A low-light image enhancement method based on channel-space composite attention

Abstract

This invention belongs to the field of digital image processing and discloses a low-light image enhancement method based on channel-space composite attention. The method comprises the following steps: constructing and training a low-light image enhancement network model; the constructed low-light image enhancement network comprises a decomposition network, a reflection restoration network, and an illumination adjustment network; the reflection restoration network employs a CBAM attention module to guide the denoising of the reflection image; the illumination adjustment network employs a GAM attention module to guide the brightness adjustment of the illumination image; and the low-light image is input into a trained convolutional neural network to obtain an enhanced result image. This method solves the problems of excessive noise in the reflection image and overexposure in the illumination image, and also achieves significant results in color restoration and global illumination consistency.

Description

Technical Field

[0001] The present invention belongs to the field of digital image processing, and in particular relates to a low-illumination image enhancement method based on channel space composite attention. Background Art

[0002] Due to environmental factors and limitations of camera equipment, especially at night, images captured can exhibit low brightness, low contrast, high noise, and color distortion, which impacts human perception. Furthermore, in advanced computer vision processing tasks, low-light images lack a significant amount of information, hindering subsequent image recognition, object detection, and tracking. Consequently, there is a need to enhance low-light images.

[0003] To effectively improve the quality of low-light images, numerous low-light image enhancement algorithms have been proposed. These algorithms have evolved through three stages: grayscale transformation, Retinex theory, and deep learning. Early grayscale transformation methods, such as histogram equalization and gamma correction, were used to stretch the intensity distribution in low-light areas to increase the brightness of dark areas. However, due to a lack of consideration of the relationships between pixels, grayscale transformation often results in enhanced images lacking realism. Low-light image enhancement methods based on Retinex theory often rely on manually designed constraints and parameters, often resulting in poor generalization capabilities.

[0004] Deep learning-based methods are the mainstream solution for low-light image enhancement, with Retinex-Net being a classic model. The Retinex-Net model uses normal-light images as a data reference and shares parameters with low-light images, eliminating the need for manual function design. This improves the model's adaptability across multiple scenarios and significantly enhances image brightness, contrast, color, and detail compared to non-deep learning methods. However, the reflectance map decomposed by the Retinex-Net model still exhibits significant noise after BM3D denoising, and the illumination map exhibits overexposure after passing through the illumination adjustment network. Zhang Yonghua et al. proposed a model that can freely adjust the brightness of the illumination map in the illumination restoration network (Zhang Y, Zhang J, Guo X. Kindling the darkness: A practical low-light image enhancer [C] / / Proce edings of the 27th ACM international conference on multimedia. 2019: 1632-1640.), namely KinD, which solves the problem of overexposure of the illumination map. However, due to the use of the U-Net structure in the reflection denoising network, the network has oversampling and over-pooling, and the final enhanced image has an unnatural halo. Zhang Yonghua et al. (Zhang Y H, Guo XJ, Ma JY, et al. Beyond Brightening Low-light Images [J]. International Journal of Computer Vision, 2021, 129 (4): 1013-1037.) proposed the KinD++ model to solve this problem, but the image still has problems such as color distortion. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a low-illumination image enhancement method based on channel-space composite attention, which can remove image noise, improve image brightness, and avoid color distortion.

[0006] In order to achieve the above object, the present invention is achieved through the following technical solutions:

[0007] The present invention is a low-light image enhancement method based on channel space composite attention, which comprises: inputting an image into a constructed low-light image enhancement network to train the network, inputting the low-light image to be enhanced into the trained model to obtain an enhanced result image; wherein the established low-light image enhancement network includes a decomposition network, an illumination adjustment network and a reflection restoration network,

[0008] The training process of the established low-light image enhancement network includes:

[0009] S1: low-light image S low and normal lighting image S normal Input the decomposition network to decompose the low-light image S low The corresponding output is the reflection map R low The illumination component L1 is filtered by the guide filter to obtain the illumination component L2, and the illumination component L1 and the illumination component L2 are added element by element to obtain the illumination map I low ;

[0010] S2: The reflection image R decomposed in step S1 low Denoising and enhancement are performed through the reflection recovery network to obtain the enhanced reflection map

[0011] S3: The light map I obtained by weighting in step S1 low Brightness enhancement is performed through the illumination adjustment network to obtain an enhanced illumination map

[0012] S4: The enhanced reflection image in step S2 and the enhanced illumination map in step S3 Multiply element by element to get the enhanced image

[0013] A further improvement of the present invention is that: in step S1, the decomposition network is composed of 6 convolutional layers and Sigmoid modules connected in series, wherein the 1st and 6th layers are ordinary convolutional layers for extracting features in the image; the 2nd to 5th layers are convolutional activation modules with activation functions for filtering areas in the image that are not related to the features and accurately distinguishing between high-frequency and low-frequency areas of the image.

[0014] The step S1 is specifically as follows:

[0015] The low-light image S low and normal lighting image S normal The decomposition networks are input separately for 6-layer convolution operations, and finally the convolution results are processed by Sigmoid function to obtain the reflection map R of the normal illumination image. normal and Light Map I normal , the reflectance map R of the low-light image low And illumination component L1; input illumination component L1 into Gaussian filter to obtain the denoised guidance map S g , the illumination component L1 is passed through the guidance map S g As the guided filter of the guided image, the illumination component L2 is obtained; the illumination component L1 and the illumination component L2 are added element by element to obtain the illumination image Ilow ; The low-light image S low and normal lighting image S normal Enter the shared network parameters after decomposing the network.

[0016] A further improvement of the present invention is that: in the step S2, the reflection restoration network is composed of 4 convolutional activation block attention modules, 1 convolutional activation module with LReLU activation function, a common convolution layer and a Sigmoid function connected in series, wherein the convolutional activation block attention module is 2 convolutional activation modules with LReLU activation function connected in series with 1 CBAM attention module, the convolutional activation module with LReLU activation function is used to filter the areas in the image that are not related to the features and pay attention to the noise information in the image, and the CBAM is used to guide the reflection map R low Denoising;

[0017] The step S2 is specifically as follows:

[0018] The reflection image R decomposed in step S1 low The input reflection recovery network performs 14 layers of convolution operations, and finally the convolution result is processed by the Sigmoid function to obtain the denoised and enhanced reflection map.

[0019] A further improvement of the present invention is that: in the step S3, the illumination adjustment network includes a downsampling stage and a corresponding upsampling stage, a common convolution layer and a Sigmoid function connected in series, wherein the downsampling stage is composed of two convolution activation modules with LReLU activation functions, one GAM attention module and one convolution activation module with LReLU activation function connected in series, and the upsampling stage is the opposite, consisting of one convolution activation module with LReLU activation function, one GAM and two convolution activation modules with LReLU activation function connected in series, the convolution activation module with activation function is used to filter the areas in the image that are not related to the features, pay attention to the brightness information in the image, and the GAM is used to guide the light map I low Brightness adjustment. In addition, the downsampling stage and the upsampling stage are connected by a GAM. The convolution layer realizes the jump connection between layers through channel splicing, guiding the downsampling layer to its corresponding upsampling layer;

[0020] A further improvement of the present invention is that the loss function of the established CNN network is:

[0021] L=ω c L c +ω d L d +ω r L r (1);

[0022] Among them L c , L d and L r They are color angle loss function, decomposition loss function and reconstruction loss function, ω c 、ω d and ω r is the corresponding weight coefficient.

[0023]

[0024] Among them S normal is a normal lighting image, is the enhanced low-light image, (S normal ) p and Indicates S normal and The pixel corresponding to the upper point p, For (S normal ) p and The color angle of , ||·||2 is the 2-norm;

[0025] L d =||R low -R normal ||1(3);

[0026] where R low is the reflection map of the low-light image, R normal This is the reflection map of the normal lighting image. According to Retinex theory, the reflection map is independent of the lighting map, so the reflection map should be kept consistent as much as possible.

[0027]

[0028]

[0029] in is the enhanced low-light image, (S normal ) c and is the value after image normalization.

[0030] The beneficial effect of the present invention is that: in guided filtering, a low-illumination image that has been denoised by Gaussian filtering is used as a guided image to decompose the low-illumination image, thereby suppressing edge noise in the illumination image.

[0031] In the decomposition network, the number of convolutional layers is reduced to reduce the amount of computation and prevent distortion of image details due to overfitting. The strategy of adding two illumination maps is used to obtain the illumination map, and the illumination map is subjected to average denoising.

[0032] In the reflection restoration network, a single-branch structure is used instead of the U-Net-like structure to reduce the noise in the reflection map and enhance the halo in the image.

[0033] The CBAM attention mechanism is adopted to focus on pixel weights while considering multi-channel correlation, which improves the network's sensitivity to image noise information and achieves denoising effect.

[0034] In the illumination adjustment network, the downsampling layer and the corresponding upsampling layer are jump-connected through channel splicing, so that the upsampling process can find the pixels lost during the downsampling process and obtain a high-resolution illumination map;

[0035] The GAM attention mechanism is adopted to focus on the weights of pixels in different channels, which reduces irrelevant feature responses, improves the network's sensitivity to brightness information and its ability to process detail information, and maintains the consistency of global illumination while adjusting the brightness due to its global interaction characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of the low-illumination image enhancement method of the present invention.

[0037] Figure 2 It is a schematic diagram of the decomposition network structure in the present invention.

[0038] Figure 3 It is a schematic diagram of the reflection recovery network structure in the present invention.

[0039] Figure 4 It is a schematic diagram of the lighting adjustment network structure in the present invention.

[0040] Figure 5 This is the first set of comparison pictures of the enhancement effects of the present invention and the comparison method on low-light LOL dataset images.

[0041] Figure 6 This is the second set of comparison pictures of the enhancement effects of the present invention and the comparison method on low-light LOL dataset images.

[0042] Figure 7 This is a comparison chart of the enhancement effects of the present invention and the comparison method on artificially taken low-light pictures of houses. DETAILED DESCRIPTION

[0043] The present invention is described clearly and completely below based on the drawings in the embodiment. Obviously, the embodiment described is only a part of the embodiment of the present invention, not all of the embodiment. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] The present invention proposes a low-light image enhancement method based on channel-space composite attention, which includes: inputting an image into a constructed convolutional neural network (CNN), i.e., a low-light image enhancement network, to train the network; inputting the low-light image to be enhanced into the trained low-light image enhancement model to obtain an enhanced result image; wherein the established low-light image enhancement network includes a decomposition network, an illumination adjustment network, and a reflection restoration network;

[0045] Among them, the established low-light image enhancement network includes a decomposition network, an illumination adjustment network and a reflection restoration network.

[0046] like Figure 1 As shown, the network designed in this embodiment is a low-light image restoration network based on channel-space composite attention, including a decomposition network, an illumination adjustment network and a reflection restoration network.

[0047] The training process of the established low-light image enhancement network includes:

[0048] S1: low-light image S low and normal lighting image S normal Input the decomposition network to decompose the low-light image S low The corresponding output is the reflection map R low The illumination component L1 is filtered by the guide filter to obtain the illumination component L2, and the illumination component L1 and the illumination component L2 are added element by element to obtain the illumination map I low ;

[0049] like Figure 2 As shown, the present invention simultaneously converts the low-light image S low and normal lighting image S normal Input the decomposition network separately to obtain the reflection map R of the normal illumination image normal and Light Map I normal , the reflectance map R of the low-light image low And the illumination component L1. Input the illumination component L1 into the Gaussian filter to obtain the denoised guidance map S g , the illumination component L1 is passed through the guidance map S g As the guided filter of the guided image, the illumination component L2 is obtained; the illumination component L1 and the illumination component L2 are added element by element to obtain the illumination image I low ; The low-light image S low and normal lighting image S normal After inputting the decomposition network, the network parameters are shared. The characteristic is that the LReLU activation function is used in the decomposition network instead of the commonly used ReLU activation function to prevent the gradient from disappearing when the ReLU activation function sets negative numbers to 0 during the learning process. low and Light Map Ilow The Sigmoid function is used to limit the image pixel range to [0, 1].

[0050] In step S1, the decomposition network is composed of 6 convolutional layers and Sigmoid modules connected in series. The first and sixth layers are ordinary convolutional layers with a convolution kernel of 3×3, which are used to extract features in the image. The second to fifth layers are convolution activation modules with a convolution kernel of 3×3 and an activation function, which are used to filter out areas in the image that are not related to the features and accurately distinguish between high-frequency and low-frequency areas of the image.

[0051] Step S1 is specifically: low-light image S low and normal lighting image S normal They are sent to the decomposition network respectively, and the feature map conv1 is obtained after passing through the 3×3 convolution layer; conv1 is extracted through 4 convolution activation modules with 3×3 convolution kernels and LReLU activation functions to obtain the processed feature map conv2; the processed feature map conv2 is passed through the 3×3 convolution layer to obtain the feature map conv_out; the feature map conv_out is processed by the Sigmoid function to obtain the reflection map R of the normal illumination image. normal and Light Map I normal , the reflectance map R of the low-light image low and illumination component L1; the illumination component L1 is input into Gaussian filter for denoising preprocessing, and the scale factor of Gaussian filter sigma = 80 to improve its image details; after Gaussian filter processing, a guide map S with clearer detail texture is obtained g However, there is still edge blur in the local area of the image. Next, the illumination component L1 is passed through the guidance map S g As the guided filter of the guided image, the scale factor of the guided filter is ε=16, the filter radius r=0.01, and the illumination component L2 is obtained; in order to obtain the reflection map of the low-illumination image with edge preservation, smooth details and low noise, the illumination component L1 and the illumination component L2 are added element by element to obtain the illumination map I low . The normal lighting image S normal As the true label is only input during training, and based on the principle of consistent Retinex reflectivity, the two share network parameters.

[0052] S2: The reflection image R decomposed in step S1 low Denoising enhancement via reflection restoration network.

[0053] like Figure 3 As shown, the reflectance map R of the low-light image obtained by the decomposition network is lowEnter the reflection restoration network. The reflection restoration network consists of four convolutional activation block attention modules, one convolutional activation module with a 3×3 convolution kernel and an LReLU activation function, a normal convolution layer with a 3×3 convolution kernel, and a Sigmoid function in series. The convolutional activation block attention module is composed of two convolutional activation modules with a 3×3 convolution kernel and an LReLU activation function in series with one convolutional block attention module (CBAM). The convolutional activation module with the LReLU activation function is used to filter out areas in the image that are not related to features and focus on the noise information in the image. The CBAM is used to guide the reflection map R low Denoising. CBAM considers both channel attention and spatial attention, calculating the corresponding attention weight map by computing the feature map. Compared to single-spatial attention or single-channel attention, CBAM considers multi-channel correlations while focusing on pixel weights, increasing the network's sensitivity to image noise and achieving denoising effects. Furthermore, to prevent halos in images, a single-branch structure is used in the reflection restoration network instead of a U-Net structure.

[0054] The step S2 is specifically as follows: the reflection image R of the low illumination image decomposed in S1 is low Input into the reflection recovery network, first extract two layers of features through two convolution activation modules with a convolution kernel of 3×3 and an LReLU activation function to obtain the feature map conv1_ref; refine the feature map conv1_ref by multiplying the attention with the feature map conv1_ref through the CBAM attention module, and the CBAM module generates a new attention weight and assigns it to the feature map conv1_ref to obtain the feature map conv2_ref to guide the reflection recovery network for denoising; pass the feature map conv2_ref through three convolution activation attention modules to obtain the feature map conv3_ref; pass the feature map conv3_ref through a convolution activation module with a convolution kernel of 3×3 and an LReLU activation function and a convolution layer with a convolution kernel of 3×3 to obtain the feature map conv_out_ref; finally, use the Sigmoid function to limit the pixel range of the feature map conv_out_ref to between [0,1] to obtain the enhanced reflection map Since the CBAM module is an end-to-end universal module, it can be flexibly embedded into the reflection restoration network.

[0055] S3: The light map I obtained by weighting in step S1 low Brightness enhancement via a lighting adjustment network.

[0056] like Figure 4 As shown, the illumination map I lowInput illumination adjustment network. The illumination adjustment network consists of a downsampling stage and a corresponding upsampling stage, a normal convolution layer and a Sigmoid function in series. The downsampling stage consists of two convolution activation modules with a convolution kernel of 3×3 and LReLU activation function, one global attention module (GAM) and one convolution activation module with a convolution kernel of 3×3 and LReLU activation function in series. The upsampling stage is the opposite, consisting of one convolution activation module with a convolution kernel of 3×3 and LReLU activation function, one GAM and two convolution activation modules with a convolution kernel of 3×3 and LReLU activation function in series. The convolution activation module with activation function is used to filter out the areas in the image that are not related to the features and focus on the brightness information in the image. GAM is used to guide the light map I low Brightness adjustment. In addition, the downsampling and upsampling stages are connected by a GAM. The convolutional layers implement skip connections between layers through channel splicing, guiding the downsampling layer to its corresponding upsampling layer. The GAM module focuses on the weights of pixels in different channels, reducing irrelevant feature responses and improving the network's sensitivity to brightness information and its ability to process detailed information. Due to its global interaction properties, the consistency of global illumination in the illumination image is taken into account when adjusting brightness. To obtain a high-resolution illumination map, channel splicing is used between the upsampling and downsampling layers of the network, allowing the upsampling process to find pixels lost during the downsampling process.

[0057] The step S3 is specifically as follows: the illumination map I of the low illumination image decomposed in S1 is converted into low Input into the illumination adjustment network, extract features through the downsampling layer, and obtain the feature map conv1_ill, where the GAM attention module multiplies the attention with the feature map to refine the feature map, and generates a new attention weight assigned to the feature map to guide the illumination adjustment network to enhance the illumination; the feature map conv1_ill is passed through a GAM to obtain the feature map conv2_ill; the feature map conv2_ill is passed through the upsampling layer to restore the spatial dimension to obtain the feature map conv3_ill. The upsampling layer is symmetrical with the downsampling layer, and the corresponding layers are spliced through channels to make up for the detail information lost during the downsampling process; the feature map conv3_ill is passed through a convolution layer with a convolution kernel of 3×3 to obtain conv_out_ill; finally, the Sigmoid function is used to limit the image pixel range of the feature map conv_out_ill to [0,1] to obtain the enhanced illumination map Since the GAM module is an end-to-end universal module, it can be flexibly embedded into the reflection recovery network.

[0058] S4: The enhanced reflection image in step S2 and the enhanced illumination map in step S3 Multiply element by element to get the enhanced image

[0059] The total loss function of the established low-light image enhancement network is:

[0060] L=ω c L c +ω d L d +ω r L r (1);

[0061] Among them L c , L d and L r They are color angle loss function, decomposition loss function and reconstruction loss function, ω c 、ω d and ω r is the corresponding weight coefficient, in this embodiment, ω c =2,ω d =1.5,ω r =1.

[0062]

[0063] Among them S normal is a normal lighting image, is the enhanced low-light image, (S normal ) p and Indicates S normal and The pixel corresponding to the upper point p, For (S normal ) p and The color angle of , ||·||2 is the 2-norm;

[0064] L d =||R low -R normal ||1(3);

[0065] where R low is the reflection map of the low-light image, R normal This is the reflection map of the normal lighting image. According to Retinex theory, the reflection map is independent of the lighting map, so the reflection map should be kept consistent as much as possible.

[0066]

[0067]

[0068] in is the enhanced low-light image, (S normal ) c and is the value after image normalization.

[0069] The present invention will be evaluated below.

[0070] Experimental environment:

[0071] The following experiments were all conducted in the same training environment: Intel i7-9750H CPU, 32GB memory and two NVIDIA GeForce RTX 1080Ti GPUs, the operating system is Ubuntu 18.04; the programming language is Python 3.6. The present invention is implemented based on the Tensorflow framework, the batch size is set to 10, the epochs is 1500, and the initial value of the learning rate is 0.001. 485 image pairs in the low-light dataset (LOL) are used as training sets, and 30 image pairs are used as test sets. In addition, 20 normal-light images were taken with a mobile phone, and the brightness was adjusted to -50, the exposure to -50, and the contrast to 10 using the software that comes with the mobile phone to obtain the corresponding low-light images, and 20 image pairs were obtained as test sets. The mobile phone model is Redmi Note 115G, and the camera pixels are rear 50MP+8MP.

[0072] Evaluation indicators:

[0073] The present invention is compared with the Retinex-Net and KinD++ methods, and the evaluation indicators used are peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and natural image quality evaluation (NIQE). The larger the PSNR and SSIM values, the better the image enhancement effect, and the smaller the NIQE value, the better the image enhancement effect. The objective indicator comparison results of the Retinex-Net and KinD++ methods and the present invention on the low-illumination image dataset are shown in Tables 1 and 2. The results take the average value of the objective evaluation indicator values obtained from the 50 image tests.

[0074] Table 1 Comparison of objective indicators of Retinex-Net, KinD++ and the present invention on the LOL dataset

[0075]

[0076] Table 2 Comparison results of objective indicators of Retinex-Net, KinD++ and the present invention on images taken by mobile phones

[0077]

[0078] As can be seen from Tables 1 and 2, the present invention has the highest SSIM index, indicating that the enhancement result of the present invention is most similar to the original image; the present invention has the highest PSNR value, indicating that the present invention can reduce the noise in the image; the present invention has the smallest NIQE value, indicating that the enhancement result of the present invention is more consistent with the visual effect of the human eye.

[0079] Figure 5 and Figure 6 These are two sets of comparison pictures of the enhancement effect of the present invention and the comparison method on the low-light LOL dataset. Figure 5 and Figure 6 It can be seen that the enhancement results obtained by the Retinex-Net method are distorted in color. Figure 5 Bookcase wood material and Figure 6 The grass and plants are too bright in color, which does not conform to the reality and the visual effect of the human eye. Although the enhancement results obtained by the KinD++ method have significantly improved the color and brightness restoration, there are still artifacts and noise in some parts. Figure 5 The wooden bookshelf and the lower right corner are unnaturally overexposed. This is caused by not taking global illumination consistency into account. Figure 6 The brightness of the windowsill is not restored enough. By comparison, it can be clearly seen that compared with the Retinex-Net method and the KinD++ method, the present invention obtains an enhanced image with less noise, more uniform brightness, more natural colors, and details that are more in line with the visual effects of the human eye. Figure 7 This is a comparison chart of the enhancement effect of the present invention and the contrast method on artificially taken low-light pictures of houses. Figure 7 It can be seen that the leaves and vines on the house in the enhanced image obtained by the Retinex-Net method have color distortion; the roof of the house in the enhanced image obtained by the KinD++ method has a halo, and the image as a whole is overexposed; although the enhanced image obtained by the present invention is not as good as the KinD++ method in improving brightness, it is better than the KinD++ method in removing noise, artifacts and halo.

[0080] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A low-light image enhancement method based on channel-space composite attention, characterized by: The method comprises inputting an image into a constructed low-light image enhancement network to train the network, and inputting the low-light image to be enhanced into the trained low-light image enhancement network to obtain an enhanced result image. The low-light image enhancement network includes a decomposition network, an illumination adjustment network, and a reflection restoration network. Specifically, the low-light image enhancement method comprises the following steps: Step 1: Transform the low-light image S low and normal lighting image S normal Input the decomposition network to decompose respectively, the low illumination image S low The corresponding output is the reflection map R low The illumination component L1 is filtered by the guide filter to obtain the illumination component L2, and the illumination component L1 and the illumination component L2 are added element by element to obtain the illumination map I low ; Step 2: Decompose the reflection map R obtained in step 1 low The enhanced reflection map is obtained by denoising and enhancing the reflection restoration network. Step 3: The light map I obtained by weighting in step 1 low Brightness enhancement is performed through the illumination adjustment network to obtain an enhanced illumination map. Step 4: Enhance the reflection image in step 2 and the enhanced lighting map in step 3 Multiply element by element to get the enhanced image in: In step 1, the decomposition network is composed of 6 convolutional layers and Sigmoid modules connected in series, wherein the 1st and 6th layers are ordinary convolutional layers for extracting features in the image, and the 2nd to 5th layers are convolutional activation modules with activation functions for filtering areas in the image that are not related to the features and distinguishing high-frequency areas from low-frequency areas of the image; The reflection restoration network is composed of four convolutional activation block attention modules, one convolutional activation module with LReLU activation function, a normal convolution layer and a Sigmoid function in series, wherein the convolutional activation block attention module is two convolutional activation modules with LReLU activation function in series with one convolutional block attention module (CBAM). The convolutional activation module with LReLU activation function is used to filter the areas in the image that are not related to the features and pay attention to the noise information in the image. The convolutional block attention module (CBAM) is used to guide the reflection map R low Denoising; The illumination adjustment network includes a downsampling stage and a corresponding upsampling stage, a common convolution layer and a Sigmoid function in series, wherein the downsampling stage is composed of two convolution activation modules with LReLU activation function, one global attention module (GAM) and one convolution activation module with LReLU activation function in series, and the upsampling stage is composed of one convolution activation module with LReLU activation function, one global attention module (GAM) and two convolution activation modules with LReLU activation function in series. The convolution activation module with activation function is used to filter the areas in the image that are not related to the features and focus on the brightness information in the image. The global attention module (GAM) is used to guide the illumination map I low Brightness adjustment, the downsampling stage and the upsampling stage are connected by a global attention module (GAM). The convolution layer realizes jump connection between layers through channel splicing, guiding the downsampling layer to its corresponding upsampling layer.

2. The low-light image enhancement method based on channel-space composite attention according to claim 1, characterized in that: The step 1 is specifically as follows: low and normal lighting image S normal The decomposition networks are input separately for 6-layer convolution operations, and finally the convolution results are processed by Sigmoid function to obtain the reflection map R of the normal illumination image. normal and Light Map I normal , the reflectance map R of the low-light image low And the illumination component L1, the illumination component L1 is input into the Gaussian filter to obtain the denoised guidance map S g , the illumination component L1 is passed through the guidance map S g As the guided filter of the guided image, the illumination component L2 is obtained, and the illumination component L1 and the illumination component L2 are added element by element to obtain the illumination image I low , where the low-light image S low and normal lighting image S normal Enter the shared network parameters after decomposing the network.

3. The low-light image enhancement method based on channel-space composite attention according to claim 1, characterized in that: The step 2 is specifically as follows: the reflection image R decomposed in step 1 is low Input reflection recovery network for 14 layers of convolution Finally, the convolution result is processed by Sigmoid function to obtain the denoised and enhanced reflection map.

4. The low-light image enhancement method based on channel-space composite attention according to any one of claims 1 to 3, characterized in that: The loss function of the low-light image enhancement network is: L=ω c L c +oh d L d +oh r L r ; Among them L c , L d and L r They are color angle loss function, decomposition loss function and reconstruction loss function, ω c 、ω d and ω r is the corresponding weight coefficient, Among them S normal is a normal lighting image, is the enhanced low-light image, (S normal ) p and Indicates S normal and The pixel corresponding to the upper point p, For (S normal ) p and The color angle of , ||·||2 is the 2-norm; L d =||R low -R normal ||1; where R low is the reflection map of the low-light image, R normal This is the reflection map of the normal illumination image. According to Retinex theory, the reflection map is independent of the illumination map, so the reflection map remains consistent. in is the enhanced low-light image, (S normal ) c and is the value after image normalization.

Images