination interference, enhances details, and also amplifies subtle changes and noise in the image. The decomposition network DecomNet in the Recurrent Decomposition Module (RDec-Net) consists of U-Net and residual blocks, with the structure as follows: Figure 5 As shown, the DecomNet decomposition network employs a hybrid long-short hop connection method, which maximizes information transmission and minimizes information loss, thereby preserving image details during upsampling. During input image downsampling, this invention uses a 4×4 Conv layer with stride=2 instead of a max-pooling layer. Conv represents convolution, allowing for flexible control of computational load by adjusting the kernel size and stride. The convolutional layers also better preserve feature map information. Furthermore, residual blocks are applied to each layer of the encoder-decoder architecture to effectively address the vanishing gradient problem and accelerate network convergence. The encoder is responsible for downsampling and feature extraction of the input image; the decoder is responsible for upsampling and image reconstruction.

[0048] Step 3: The illumination adjustment module (IENet) based on the U-Net architecture, guided by the M-LBP feature map obtained in step 1 and the multi-level feature attention module (MFAB), adjusts the initial illumination components after superposition. Enhancement is performed, and then feature fusion is performed through residual dense blocks (RDB) to obtain the enhanced illumination component;

[0049] The structure of IENet is as follows Figure 6 As shown, the processing procedure is as follows: using the superimposed initial illumination components and final illumination component As input to the illumination adjustment module, basic features are first extracted through a single-layer convolutional layer and a Channel Attention Block (CAB). The CAB learns channel weights and highlights high-information feature channels through a squeeze-excitation-attention process. Subsequently, a U-Net-style encoder-decoder subnetwork is used. The encoder downsamples features to a small scale to capture a large-scale illumination distribution, enabling the subnetwork to adaptively adjust. The decoder reconstructs the local illumination distribution using large-scale illumination information. To compensate for the loss of detail during downsampling, the subnetwork maintains the feature mapping at the original image resolution to preserve local details. The M-LBP feature map (Ln) scaled at the same scale is fed together with the encoder / decoder features (Fn) into a multi-level Feature Attention Module (MFAB). Three MFABs sequentially perform targeted texture detail extraction and enhancement. Finally, the features processed by MFAB are fused through Residual Dense Blocks (RDB). The fused features are added to the residual of the original low-quality input image to generate the enhanced illumination component. The entire process achieves adaptive adjustment of global illumination while preserving local and texture details through multi-module collaboration.

[0050] The structure of MFAB is as follows: Figure 7 As shown, high-resolution features are first extracted using five-channel attention blocks (CABs), and then the resized contextual features are... The M-LBP feature maps of different sizes after being scaled at the same scale are added to the original resolution context features to refine the feature mapping of the entire image. It is also connected to the feature results processed by the five-channel attention blocks to highlight its texture features. CAB learns the channel weights through global pooling and fully connected layers to achieve adaptive brightening of dark areas.

[0051] Step 4: The dual-branch denoising module (DBDNet) uses the M-LBP feature map obtained in Step 1 and the final reflection component... As input, the upper branch extracts positive features through the Multi-Functional Feature Perception (DFPM) module, and the lower branch inverts the M-LBP feature map and mines low-contrast information through the Multi-Modal Attention Fusion (MAFM) module. The extracted positive features and low-contrast information are superimposed and then processed by the Noise Integration (NIM) module and the Feature Refinement (FRM) module in sequence to remove noise from the reflection component, resulting in the denoised reflection component.

[0052] The network structure of DBDNet is as follows: Figure 8 As shown. In the dual-branch denoising module (DBDNet), the M-LBP feature map containing texture information is combined with the noisy final reflection component. Both are used as input to preserve the texture details of the image and improve the performance of the noise reduction process. Preferably, this dual-branch noise reduction module consists of two branches, such as... Figure 9 As shown, the DFPM in the upper branch uses 1-Dilated Conv (one-fold dilated convolution), 2-Dilated Conv (two-fold dilated convolution), and 3-Dilated Conv (three-fold dilated convolution) in parallel to extract multi-scale features. After channel attention and residual processing, it can more accurately distinguish noise and real image information based on features at different scales, removing noise while preserving various details and structures of the image to the greatest extent. Figure 10As shown, the MAFM in the lower branch includes sequentially connected local residuals, channel attention, and pixel attention. Channel weights are generated through global average pooling, and pixel weights are generated by convolutional layers. This precisely focuses on pixel information in low-contrast regions, helping the network better capture and process features in these key areas, laying the foundation for subsequent denoising and image quality improvement. The noise integration module consists of convolution, ReLU, and Tanh hyperbolic functions, while the feature refinement module consists of multiple sequentially connected ReLU, BN, and convolution functions. The noise integration module deeply fuses and refines the noise-related features extracted from the upper and lower branches, generating a more accurate noise feature representation.

[0053] Step 5: Multiply the enhanced illumination component and the denoised reflection component element by element to output the final enhanced image.

[0054] To verify the effectiveness and feasibility of the low-light image enhancement method of this invention, the original low-light image selected is as follows: Figure 11 As shown in (a)-(b), according to Figure 2 The processing flow shown was validated. Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), learned perceptual patch similarity (LPIPS), and no-reference image evaluation metric (NIQE) were used to compare the results with five other state-of-the-art methods on four public datasets: LOL, LIME, DICM, and MEF.

[0055] The expression for Peak Signal-to-Noise Ratio (PSNR) is:

[0056] ,

[0057] in, The maximum pixel value of the image. To enhance the mean square error between the image and the reference image.

[0058] The expression for structural similarity (SSIM) is:

[0059] ,

[0060] Where x and y represent two normal images to be compared. express The average value, express The average value, It means variance represent variance express and covariance, , and represent the stability constants, respectively.

[0061] The expression for learning perceptual image patch similarity LPIPS is:

[0062] ,

[0063] in, for and The distance between them is used to measure the enhanced image. Compared with normal lighting images The degree of perceptual similarity between them; , The first The feature map height and width of the layer. For spatial normalization coefficients, For the feature extraction layer index, For the reference image patch feature vector, For the feature vector of the distorted image patch, This represents the square of the L2 norm.

[0064] The expression for the No-Reference Image Evaluation Metric (NIQE) is:

[0065] ,

[0066] in, , Let represent the mean vectors of the multi-valued Gaussian models of the natural image and the input image, respectively. , Let represent the covariance matrices of the multivalued Gaussian models of the natural image and the input image, respectively.

[0067] The following is combined Figures 12-15 Detailed explanation: Compared with LIME, Retinex-Net, KinD, RRDNet, and Zero-DCE enhancement models, the image enhancement method provided in this embodiment has a better effect on image enhancement.

[0068] Depend on Figures 12-15 It can be seen that in the LOL dataset (such as...) Figure 12 In (a)-(f) shown in the figure, the present invention can clearly restore the texture of books in the dark areas of the bookshelf without excessive brightness or noise. The KinD method has problems of excessive brightness and noise amplification. In the LIME dataset (such as Figure 13 As shown in (a)-(f) in the figure, the present invention can restore the fracture texture at the cliff block and the greenery in the lower right corner. The RRDNet method is generally dark and has severe edge distortion. In the DICM dataset (such as Figure 14In (a)-(f) shown in the figure, the method of the present invention has suitable lighting, and while preserving strong backlight, the color restoration of the church is closer to the original image and without distortion. The LIME method has poor restoration effect in strong backlight, with excessive local brightness, resulting in sharpening, color distortion and other problems. In the MEF dataset (such as... Figure 15 As shown in (a)-(f), the present invention can restore the details of the lawn and balloons in backlit scenes with natural colors, while the Zero-DCE method has residual dark areas.

[0069] As can be seen from the objective index values ​​in Tables 1 to 4, the image enhancement method proposed in this invention achieves relatively high-quality results in the four indicators of PSNR, SSIM, LPIPS, and NIQE.

[0070] Table 1. Quantitative evaluation results on the LOL dataset.

[0071]

[0072] Table 2. Results of quantitative evaluation on the LIME dataset.

[0073]

[0074] Table 3. Results of quantitative evaluation on the DICM dataset.

[0075]

[0076] Table 4. Results of quantitative evaluation on the MEF dataset.

[0077]

[0078] The results in Tables 1-4 clearly and effectively demonstrate that the M-LBP-guided multi-attention low-light image enhancement method proposed in this invention is superior both in terms of subjective human observation and in terms of objective image quality evaluation metrics. The method of this invention effectively enhances the input image, significantly improving image brightness, highlighting details in darker areas, preserving more feature details, and effectively removing image noise. This makes the enhanced image closer to the original high-quality image, which is more beneficial for practical applications.

[0079] A computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the low-light image enhancement method of the present invention. The processor includes a kernel that retrieves corresponding program units from memory.

[0080] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed, it implements the steps of the low-light image enhancement method of the present invention.

[0081] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0082] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the low-light image enhancement method of the present invention.

[0083] The steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular hardware and software combination.

[0084] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A multi-attention low-light image enhancement method based on M-LBP guidance, characterized in that, Comprising the following steps: S1, using an improved M-LBP feature extraction method to generate an M-LBP feature map containing texture information for low-illumination images; S2, secondary decompose the low-illumination image into illumination component and reflection component by the cycle decomposition module, to obtain initial illumination component and initial reflection component , and then decompose the initial reflection component to obtain final illumination component and final reflection component ; M-LBP feature maps containing texture information, superimposed initial illumination component with final illumination component M-LBP feature maps containing texture information and final reflectance component as input of the illumination adjustment module as input of the two-branch denoising module; S3, using the illumination adjustment module to adjust the initial illumination component after superposition enhanced illumination component is obtained. adopting a double-branch denoising module to the final reflection component performing denoising processing to obtain a denoised reflection component; S4, element-wise multiplication of the enhanced illumination component and the de-noised reflection component to output the final enhanced image; The implementation steps of the improved M-LBP feature extraction method are as follows: S11, the pixels in the selected window region are arranged radially in 8 directions, and the average value of the center pixel and the 8-direction neighborhood pixels is calculated , respectively , the average value of 8 directions is taken as the center pixel S12, the 8 respectively, with the center pixel are compared: if then the corresponding neighborhood is assigned the value 1; otherwise, the value 0. S13, the results obtained in step S12 are arranged in a clockwise direction to form an eight-bit binary value, and the eight-bit binary value is converted into a decimal number of 0-255 as an M-LBP value; S14, repeating steps S11-S13 for each pixel point of the image, and finally forming an M-LBP feature map from all M-LBP values; The light adjustment module is composed of a multi-stage feature attention module and a U-Net, the U-Net is an encoder-decoder structure, and the processing process is as follows: taking the superimposed initial light component and the final light component as input, the basic features are extracted through a single convolutional layer and a channel attention block, the channel attention block learns the channel weight and highlights the high information feature channel through the squeeze-excitation-attention process; Subsequently, the encoder sequentially down-samples the features to a small scale, and the decoder reconstructs the local illumination distribution using large-scale illumination information; at the same time, the M-LBP feature map after scaling of the same scale is sent to the multi-level feature attention module together with the encoder / decoder features, and the texture details are extracted and enhanced through the three multi-level feature attention modules in turn; finally, the features processed by the multi-level feature attention module are fused through the residual dense block, and the fused features are added to the original input image residual to generate the enhanced illumination component. The double-branch denoising module takes the M-LBP feature map and the final reflection component As input, wherein the upper branch extracts forward features through a multi-element feature perception module, and the lower branch takes the M-LBP feature map after negation and excavates low-contrast information through a multi-modal attention fusion module; after superimposing the extracted forward features and low-contrast information, sequentially passing through a noise integration module and a feature refinement module, removing the reflection component noise, obtaining the denoised reflection component.

2. The method of claim 1, wherein the method is based on M-LBP guided multi-attention low-light image enhancement. The decomposition network DecomNet in the cycle decomposition module is composed of a U-Net and a residual block, and the decomposition network DecomNet adopts a long-short skip hybrid connection mode; in the down-sampling process of the input image, Conv 4x4, stride=2 is used instead of the maximum pooling layer, and a residual block is applied to each layer of the encoder-decoder, wherein Conv represents convolution.

3. The method of claim 1, wherein the method is based on M-LBP guided multi-attention low-light image enhancement. The multi-element feature perception module extracts multi-scale features in parallel through one-time expansion convolution, two-time expansion convolution and three-time expansion convolution, and then performs channel attention and residual processing; the multi-modal attention fusion module includes local residual, channel attention and pixel attention connected in turn, generates channel weights through global average pooling, and generates pixel weights through convolution layers; the noise integration module is composed of a Tanh hyperbolic function, a ReLU and a convolution connected in turn, and the feature refinement module is composed of multiple groups of ReLU, BN and convolution connected in turn; the noise integration module deeply fuses the noise-related features extracted by the upper branch and the lower branch to generate noise feature representation.

4. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the multi-attention low-illumination image enhancement method of any one of claims 1 to 3.

5. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the computer program is executed, the steps of the multi-attention low-illumination image enhancement method of any one of claims 1 to 3 are implemented.

6. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the multi-attention low-illumination image enhancement method of any one of claims 1 to 3.

Images