Low-light image enhancement network model based on attention mechanism

By introducing multi-scale self-attention and local global attention modules into the low-light image enhancement network, the problems of contrast drop and color distortion of the low-light image enhancement method in the existing technology are solved, and better details recovery and brightness improvement are achieved. , significantly improve the visual effect of the image.

CN118154446BInactive Publication Date: 2024-11-05SICHUAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410383515.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-01
Publication Date
2024-11-05
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing low-light image enhancement methods have problems such as decreasing contrast, color distortion, halo causing blurred areas or unclear edges, making it difficult to restore details in dark areas, and the existing deep learning methods have increased complexity but are not effective.

Method used

Using a low-light image enhancement network model based on attention mechanism, the image decomposition module is used to decompose the image into reflection maps and lighting maps. Multi-scale self-attention, local and global attention modules are added to the reflection recovery module to enhance the recovery of dark areas. and output low light enhancement images through feature stitching and convolutional layer processing.

Benefits of technology

Effectively integrate multi-scale information, reduce halo and color distortion, improve the detail recovery ability of low-light images, significantly improve brightness and fidelity, and make the image more natural.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118154446B_ABST
    Figure CN118154446B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of image processing technology, and discloses a low-light image enhancement network model based on an attention mechanism, the model includes an image decomposition module, which obtains an original image and decomposes the original image into a reflection map and an illumination map; a reflection recovery module, which performs restoration processing on the reflection map and the illumination map to obtain a restored feature map; an illumination adjustment module, which performs illumination adjustment on the illumination map to obtain an illumination feature map; and an image output module, which outputs a low-light enhanced image after merging the restored feature map and the illumination feature map. The present invention adopts a multi-scale adaptive attention mechanism to complete the restoration of color and texture, and a local and global attention association mechanism to enhance the detail restoration of dark areas in low-light images. The global attention captures the global information of the input feature map, which helps the model better understand the overall scene; the local attention is used to highlight the dark details in the input feature to make it more clearly visible.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a low-light image enhancement network model based on an attention mechanism. Background Art

[0002] Due to the limitations of the shooting environment, images taken under low-light conditions will have some problems of low visibility. Image processing methods can solve this problem, such as the histogram equalization method, which is simple to calculate and easy to implement, but has problems such as over-enhancement and local contrast distortion; filtering-based low-light enhancement methods, which are suitable for some low-light images with more noise, but may cause image details to be lost, and are not obvious enough for some weak target signals; deep learning-based methods; which have shown superior performance in low-level visual tasks such as denoising and super-resolution, but most tasks require reference training examples; with the development of data-driven methods and the improvement of performance, the complexity of low-light enhancement networks based on deep learning is also increasing, but existing deep learning-based methods still have the following problems: such as contrast reduction, color distortion, a large number of halos, which lead to blurred areas or unclear edges in the image, and it is difficult to restore details in the dark or missing details.

[0003] Ideal low-light image enhancement requires not only changing the brightness distribution of the image to increase the brightness of the underexposed areas in the image, but also restoring the detail information in these areas as much as possible to improve the visual effect of the image. Obviously, the existing methods do not achieve the expected results in the process of image enhancement and detail restoration. Summary of the invention

[0004] The purpose of the present invention is to provide a low-light image enhancement network model based on an attention mechanism, which adds a new attention module to the reflection recovery module to enhance the ability to restore details in dark areas in low-light images.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A low-light image enhancement network model based on attention mechanism, including:

[0007] An image decomposition module, which obtains an original image and decomposes the original image into a reflection map and an illumination map;

[0008] A reflection recovery module, which recovers the reflection map and the illumination map into a restored feature map; the reflection recovery module includes a pre-processing module, a processing module and an output module, the pre-processing module includes two image processing units consisting of a convolution + activation function layer and an attention module, and a splicing module, the reflection map and the illumination map are respectively input to the splicing module after passing through the convolution + activation function layer and the attention module, the splicing module splices the reflection feature map and the illumination feature map and then inputs them to the processing module, the processing module includes four sequentially connected processing sub-units, each processing sub-unit includes two convolution + ReLU layers and an attention module, and the output module includes a convolution + ReLU layer, a convolution layer and a Sigmoid layer;

[0009] A lighting adjustment module performs lighting adjustment on the lighting diagram to obtain a lighting characteristic diagram;

[0010] An image output module combines the restored feature map and the illumination feature map to output a low-light enhanced image;

[0011] Among them, the attention module includes a multi-scale self-attention module, a local attention module and a global attention module. The feature maps output by the convolutional layer are respectively input into the multi-scale self-attention module and the local attention module. The feature maps output by the multi-scale self-attention module are used as the input feature maps of the global attention module. The feature maps output by the global attention module and the feature maps output by the local attention module are combined and output.

[0012] Furthermore, the image decomposition module includes a reflection map decomposition operation and an illumination map decomposition operation, and the reflection map decomposition operation includes a first downsampling module, a second downsampling module, a first convolutional layer, a first upsampling module, a second upsampling module and a second convolutional layer connected in sequence. The first downsampling module is connected to the second upsampling module, and the second downsampling module is connected to the first upsampling module. The downsampling module is composed of a 3×3 convolution + ReLU layer and a 2×2 maximum pooling layer, and the upsampling module is composed of a deconvolution layer, feature splicing, and a 3×3 convolution + ReLU layer.

[0013] The illumination graph decomposition operation includes a first downsampling, two convolution+ReLU layers, a feature splicing layer, and a convolution+Sigmoid layer connected in sequence, and the input end of the feature splicing layer is also connected to the output end of the second upsampling module.

[0014] Furthermore, the multi-scale adaptive attention module includes a channel self-attention module and a spatial self-attention module. The channel self-attention module first performs a multi-branch separation convolution layer on the original input feature map. The input feature map is processed by three parallel convolution layers to obtain three feature maps with different receptive fields. The three feature maps are combined in pairs and the global average value of the entire feature map is obtained by global average pooling. The pooled feature maps are respectively passed through two fully connected layers to obtain three new feature maps; then the sigmoid activation function is used to obtain the channel attention weight; the weight is multiplied and merged with the input feature map, and normalized to obtain the channel attention feature map of each branch, and the channel attention feature maps of the three branches are fused to obtain the channel attention map;

[0015] A 1×1 convolution layer is used on the original input feature map to obtain the convolution layer feature map. The convolution layer feature map is added to the channel attention map as the input of the spatial self-attention module. The features after channel attention processing and the spatial attention features are combined to generate the spatial attention weight using a 1×1 convolution layer and a sigmoid activation function, which is then multiplied with the input feature map of the spatial self-attention module to obtain the spatial self-attention map.

[0016] Furthermore, the global attention module is used to capture the global information of the input feature map; first, a global average pooling operation is performed on the spatial self-attention map to obtain the global average of the entire feature map; two fully connected layers are used for feature conversion, and the first fully connected layer reduces the dimension of the features after global average pooling, reducing the feature dimension to 1 / 16 of the original feature dimension; the second fully connected layer maps the output of the first fully connected layer back to the original feature dimension to generate attention weights; finally, the output of the second fully connected layer is normalized by the softmax activation function to obtain the attention weights, and the attention weights are multiplied by the input feature map to obtain the weighted global attention map.

[0017] Furthermore, the local attention module is used to highlight the dark details in the input features; the original feature map first extracts the features of the input data through three convolutional layers, and then generates an attention map with the same spatial dimension but only one channel through a sigmoid activation function; the attention map is multiplied with the original feature map to obtain a feature map; the feature map and the original feature map are merged along the channel dimension to obtain a local attention feature map, the local attention feature map is processed by a convolutional layer, and then the output local attention map is generated through a ReLU activation function.

[0018] Furthermore, the lighting adjustment module includes three convolutional layers, which are connected in sequence.

[0019] The present invention has the following beneficial effects:

[0020] (1) Multi-scale information fusion for color and texture restoration: By selecting convolutional layer kernels of different scales, the feature information of multi-scale colors and textures can be effectively fused. This helps the network capture visual features at different levels, improves the feature representation ability, and is more conducive to learning the mapping relationship between low-light images and normal-light images, thereby reducing halo and color distortion.

[0021] (2) The adaptive selection module adapts to the different lighting characteristics of the image, enabling the network to adaptively select the scale of the convolutional layer kernel, thereby adapting to the different lighting characteristics of different images; at the same time, although multi-scale convolutional layer kernels are introduced, due to the existence of the selection module, only a part of the convolutional layer kernels will be selected for calculation, which reduces the computational cost to a certain extent and maintains the high efficiency of the network.

[0022] (3) The local and global attention association mechanism enhances the detail recovery of dark areas in low-light images. Global attention captures the global information of the input feature map, which helps the model better understand the overall scene. Local attention is used to highlight the dark details in the input features, making them more clearly visible. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a structural schematic diagram of the present invention.

[0024] Figure 2 It is a structural schematic diagram of the image decomposition module of the present invention.

[0025] Figure 3 It is a schematic structural diagram of the reflection recovery module of the present invention.

[0026] Figure 4 Schematic diagram of the structure of the attention module of the present invention.

[0027] Figure 5 Schematic diagram of the structure of the multi-scale self-attention module of the present invention.

[0028] Figure 6 For the structural comparison of the present invention and other methods Figure 1 .

[0029] Figure 7 For the structural comparison of the present invention and other methods Figure 2 . DETAILED DESCRIPTION

[0030] like Figure 1 As shown, a low-light image enhancement network model based on an attention mechanism provided in this embodiment includes an image decomposition module, a reflection recovery module, a lighting adjustment module and an image output module.

[0031] The image decomposition module is used to obtain the original image and decompose the original image into a reflection image and an illumination image. The image decomposition module of this embodiment is implemented by an encoding-decoding structure operation; specifically, it includes a reflection image decomposition operation and an illumination image decomposition operation. The reflection image decomposition operation is a typical five-layer U-net structure, such as Figure 2 As shown, it includes a first downsampling module, a second downsampling module, a first convolutional layer, a first upsampling module, a second upsampling module and a second convolutional layer connected in sequence. The first downsampling module is connected to the second upsampling module, and the second downsampling module is connected to the first upsampling module. The downsampling module consists of a 3×3 convolution + ReLU layer and a 2×2 maximum pooling layer. The upsampling module consists of a deconvolution layer, a feature splicing, and a 3×3 convolution + ReLU layer. The original image is downsampled twice and upsampled twice, and finally outputs a reflection map through a convolutional layer. The illumination map decomposition operation includes the first downsampling, two layers of convolution + ReLU layers, a feature splicing layer, and a convolution + Sigmoid layer connected in sequence. The input end of the feature splicing layer is also connected to the output end of the second upsampling module. The original image outputs an illumination map through upsampling, convolution, and splicing.

[0032] The reflection recovery module is used to perform a recovery operation on the reflection image and the illumination image and output a recovery feature image. This module is also a key improvement module of this embodiment. Figure 3 As shown, the reflection recovery module includes a first convolution + ReLU layer and a first attention module connected in sequence, and a first convolution + sigmoid layer and a second attention module connected in sequence; the first attention module and the second attention module are both connected to the splicing module, and the splicing module is sequentially connected with the second convolution + ReLU layer, the third convolution + ReLU layer, the third attention module, the fourth convolution + ReLU layer, the fifth convolution + ReLU layer, the fourth attention module, the sixth convolution + ReLU layer, the seventh convolution + ReLU layer, the fifth attention module, the eighth convolution + ReLU layer, the ninth convolution + ReLU layer, the sixth attention module, the tenth convolution + ReLU layer and the third convolution layer, the reflection map After passing through the first convolution + ReLU layer and the first attention module, it is input into the splicing module. After passing through the first convolution + sigmoid layer and the second attention module, the illumination map is input into the splicing module. The splicing module splices the emission feature map and the illumination feature map and passes through the second convolution + ReLU layer, the third convolution + ReLU layer, the third attention module, the fourth convolution + ReLU layer, the fifth convolution + ReLU layer, the fourth attention module, the sixth convolution + ReLU layer, the seventh convolution + ReLU layer, the fifth attention module, the eighth convolution + ReLU layer, the ninth convolution + ReLU layer, the sixth attention module, the tenth convolution + ReLU layer, the third convolution layer and the sigmoid layer in turn to output the restored feature map.

[0033] The structures of the first to sixth attention modules are the same. In this embodiment, the first, second, third, and fourth are added before the attention modules to distinguish them, which has no special meaning. The following describes the structure of the attention module as an example. Figure 4 As shown, the attention module HAIA includes a multi-scale self-attention module, a local attention module and a global attention module.

[0034] The multi-scale self-attention module includes a channel self-attention module and a spatial self-attention module. The channel self-attention module emphasizes the importance of different channels in the image and allocates more attention to more important feature channels by learning the weights of different channels. The spatial self-attention module is used to highlight the important spatial areas in the feature map. On the basis of the attention enhancement of a given channel, different weights are allocated to further emphasize the importance of specific spatial positions in the image. Figure 5 As shown in the figure, the channel self-attention block first performs a multi-branch separation convolution layer on the feature map, and each branch uses a different convolution layer kernel (with different receptive fields) for feature extraction. The input feature map is processed by three parallel convolution layers to obtain three feature maps with different receptive fields. The three feature maps are combined in pairs to obtain three new feature maps. The new feature maps are respectively passed through the global average pooling (GAP) to obtain the global average feature map of the entire feature map. After entering the fully connected layer, they are passed through the BN layer and the ReLU layer to generate a new feature map. The three feature maps are passed through the fully connected layer again to obtain three new feature maps. Then the sigmoid activation function is used to obtain the channel attention map. Figure 5 In the figure, a and b represent the channel attention weight map of one of the branches respectively. The attention maps of these two channels help to adaptively extract more representative feature maps from receptive fields of different scales. The operations of the other two channels are the same as this branch. In order to realize automatic feature selection, this embodiment uses a calibration feature map. Respectively represent the calibrated feature maps; the calibrated feature maps are added to obtain one of the three branches, and the channel dimension is adjusted to obtain the feature map, and finally the three branch results are fused to obtain the output result of the channel self-attention module. The channel attention mechanism focuses on the category of features, and the spatial attention mechanism focuses on the location of features. In order to refine the location information of the target, a 1×1 convolution layer is used for the original feature map to obtain a convolution layer feature map, and the convolution layer feature map is added to the channel attention feature map as the input of the spatial self-attention block. The features after channel attention processing and the spatial attention features are combined to generate spatial attention weights using a 1×1 convolution layer and a sigmoid activation function. These weights are multiplied with the features enhanced by channel attention to highlight the features of a specific spatial area, and the weight dimension is adjusted using spatial tensor replication to match the original features, and then element-wise multiplication is performed. Figure 5In the figure, c and d represent the spatial attention weight maps, which are used to weight the sum. Finally, a 1×1 convolutional layer and batch normalization are used to complete the feature fusion and obtain the spatial self-attention feature map.

[0035] The local attention module is used to highlight the dark details in the input features. Through a series of convolutional layers and attention mechanisms, the module can enhance the details of the dark areas in the input features to make them more clearly visible. The local attention mechanism helps the model focus on the dark areas; first, through the initial convolutional layer, the first to third layers use a combination of convolutional layers and batch normalization plus ReLU activation functions to gradually extract the features of the input image; keep the spatial dimension of the feature map unchanged to obtain the feature map, and then generate an attention map with the same spatial dimension but only one channel through the sigmoid activation function; the attention map indicates the importance of the position of each pixel; the attention map is multiplied with the input feature map to obtain the feature map. Emphasize the areas that are considered important in the original input, that is, the dark areas, because the places where the value of the attention map is close to 1 will retain the original features, while the places close to 0 will be suppressed; then feature merging is performed, the original input feature map and the attention-weighted data are merged along the channel dimension, while retaining the original features and the emphasized features to obtain the feature map, and finally the merged feature map is processed through a convolutional layer, and finally the output local attention map is generated through the ReLU activation function.

[0036] The global attention module is used to capture the global information of the input feature map; by weighting the attention of the entire feature map, the global attention mechanism helps the model better understand the overall scene and guides the model to focus on important feature areas. First, the input features are subjected to a global average pooling GAP operation to obtain the global average of the entire feature map; then two fully connected layers are used for feature conversion; the first fully connected layer (FC1) reduces the dimension of the features after global average pooling and reduces the feature dimension to 1 / 16 of the original feature dimension; the second fully connected layer (FC2) maps the output of the first fully connected layer back to the original feature dimension to generate attention weights; the attention weight n is obtained through the softmax activation function; finally, the attention weight is multiplied by the input feature map to obtain the weighted feature map, i.e., the global attention map.

[0037] The input feature map of the global attention module of this embodiment is the output feature map of the multi-scale self-attention module, that is, the spatial self-attention feature map. The feature map output by the global attention module is combined with the feature map output by the local attention module to output the feature map.

[0038] The lighting adjustment module is used to perform lighting adjustment on the lighting map to obtain a lighting feature map, and is mainly composed of a feature splicing layer, four convolution + ReLU layer operations, and a sigmoid activation layer operation. The feature splicing layer splices the lighting feature map and the input tensor Ratio, and then sequentially inputs the four convolution + ReLU layers and the sigmoid activation layer to output the lighting feature map.

[0039] The image output module is used to merge the restoration feature map and the illumination feature map and then output the low-light enhanced image.

[0040] To verify the effectiveness of this embodiment, this embodiment is evaluated qualitatively and quantitatively.

[0041] Taking the LOL dataset as an example, the visual comparison between this embodiment and KinD, KinD++, Night, Retinex-Net, URetinexNet, and SCI methods is shown. Figure 6 and Figure 7 As shown in the figure, the enhancement results of Retinex-Net have serious noise and color distortion, and the effect is average. Although KinD and KinD++ have suppressed noise and color distortion to a certain extent, they are prone to halo, blurred details in dark places, and insufficient brightness enhancement. Night and SCI also have problems such as blurred details in dark places. Although URetinexNet has improved brightness enhancement, it has problems such as excessive brightness enhancement leading to image color distortion.

[0042] In order to effectively evaluate the proposed method, this embodiment uses the two most commonly used indicators in this field as the evaluation results, namely structural similarity (SSIM) and peak signal-to-noise ratio (PSNR). Among them, SSIM is used to evaluate the similarity between the enhanced image and the label. The closer the SSIM value is to 1, the higher the similarity and the better the restoration result. PSNR is used to measure the fidelity between the enhanced image and the label. The larger its value, the better the fidelity of the restored image. Table 1 shows the indicator comparison of the present invention with other methods on the LOL dataset. The experimental results show that this embodiment exhibits better performance and is superior to other advanced methods in both PSNR and SSIM;

[0043] Table 1 Quantitative measurement results on the LOL dataset

[0044] name Retinex-Net KinD KinD++ Night SCI URetinexNet This embodiment PSNR 18.06 20.87 21.3 21.32 14.78 19.84 22.82 SSIM 0.74 0.8 0.82 0.85 0.71 0.86 0.87

[0045] In general, this embodiment not only has a good improvement in brightness, but also shows better effects in noise and degradation processing, especially in detail recovery in dark areas. In addition, the enhanced image appears very natural and the fidelity is also guaranteed.

[0046] The above description is only a preferred implementation manner of the present invention, but the protection scope of the present invention is not limited thereto, and any modification and replacement based on the technical solution and inventive concept provided by the present invention should be included in the protection scope of the present invention.

Claims

1. A low-light image enhancement network model device based on attention mechanism, characterized in that: include: An image decomposition module, which obtains an original image and decomposes the original image into a reflection map and an illumination map; A reflection recovery module restores the reflection map and the illumination map into a restored feature map; it includes a first convolution + ReLU layer and a first attention module connected in sequence, and a first convolution + sigmoid layer and a second attention module connected in sequence; the first attention module and the second attention module are both connected to the splicing module, and the splicing module is followed by a second convolution + ReLU layer, a third convolution + ReLU layer, a third attention module, a fourth convolution + ReLU layer, a fifth convolution + ReLU layer, a fourth attention module, a sixth convolution + ReLU layer, a seventh convolution + ReLU layer, a fifth attention module, an eighth convolution + ReLU layer, a ninth convolution + ReLU layer, a sixth attention module, a tenth convolution + ReLU layer, a third convolution layer and sigmoid layer, the reflection image passes through the first convolution + ReLU layer and the first attention module and is input into the splicing module. The illumination image passes through the first convolution + sigmoid layer and the second attention module and is input into the splicing module. The splicing module splices the reflection image and the illumination image and passes through the second convolution + ReLU layer, the third convolution + ReLU layer, the third attention module, the fourth convolution + ReLU layer, the fifth convolution + ReLU layer, the fourth attention module, the sixth convolution + ReLU layer, the seventh convolution + ReLU layer, the fifth attention module, the eighth convolution + ReLU layer, the ninth convolution + ReLU layer, the sixth attention module, the tenth convolution + ReLU layer, the third convolution layer and the sigmoid layer in sequence to output the restored feature map; A lighting adjustment module performs lighting adjustment on the lighting diagram to obtain a lighting characteristic diagram; An image output module combines the restored feature map and the illumination feature map to output a low-light enhanced image; Among them, the structures of the first attention module to the sixth attention module are the same, and the attention modules include a multi-scale self-attention module, a local attention module and a global attention module. The multi-scale self-attention module includes a channel self-attention module and a spatial self-attention module. The channel self-attention block first performs a multi-branch separation convolution layer on the feature map. The input feature map is processed by three parallel convolution layers to obtain three feature maps with different receptive fields. The three feature maps are combined in pairs to obtain three new feature maps. The new feature maps are respectively passed through the global average pool to obtain the global average feature map of the entire feature map. The global average feature map is input into the fully connected layer and then passed through the BN layer and the ReLU layer to generate a new feature map; the three new feature maps are again passed through the fully connected layer and the sigmoid activation function to obtain the channel attention map, and finally the three-branch results are fused to obtain the output result of the channel self-attention module. ; Apply a 1×1 convolution layer to the original feature map to obtain the convolution layer feature map , the convolutional layer feature map With channel attention feature map The spatial attention weights are generated by combining the features after channel attention processing and the spatial attention features using a 1×1 convolutional layer and a sigmoid activation function. The spatial attention weights are multiplied with the features enhanced by channel attention to highlight the features of a specific spatial region. The spatial tensor copy is used to adjust the weight dimension to match the original features, and then the spatial attention weight map is obtained by element-wise multiplication. The spatial attention weight map is used to and After weighting, the features are fused through a 1×1 convolution layer and batch normalization to obtain a spatial self-attention feature map. The local attention module is used to highlight the dark details in the input features. First, the features of the input image are gradually extracted through three convolutional layers, keeping the spatial dimension of the feature map unchanged, and obtaining the feature map. , and then through the sigmoid activation function, generate an attention map with the same spatial dimensions but only one channel ; Attention map Multiply the input feature map to get the feature map ; Then the original input feature map and the attention-weighted data are merged along the channel dimension, retaining the original features and the emphasized features to obtain the feature map , and finally the merged feature map is processed by a convolution layer, and finally the output local attention map is generated by the ReLU activation function; The global attention module is used to capture the global information of the input feature map; firstly, the input feature is subjected to a global average pooling GAP operation to obtain the global average value of the entire feature map; then two fully connected layers are used to perform feature conversion; the first fully connected layer performs dimensionality reduction on the features after global average pooling, reducing the feature dimension to 1 / 16 of the original feature dimension; the second fully connected layer maps the output of the first fully connected layer back to the original feature dimension, and obtains the attention weight n through the softmax activation function; finally, the attention weight is multiplied by the input feature map to obtain a weighted global attention map; The feature map input to the global attention module is the spatial self-attention feature map output by the multi-scale self-attention module; the feature map output by the global attention module is combined with the feature map output by the local attention module to output the feature map.

2. The low-light image enhancement network model device based on the attention mechanism according to claim 1 is characterized in that: The image decomposition module includes a reflectance map decomposition operation and an illumination map decomposition operation; The reflection map decomposition operation includes a first downsampling module, a second downsampling module, a first convolution layer, a first upsampling module, a second upsampling module and a second convolution layer connected in sequence, the first downsampling module is connected to the second upsampling module, the second downsampling module is connected to the first upsampling module, the downsampling module is composed of a 3×3 convolution + ReLU layer and a 2×2 maximum pooling layer, the upsampling module is composed of a deconvolution layer, feature splicing, and a 3×3 convolution + ReLU layer, the original image is downsampled twice and upsampled twice, and finally a convolution layer is used to output a reflection map; The illumination map decomposition operation includes a first downsampling module, two convolution+ReLU layers, a feature splicing layer, and a convolution+Sigmoid layer connected in sequence. The input end of the feature splicing layer is also connected to the output end of the second upsampling module. The original image is output as an illumination map through upsampling, convolution, and splicing.

3. The low-light image enhancement network model device based on the attention mechanism according to claim 1 is characterized in that: The lighting adjustment module consists of a feature splicing layer, four convolution + ReLU layer operations and a sigmoid activation layer operation. The feature splicing layer splices the lighting feature map and the input tensor Ratio, and then sequentially inputs the four convolution + ReLU layers and the sigmoid activation layer to output the lighting feature map.