Image enhancement model, method, device and storage medium with multi-branch fusion attention mechanism

Abstract

The present invention discloses an image enhancement model, method, device and storage medium of a multi-branch fusion attention mechanism. The image enhancement model consists of a feature extraction stage, an enhancement stage and a fusion stage. The feature extraction stage is responsible for identifying different areas of low-light images and extracting effective information such as edges and textures; the enhancement stage adaptively sets different weights for the extracted features by combining the attention mechanism to enhance the image and achieve denoising; the fusion stage further fuses and enhances the output results using a multi-branch fusion network, and finally outputs an enhanced image. The present invention can adaptively enhance image brightness while reducing image noise and removing artifacts, and can solve the problem of overexposure caused by low contrast, loss of detail information, color distortion and noise artifacts in low-light images.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology and relates to a low-light image enhancement model, method, device and storage medium of a multi-branch fusion attention mechanism. Background Art

[0002] With the rapid development of computer vision, high-quality images not only record daily life but also provide convenience for scientific research. In daily life, due to the influence of weather, light intensity or shooting method, the obtained images have low contrast, serious noise, artifacts and information loss. This affects subsequent operations such as target recognition, image classification and detection. Therefore, low-light image enhancement technology came into being.

[0003] For low-light image enhancement, people have proposed various algorithms in recent years, which can be mainly divided into three categories: based on traditional algorithms, based on physical model prior algorithms, and based on deep learning algorithms.

[0004] Compared with traditional grayscale enhancement methods such as histogram equalization and gamma correction, it enhances dark areas by stretching the grayscale of bright areas. LIME [see "Guo X, Li Y, Ling H. LIME: Low-light image enhancement via illumination map estimation [J]. IEEE Transactions on image processing, 2016, 26 (2): 982-993."] optimizes the bright channel map of the low-light image to obtain the illumination image, and then obtains the enhanced image through gamma transformation; these methods can effectively improve the global and local contrast, but they are not sensitive to data processing, which may cause some bad visual effects in the local area.

[0005] For physical model prior algorithms, such as the Retinex model, which is an algorithm built by imitating the human visual system. MSR [see "Jobson DJ, Rahman Z, Woodell G AA multiscale retinex for bridging the gap between color images and the human observation of scenes [J]. IEEE Transactions on Image processing, 1997, 6 (7): 965-976."] and SRIE [see "Fu X, Zeng D, Huang Y, et al. A weighted variational model for simultaneous reflectance and illumination estimation [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 2782-2790."] can improve the quality of low-light images to a certain extent, but in solving the complex prior regularization terms constructed, multiple iterations are often required, which is time-consuming.

[0006] Recently, deep learning has become a popular research direction. For example, Retinex-Net [see "Wei C, Wang W, Yang W, et al. Deep retinex decomposition for low-light enhancement [J]. arXiv preprint arXiv: 1808.04560, 2018."] passes the input image through a decomposition module and a low-light enhancement module respectively, and then enhances it through constraints; Kind [see "Zhang Y, Zhang J, Guo X. Kindling the darkness: A practical low-light image enhancer [C] / / Proceedings of the 27th ACM international conference on multimedia. 2019: 1632-1640."] The network includes a decomposition module, a restoration module, and a reconstruction module. It is trained by using image blocks with different exposure levels, so that the enhanced image has a good effect; Zero-DCE [see "Li C, Guo C, Chen C L. Learning to Enhance Low-Light Image via Zero-Reference Deep Curve Estimation [J]. IEEE Transactions onSoftware Engineering,2021”] builds an unsupervised training network that does not rely on training data, and enhances low-light images by estimating the brightness curve of the input image in an iterative and progressive manner. GLADNet [see “Wang W, Wei C, Yang W, et al. Gladnet: Low-light enhancement network with globalawareness[C] / / 201813th IEEE international conference on automatic face&gesture recognition,2018:751-755.”] proposes a global perception and detail preservation network, which estimates the illumination image of the low-light image through global illumination estimation, and then enhances the low-light image through a detail reconstruction module.

[0007] Although the above low-light image enhancement methods have achieved good results, they still have the following problems:

[0008] (1) It is impossible to denoise low-light images, suppress artifacts, and enhance different regions at the same time;

[0009] (2) There is no denoising module and the impact of noise on the image is not considered.

[0010] Therefore, the present invention proposes a multi-branch low-light enhancement network (MANet), which first effectively extracts the features of the input image, then sets different weights for the extracted images in different regions in combination with the attention mechanism for enhancement, and finally fuses the enhanced outputs to further improve the image. Summary of the invention

[0011] In order to solve the above problems, the present invention provides an image enhancement model with a multi-branch fusion attention mechanism, which can improve the brightness of low-light images while reducing image noise and removing artifacts. It can solve the overexposure problems caused by low contrast, loss of detail information, color distortion and noise artifacts in low-light images.

[0012] The second object of the present invention is to provide an image enhancement method with a multi-branch fusion attention mechanism.

[0013] A third object of the present invention is to provide an electronic device.

[0014] A fourth object of the present invention is to provide a computer storage medium.

[0015] The technical solution adopted by the present invention is an image enhancement model with a multi-branch fusion attention mechanism, including: a feature extraction stage, an enhancement stage and a fusion stage;

[0016] The feature extraction stage is responsible for identifying different areas of the low-light image and extracting effective information such as edges and textures;

[0017] In the enhancement stage, the extracted features are adaptively set with different weights by combining the attention mechanism to enhance the image and achieve denoising;

[0018] The fusion stage uses a multi-branch fusion network to further fuse and enhance the output results, and finally outputs an enhanced image.

[0019] Furthermore, the feature extraction stage consists of three groups of depth-wise separable convolutional layers;

[0020] The structure of each group of depth-separable convolutional layers is Depthwise+BN+ReLu, the convolution kernel size is 3×3, and the number of channels is 32;

[0021] Each group of depth-separable convolutional layers is responsible for extracting the brightness of different areas and sending them to subsequent operations;

[0022] The depthwise separable convolutional layer helps to extract effective feature information such as edges and textures of low-light images while reducing the amount of computation.

[0023] Further, the enhancement stage is composed of three EM enhancement modules;

[0024] The EM enhancement module first uses a 3×3 convolutional layer to extract effective features of the input image, and then passes through the CBAM module to identify and enhance different regions;

[0025] The EM enhancement module designs the number of downsampling blocks according to the width and height of the feature map output by the downsampling, and uses three downsampling blocks in total. After a series of symmetrical upsampling blocks, an enhanced feature map is output.

[0026] The downsampling block consists of a convolutional layer with a step size of 2 and a ReLU;

[0027] The downsampling block replaces the maximum pooling layer with the CBAM attention module; first, the input features are channel pooled respectively, and then they are passed through the convolution network respectively and the output results are element-wise concatenated to obtain the weight coefficient Mc, and finally multiplied with the input features to obtain the output of the channel attention module; the output of the channel attention module is then used as the input of the spatial attention module, and then passed through a 3×3 convolution layer to obtain the weight coefficient Ms, and finally multiplied with the initial input feature to obtain the final result; all layers of this module use 3×3 convolution layers, and the number of channels are 16, 32, 64, and 128 respectively;

[0028] The EM enhancement module uses skip connections to introduce the features from the down-sampled image into its corresponding up-sampling module to compensate for the lost detail information and enhance the clarity.

[0029] In the enhancement stage, the outputs EM1o, EM2o, and EM3o of the three EM enhancement modules are respectively convolved by 1×1, and then fused to obtain the result of low-light image enhancement.

[0030] Furthermore, the fusion stage is responsible for fusing the output results of the enhancement stage;

[0031] The fusion stage adopts a four-scale fusion structure, and uses a multi-scale convolution structure with four parallel branches to further merge and enhance the output EMo of the enhancement module;

[0032] The enhanced image outputted by the fusion stage is connected in the color channel dimension to all outputs from the three modules EM1o, EM2o and EM3o;

[0033] The fusion stage can fully retain the information of the enhancement module and enhance it.

[0034] A multi-branch fusion attention mechanism image enhancement method, image enhancement is performed in the following steps:

[0035] Step S1, input the low-light image into the image enhancement model of the multi-branch fusion attention mechanism;

[0036] Step S2: an image enhancement model with a multi-branch fusion attention mechanism uses a feature extraction module to identify different areas of the low-light image and extract effective information such as edges and textures;

[0037] Step S3, an image enhancement model of a multi-branch fusion attention mechanism, using an enhancement module to set different weights to enhance the features such as edges and textures extracted in step S2, and complete denoising to obtain an enhanced feature map;

[0038] Step S4, the image enhancement model of the multi-branch fusion attention mechanism uses the fusion module to further merge and enhance the enhanced feature map output by step S3 to obtain the final enhanced image.

[0039] Furthermore, the image enhancement method must be implemented with the aid of an image enhancement model trained under the constraints of a loss function;

[0040] The loss function can reduce the distance between the enhanced image estimated by the image enhancement model and the real enhanced image corresponding to the input low-light image during the training process;

[0041] The loss function includes structural loss, perceptual loss and regional loss, and its calculation formula is:

[0042] Loss = αL SSIM +βL VGG +γL Region (1)

[0043] Where L SSIM , L VGG and L Region They represent structural loss, perceptual loss and regional loss respectively.

[0044] Furthermore, the structural loss function is used to measure the difference between the real image and the enhanced image to improve the visual quality of the output image. The calculation formula is:

[0045]

[0046] Among them, μ x and μ y are the average values ​​of image pixels x and y, respectively, x 2 , σ y 2 With σ xy denote the variance and covariance of x and y respectively, C1 and C2 are arbitrary non-zero constants whose values ​​are to prevent the denominator from being zero;

[0047] The structural loss only focuses on the low-level information of the image, but high-level information is also needed to enhance the image quality;

[0048] The perceptual loss uses an image style conversion algorithm to compare the features of the low-light image obtained by convolution with the features of the enhanced image. The calculation formula is:

[0049]

[0050] Among them, E and G represent the enhanced image and the real image respectively, W i,j H i,j C i,j Respectively represent the width, height and number of channels of the feature map, is the feature map obtained by the jth convolutional layer of the i-th block, i=3, j=4;

[0051] The regional loss is to constrain the degree of enhancement between the low-light area and other areas in the image, and avoid the overexposure problem caused by the overall enhancement of different areas of the whole image. The calculation formula is:

[0052]

[0053] Where E l and G l is the low-light area of ​​the enhanced image and the real image, E h and G h is the rest of the image;

[0054] The weight value is ω l =5,ω h =1. The basis for this value is: in order to pay more attention to the low-light area in the image during training, a larger weight value is set for the low-light area.

[0055] An electronic device, characterized in that the method described in any one of claims 5 to 7 is used to enhance low-light images.

[0056] A computer storage medium, characterized in that at least one program instruction is stored in the storage medium, and the at least one program instruction is loaded and executed by a processor to implement the low-light image enhancement method as described in any one of claims 5-7.

[0057] The beneficial effect of the embodiment of the present invention is that by replacing ordinary convolution with separable convolution, the amount of calculation is reduced while extracting effective features such as edges and textures. Different weights can be set according to different characteristic information to achieve local enhancement of feature information in different regions, which can reduce the loss of semantic features and improve image clarity. The present invention can adaptively increase image brightness while reducing image noise and removing artifacts, and can solve the problem of overexposure caused by low contrast, loss of detail information, color distortion, and noise artifacts in low-light images. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0059] Figure 1 It is a structural diagram of an image enhancement model of a multi-branch fusion attention mechanism in an embodiment of the present invention.

[0060] Figure 2 It is a schematic diagram of the enhancement module structure of the enhancement stage of an image enhancement model of a multi-branch fusion attention mechanism in an embodiment of the present invention.

[0061] Figure 3 It is a structural schematic diagram of the fusion stage of an image enhancement model of a multi-branch fusion attention mechanism in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] Embodiment 1,

[0064] A multi-branch fusion attention mechanism image enhancement model, whose structure is as follows Figure 1 As shown, it includes: feature extraction stage, enhancement stage and fusion stage;

[0065] The feature extraction stage is responsible for identifying different areas of low-light images and extracting effective information such as edges and textures; the feature extraction stage consists of three groups of depthwise separable convolutional layers; the structure of each group of depthwise separable convolutional layers is Depthwise+BN+ReLu, the convolution kernel size is 3×3, and the number of channels is 32; each group of depthwise separable convolutional layers is responsible for extracting the brightness of different areas and sending it to subsequent operations; while reducing the amount of computation, the depthwise separable convolutional layers help to extract effective feature information such as edges and textures of low-light images.

[0066] like Figure 2 As shown,

[0067] In the enhancement stage, the extracted features are adaptively weighted by combining the attention mechanism to enhance the image and achieve denoising.

[0068] The enhancement stage consists of three EM enhancement modules. The EM enhancement module first uses a 3×3 convolutional layer to extract effective features of the input image, and then passes through the CBAM module to identify and enhance different regions. The EM enhancement module designs the number of downsampling blocks according to the width and height of the feature map output by the downsampling, and shares three downsampling blocks. After a series of symmetrical upsampling blocks, the enhanced feature map is output.

[0069] The downsampling block consists of a convolutional layer with a stride of 2 and a ReLU. The downsampling block replaces the maximum pooling layer with a CBAM attention module. First, the input features are channel pooled separately, and then they are passed through the convolutional network and the output results are element-wise concatenated to obtain the weight coefficient Mc, which is finally multiplied with the input features to obtain the output of the channel attention module. The output of the channel attention module is then used as the input of the spatial attention module, and then passed through a 3×3 convolutional layer to obtain the weight coefficient Ms, which is finally multiplied with the initial input feature to obtain the final result. All layers of this module use 3×3 convolutional layers, and the number of channels are 16, 32, 64, and 128 respectively.

[0070] The EM enhancement module uses skip connections to introduce the features from the down-sampled image into its corresponding up-sampling module to compensate for the lost detail information and enhance the clarity.

[0071] In the enhancement stage, the outputs EM1o, EM2o, and EM3o of the three EM enhancement modules are respectively subjected to 1×1 convolution and then fused to obtain the result of low-light image enhancement.

[0072] like Figure 3 As shown,

[0073] In the fusion stage, a multi-branch fusion network is used to further fuse and enhance the output results, and finally an enhanced image is output.

[0074] In the fusion stage, a four-scale fusion structure is used, and the output EMo of the enhancement module is further merged and enhanced using a multi-scale convolution structure with four parallel branches;

[0075] The enhanced image outputted by the fusion stage is connected in the color channel dimension to all outputs from the three modules EM1o, EM2o and EM3o;

[0076] The fusion stage can fully retain the information of the enhancement module and enhance it.

[0077] Embodiment 2,

[0078] A multi-branch fusion attention mechanism image enhancement method, according to Figure 1 The following steps are shown for image enhancement:

[0079] Step S1, input the low-light image into the image enhancement model of the multi-branch fusion attention mechanism;

[0080] Step S2: an image enhancement model with a multi-branch fusion attention mechanism uses a feature extraction module to identify different areas of the low-light image and extract effective information such as edges and textures;

[0081] Step S3, an image enhancement model of a multi-branch fusion attention mechanism, using an enhancement module to set different weights to enhance the features such as edges and textures extracted in step S2, and complete denoising to obtain an enhanced feature map;

[0082] Step S4, the image enhancement model of the multi-branch fusion attention mechanism uses the fusion module to further merge and enhance the enhanced feature map output by step S3 to obtain the final enhanced image.

[0083] The image enhancement method must be implemented with the help of an image enhancement model trained under the constraints of a loss function; the loss function can reduce the distance between the enhanced image estimated by the image enhancement model and the real enhanced image corresponding to the input low-light image during the training process;

[0084] The loss function includes structural loss, perceptual loss and regional loss, and its calculation formula is:

[0085] Loss = αL SSIM +βL VGG +γL Region

[0086] Where L SSIM , L VGG and L Region They represent structural loss, perceptual loss and regional loss respectively.

[0087] The structural loss function is used to measure the difference between the real image and the enhanced image and improve the visual quality of the output image. Its calculation formula is:

[0088]

[0089] Among them, μ x and μ y are the average values ​​of image pixels x and y, respectively, x 2 , σ y 2 With σ xy denote the variance and covariance of x and y respectively, C1 and C2 are arbitrary non-zero constants whose values ​​are to prevent the denominator from being zero;

[0090] Structural loss only focuses on low-level information of the image, but high-level information is also needed to enhance image quality;

[0091] Perceptual loss, using the image style transfer algorithm to compare the features of the low-light image obtained by convolution with the features of the enhanced image, is calculated as follows:

[0092]

[0093] Among them, E and G represent the enhanced image and the real image respectively, W i,j H i,j C i,j Respectively represent the width, height and number of channels of the feature map, is the feature map obtained by the jth convolutional layer of the i-th block, i=3, j=4;

[0094] The regional loss is to constrain the degree of enhancement between the low-light area and other areas in the image, avoiding the overexposure problem caused by the overall enhancement of different areas of the entire image. Its calculation formula is:

[0095]

[0096] Where E l and G l is the low-light area of ​​the enhanced image and the real image, E h and G h is the rest of the image;

[0097] The weight value is ω l =5,ω h =1. The basis for this value is: in order to pay more attention to the low-light area in the image during training, a larger weight value is set for the low-light area.

[0098] In order to verify the effectiveness of the image enhancement method MANet of the multi-branch fusion attention mechanism of the embodiment of the present invention, experiments were conducted on a dataset of 144 synthetic low-light images and a paired dataset SCIE of different exposure compositions. The 144 synthetic low-light image dataset uses a large number of low-light images synthesized from the PASCAL VOC image dataset based on MBLLE. There are 16,925 training sets, 56 validation sets and 144 test sets, and the image size is 256×256×3. The dataset SCIE selects paired effects composed of different exposure compositions in different scenes. Each sequence has 3 to 18 low-contrast images of different exposures, and a total of 4413 multiple exposure images.

[0099] MANet was used to carry out image enhancement experiments on the above two datasets, and compared with 9 low-light enhancement networks: LIME, MSR, BIMEF [see "Ying Z, Ge L, Wen GA Bio-Inspired Multi-Exposure Fusion Framework for Low-light Image Enhancement[J]. 2017.DOI:10.48550 / arXiv.1711.00591"], SRIE, Zero-DCE, KinD, GLADNet, RetinexNet, R2RNet [see "Hai J, Xuan Z, Yang R, et al. R2RNet: Low-light Image Enhancement via Real-low to Real-normal Network[J]. 2021.DOI:10.48550 / arXiv.2106.14501"] for details.

[0100] Tables 1 and 2 were tested on 144 synthetic low-light images, and Table 3 was tested on 4413 SCIE images. “↑” means the higher the value, the better the image quality, and “↓” means the lower the value, the better the image quality.

[0101] Table 1 shows the image without noise, which shows that the proposed algorithm has a good advantage in PSNR, SSIM and NIQE, and the PSNR is improved by 17% compared with R2RNet. Table 2 shows the image with noise, which shows that the MANet model has a certain effect on noise suppression in PSNR, SSIM and NIQE compared with other methods, and the PSNR is improved by about 20% compared with R2RNet.

[0102] Table 1 Quantitative evaluation of synthetic low-light image (noise-free) enhancement

[0103]

[0104] Note: Bold indicates the best, italics indicates the second best.

[0105] Table 2 Quantitative evaluation of synthetic low-light image (noise-containing) enhancement

[0106]

[0107] Note: Bold indicates the best, italics indicates the second best.

[0108] Table 3 shows the test results on the SCIE dataset. It can be seen from the table that the MANet model has achieved better results in PSNR, SSIM and NIQE compared with other methods, and compared with R2RNet, the PSNR and SSIM are improved by 7% and 12% respectively.

[0109] Table 3 Quantitative evaluation on SCIE dataset

[0110]

[0111] Note: Bold indicates the best, italics indicates the second best.

[0112] In summary, the MANet model of the present invention achieves the best results in all indicators and also has a certain enhancement effect on low-light images.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. An image enhancement method with multi-branch fusion attention mechanism, characterized in that: To perform image enhancement, follow these steps: Step S1, input the low-light image into the image enhancement model of the multi-branch fusion attention mechanism; Step S2: an image enhancement model with a multi-branch fusion attention mechanism uses a feature extraction module to identify different areas of the low-light image and extract effective information of edges and textures; The feature extraction module consists of three groups of depthwise separable convolutional layers. The structure of each group of depthwise separable convolutional layers is Depthwise+BN+ReLu, the convolution kernel size is 3×3, and the number of channels is 32. Each group of depthwise separable convolutional layers is responsible for extracting the brightness of different areas and sending them to subsequent operations; The depthwise separable convolutional layer helps to extract effective feature information of edges and textures of low-light images while reducing the amount of computation; Step S3: an image enhancement model of a multi-branch fusion attention mechanism, using an enhancement module to set different weights to enhance the edge and texture features extracted in step S2, and to perform denoising to obtain an enhanced feature map; The enhancement module is composed of three EM enhancement modules. The EM enhancement module first uses a 3×3 convolution layer to extract effective features of the input image, and then passes through the CBAM module to identify and enhance different regions; The EM enhancement module designs the number of downsampling blocks according to the width and height of the feature map output by the downsampling, and uses 3 downsampling blocks in total. After a series of symmetrical upsampling blocks, the enhanced feature map is output. The EM enhancement module uses skip connections to introduce the features from the down-sampled image into its corresponding up-sampling module to compensate for the lost detail information and enhance the clarity. The downsampling block consists of a convolutional layer with a step size of 2 and a ReLU. The downsampling block replaces the maximum pooling layer with a CBAM attention module. First, the input features are channel pooled respectively, and then they are passed through the convolutional network respectively and the output results are element-wise concatenated to obtain the weight coefficient Mc, which is finally multiplied with the input features to obtain the output of the channel attention module. The output of the channel attention module is then used as the input of the spatial attention module, and then passed through a 3×3 convolutional layer to obtain the weight coefficient Ms, which is finally multiplied with the initial input feature to obtain the final result. All layers of this module use 3×3 convolutional layers, and the number of channels are 16, 32, 64, and 128 respectively. The enhancement module performs 1×1 convolution on the outputs EM1o, EM2o, and EM3o of the three EM enhancement modules respectively, and then fuses them to obtain the result of low-light image enhancement; Step S4: an image enhancement model of a multi-branch fusion attention mechanism, using a fusion module to further merge and enhance the enhanced feature map outputted from step S3 to obtain a final enhanced image; The fusion module is responsible for fusing the output results of the enhancement phase; The fusion stage adopts a four-scale fusion structure, and uses a multi-scale convolution structure with four parallel branches to further merge and enhance the output EMo of the enhancement module; the enhanced image output in the fusion stage connects all outputs from the three modules EM1o, EM2o, and EM3o in the color channel dimension; the fusion stage can completely retain the information of the enhancement module and enhance it.

2. According to claim 1, the image enhancement method of the multi-branch fusion attention mechanism is characterized in that: The image enhancement method must be implemented with the help of an image enhancement model trained under the constraints of a loss function; The loss function can reduce the distance between the enhanced image estimated by the image enhancement model and the real enhanced image corresponding to the input low-light image during the training process; The loss function includes structural loss, perceptual loss and regional loss, and its calculation formula is: Loss=αL SSIM +βL VGG +γL Region (1) Where L SSIM , L VGG and L Region They represent structural loss, perceptual loss and regional loss respectively.

3. According to claim 2, the image enhancement method of the multi-branch fusion attention mechanism is characterized in that: The structural loss function is used to measure the difference between the real image and the enhanced image and improve the visual quality of the output image. Its calculation formula is: Among them, μ x and μ y are the average values ​​of image pixels x and y, respectively, x 2 , σ y 2 With σ xy denote the variance and covariance of x and y respectively, C1 and C2 are arbitrary non-zero constants whose values ​​are to prevent the denominator from being zero; The structural loss only focuses on the low-level information of the image, but high-level information is also needed to enhance the image quality; The perceptual loss uses an image style conversion algorithm to compare the features of the low-light image obtained by convolution with the features of the enhanced image. The calculation formula is: Among them, E and G represent the enhanced image and the real image respectively, W i,j H i,j C i,j Respectively represent the width, height and number of channels of the feature map, It is the feature map obtained by the jth convolutional layer of the i-th block; The regional loss is to constrain the degree of enhancement between the low-light area and other areas in the image, and avoid the overexposure problem caused by the overall enhancement of different areas of the whole image. The calculation formula is: Where E l and G l is the low-light area of ​​the enhanced image and the real image, E h and G h is the rest of the image; The weight value is ω l =5,ω h =1. The basis for this value is: in order to pay more attention to the low-light area in the image during training, a larger weight value is set for the low-light area.

4. An electronic device, characterized in that: The method according to any one of claims 1 to 3 is used to enhance the low-light image.

5. A computer storage medium, characterized in that: At least one program instruction is stored in the storage medium, and the at least one program instruction is loaded and executed by the processor to implement the low-light image enhancement method according to any one of claims 1 to 3.

Images