A low-light image enhancement method based on HSV and attention mechanism

Abstract

The present invention provides a low-light image enhancement method based on HSV and attention mechanism. The method is based on the characteristic model of the human visual system. First, in view of the problem of insufficient brightness, a Retinex theory variation is proposed. This method can better process the brightness, effectively capture the brightness characteristics of the image, and improve the brightness while maintaining the hue. Secondly, in view of the problem that the saturation human eye visual effect is reduced due to the sudden change in brightness, an adaptive block saturation enhancement function is designed. The pixel is adaptively blocked through the input saturation feature, and the part with lower pixels is linearly enhanced. Finally, in view of the color unevenness, exposure and noise problems in the preliminary enhanced image, a guided attention mechanism and a fusion module are designed to fuse the two features to achieve image enhancement that is more in line with human eye vision. The present invention achieves excellent performance with a small number of parameters, and the network has good robustness.

Description

Technical Field

[0001] The present invention relates to image processing technology, and in particular to a low-illumination image enhancement method based on HSV and attention mechanism. Technical Background

[0002] Low-light image enhancement based on deep learning is an important research direction in the field of computer vision, which aims to improve the perceived quality of images collected in low-light environments. Low-light image enhancement methods have been a long-standing concern in the field of computer vision. Early low-light image enhancement methods mainly used single-channel or multi-channel enhancement techniques to extract information through a single channel or multiple channels to improve the quality of low-light images. With the development of machine learning, deep learning can learn and automate low-light image enhancement from large-scale data. In the fields of image denoising and image detection, low-light image enhancement tasks show corresponding correlations and have excellent performance in denoising and detection in low-light areas.

[0003] However, the low-light image enhancement method enhances the overall brightness, resulting in insufficient saturation and overexposure and underexposure observed by the human eye due to the brightness enhancement. Specifically, the low-light image enhancement method has the following problems: First, brightness enhancement. In the low-light image enhancement task, it is necessary to enhance the brightness from the low-light image so that the visual system can capture effective information. This task is the focus of the entire study, but the sudden increase in brightness may lead to over-enhancement and insufficient visual saturation. Therefore, how to enhance the appropriate low-light image brightness and improve saturation is one of the challenges we are currently facing. Second, image denoising. The noise in the low-light image itself and the noise caused by the brightness increase have an impact on subsequent image tasks. This requires an effective denoising process to remove the noise in the low-light image enhancement task. Third, the model parameters are large. In the low-light image enhancement task, a deep neural network model is usually required to learn the mapping from the low-light image to the normal image. Since the normal light image has a complex structure and it is necessary to judge whether the normal light image obtained is the optimal solution, in order to capture these features, the training and reasoning process requires more computing resources.

[0004] In response to the above-mentioned problems of low-light image enhancement, RetinexFormer recently designed an end-to-end model with a single-stage framework that can model well and use IGT to model long-term dependencies. CIDNet uses a bidirectional U-net structure to achieve more efficient feature extraction and fusion. SIDNet uses Retinex theory to remove degraded illumination from mixed features and generate features for enhanced images. Although the existing technology has made substantial progress, the problem of insufficient visual saturation caused by a sudden increase in brightness has not been solved. Secondly, when the attention mechanism is used for denoising tasks, it does not pay attention to the connection between brightness and saturation and the problem of feature information loss during information fusion. Therefore, designing an efficient low-light image enhancement model is of great research significance. Summary of the invention

[0005] The purpose of the present invention is to solve the problems of color saturation deviation and information fusion in low-light image enhancement, and to provide a low-light image enhancement method based on HSV and attention mechanism, which can restore the normal light image through a deep learning network according to the input low-light image.

[0006] Taking the example of realizing the above-mentioned purpose, the present invention provides a low-light image enhancement method based on HSV and attention mechanism, which mainly includes the following seven parts: the first part is to preprocess the low-light image data set; the second part is to enhance the brightness and saturation of the low-light image and extract shallow features; the third part is to guide the brightness attention of the enhanced image; the fourth part is to guide the saturation attention of the features after shallow feature extraction; the fifth part is to iterate the operation from the third part to the fourth part; the sixth part is to perform image fusion operation on the image; the seventh part is to train and test the low-light image enhancement network model based on HSV and attention mechanism, and finally obtain the enhanced normal light image, and the specific steps are as follows:

[0007] The first part consists of two steps:

[0008] Step 1: Download the public dataset of low-light images and select data with low-light image pairs, which include indoor and outdoor scenes, as training data and test data;

[0009] Step 2: Crop the low-light image and the normal-light image into image blocks of one-to-one corresponding size according to the sampling size of 160×160, perform data enhancement operation on the sampled data to expand the data samples and form the final training set samples;

[0010] The second part consists of five steps;

[0011] Step 3: Input the training sample image in step 2 into the network, and separate three channels by transforming the features from RGB space to HSV space, namely, hue channel feature H1, saturation channel feature S1, and lightness channel feature V1;

[0012] Step 4: The saturation channel feature S1 separated in step 3 is subjected to an adaptive block saturation enhancement method BSE (block saturation enhancement), that is, low saturation is linearly enhanced and high saturation is not processed to avoid oversaturation, thereby obtaining an enhanced saturation feature S2;

[0013] Step 5: The hue channel feature H1 in step 3 is fused with the lightness channel feature V1 and the saturation feature S1 to obtain the HSV image feature L1, which is then fused with the lightness channel feature V1 in step 3 to obtain L2, and a shallow feature extraction operation is performed to obtain the brightness feature V2, and the brightness feature V2 is subjected to a convolution operation to obtain the brightness feature map V3;

[0014] Step 6: Based on the variation of Retinex theory, perform Retinex variation operation on the brightness feature V3 obtained in step 5 and the HSV low-light image feature L2 to obtain a preliminary normal-light image G1;

[0015] Step 7: The HSV image feature L1 obtained in step 5 is fused with the enhanced saturation feature S2 in step 4 and a shallow feature extraction operation is performed to obtain a saturation feature S3. The saturation feature S3 is subjected to a convolution operation to obtain a saturation feature map S4;

[0016] The third part consists of two steps:

[0017] Step 8: The preliminary normal illumination image G1 obtained in step 6 is subjected to a convolution process to obtain a normal illumination image shallow feature G2 having the same BatchSize, Channels, Height, and Weight as the brightness feature V2 in step 5;

[0018] Step 9: Take the shallow feature G2 of the normal illumination image in step 8 as input, use the U-shaped structure network, and process it through the brightness guided attention module IGAB to obtain the normal illumination image brightness deep feature G3.1, where the brightness guided attention uses the brightness feature V2 in step 5 as a guide;

[0019] Part 4 consists of two steps:

[0020] Step 10: The feature S4 obtained in step 7 is subjected to a convolution process to obtain a shallow feature S5 with the same BatchSize, Channels, Height, and Weight as the saturation feature S3 in step 7;

[0021] Step 11: Take the shallow image feature S5 in step 10 as input, use the U-shaped structure network, and process it through the saturation guided attention SGAB to obtain the normal illumination image saturation deep feature S6.1, where the saturation guided attention uses the saturation feature S3 in step 7 as a guide;

[0022] Part 5 consists of two steps:

[0023] Step 12: The encoder consists of three combination blocks. The output of step 9 is used as the input of step 9. The three combination blocks are IGAB iterated 2 times, 2 times, and 1 times; the decoder consists of two combination blocks. The second combination block and the third combination block in the encoder are fused as the input of the first combination block in the decoder. The output of the first combination block in the encoder and the first combination block of the decoder are fused as the input of the second combination block in the decoder. The two combination blocks are IGAB iterated 2 times and 2 times respectively. The intermediate feature G3.n output in step 9 is obtained, where n = 2, 3, 4, 5, 6, 7, 8, 9;

[0024] Step 13: The encoder consists of three combination blocks. The output of step 11 is used as the input of step 11. The three combination blocks are SGAB iterated 2 times, 2 times, and 1 times; the decoder consists of two combination blocks. The second combination block and the third combination block in the encoder are fused as the input of the first combination block in the decoder. The output of the first combination block in the encoder and the first combination block of the decoder are fused as the input of the second combination block in the decoder. The two combination blocks are SGAB iterated 2 times and 2 times respectively. The intermediate feature S6.n output in step 9 is obtained, where n = 2, 3, 4, 5, 6, 7, 8, 9;

[0025] Part 6 includes a step:

[0026] Step 14: The G3.9 obtained in step 12 and the S6.9 obtained in step 13 are fused into a final normal illumination map using a fusion module;

[0027] Part 7 consists of two steps:

[0028] Step 15: Input the training set samples in step 2 into the network from step 3 to step 14, set the network hyperparameters: learning rate is 2e-4, batch size is 8, optimizer is Adam, loss function is L1Loss, train the network to obtain the final low-light image enhancement pre-training model;

[0029] Step 16: Input the low-light image of the public test set into the pre-trained model obtained in step 15, and the network can restore the normal-light image.

[0030] The present invention provides a low-light image enhancement method based on HSV and attention mechanism. The method is based on the characteristic model of the human visual system. First, in view of the problem of insufficient brightness, the brightness component of the low-light image is extracted, and the brightness component is fused with the HSV image of the low-light image. The brightness feature is obtained by shallow feature extraction, and the Retinex theory variation operation is performed on it and the HSV image of the low-light image. This method can better process the brightness, effectively capture the brightness feature of the image, and improve the brightness while maintaining the hue. Secondly, in view of the problem that the saturation human eye visual effect is reduced due to the sudden change of brightness, an adaptive block saturation enhancement function BSE is designed. BSE performs adaptive pixel block by input saturation feature, linearly enhances the lower pixel part, and keeps the higher pixel part unchanged. Then, in view of the color unevenness, exposure and noise problems of the preliminary enhanced image, according to the U-type network, the encoder and decoder are used for processing respectively, the brightness enhancement image feature and the saturation feature are respectively introduced into the denoising module as input, and the brightness feature and the saturation feature are respectively added as guides, and the two features are fused to achieve image enhancement that is more in line with human eye vision. Finally, in order to solve the problem of detail loss, a fusion module is designed to fuse multiple features to enrich the image information. The present invention realizes the low-light image brightness enhancement operation by performing Retinex processing on the brightness component of the image, adopts an adaptive saturation enhancement function, introduces brightness attention and saturation attention mechanisms, makes the network pay more attention to the image brightness and saturation features, achieves excellent performance with fewer parameters, and has good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is the overall framework diagram of the network of the present invention;

[0032] Figure 2 This is a flow chart of the adaptive saturation enhancement of the present invention;

[0033] Figure 3 The guided attention mechanism GBA of the present invention;

[0034] Figure 4 The fusion module Fusion of the present invention;

[0035] Figure 5 For low-light images;

[0036] Figure 6 To utilize the present invention to process Figure 5 The enhanced image after. DETAILED DESCRIPTION

[0037] In order to better understand the present invention, the low-light image enhancement method based on HSV and attention mechanism of the present invention is described in more detail below in combination with specific embodiments. In the following description, the detailed description of the current existing technology may dilute the subject content of the present invention, and these descriptions will be ignored here.

[0038] Step 1: Download the public dataset of low-light images and select data with low-light image pairs, which include indoor and outdoor scenes, as training data and test data;

[0039] Step 2: Crop the low-light image and the normal-light image into image blocks of one-to-one corresponding size according to the sampling size of 160×160, perform data enhancement operation on the sampled data to expand the data samples and form the final training set samples;

[0040] Figure 1 This is the overall network framework diagram of the low-light image enhancement method based on HSV and attention mechanism of the present invention. In this embodiment, the following steps are performed:

[0041] Step 3: Input the training sample image in step 2 into the network, and separate three channels by transforming the features from RGB space to HSV space, namely, hue channel feature H1, saturation channel feature S1, and lightness channel feature V1;

[0042] Step 4: Adaptive Saturation Enhancement BSE process as follows Figure 2 As shown, the saturation channel feature S1 separated in step 3 is subjected to an adaptive block saturation enhancement method BSE, that is, low saturation is linearly enhanced, and high saturation is not processed to avoid oversaturation, thereby obtaining an enhanced saturation feature S2;

[0043] Step 4-1: The linear enhancement function first performs mean processing on the saturation channel of the input image, then divides the entire feature into four equal parts, extracts the pixels contained in the four parts of the input image, and obtains the corresponding ratios μ1, μ2, μ3 and μ4.

[0044] Step 4-2: The corresponding enhancement function required for the saturation of each image is obtained by setting certain parameters. The specific formula is:

[0045] α=-0.40μ1-0.12μ2+0.19μ3+0.83μ4

[0046] S en =(3.354-4.427α)·S

[0047] Among them, μ1, μ2, μ3 and μ4 are four block ratios of the input image features, and S is the saturation tensor of the low-saturation part of the input image.

[0048] Step 5: The hue channel feature H1 in step 3 is fused with the lightness channel feature V1 and the saturation feature S1 to obtain the HSV image feature L1, which is then fused with the lightness channel feature V1 in step 3 to obtain L2, and a shallow feature extraction operation is performed to obtain the brightness feature V2, and the brightness feature V2 is subjected to a convolution operation to obtain the brightness feature map V3;

[0049] Step 6: Based on the variation of Retinex theory, the brightness feature V3 obtained in step 5 is subjected to variation operation with the HSV low-light image feature L2 to obtain a preliminary normal-light image G1. The specific formula is:

[0050]

[0051] Among them, R is the reflection component, L is the illumination component under ideal conditions, are different illumination components in actual situations, α i is the impact factor, and n is the noise.

[0052] Step 7: The HSV image feature L1 obtained in step 5 is fused with the enhanced saturation feature S2 in step 4 and a shallow feature extraction operation is performed to obtain a saturation feature S3. The saturation feature S3 is subjected to a convolution operation to obtain a saturation feature map S4;

[0053] Step 8: The preliminary normal illumination image G1 obtained in step 6 is subjected to a convolution process to obtain a normal illumination image shallow feature G2 having the same BatchSize, Channels, Height, and Weight as the brightness feature V2 in step 5;

[0054] Step 9: Take the shallow feature G2 of the normal illumination image in step 8 as input, use the U-shaped structure network, and process it through the brightness guided attention module IGAB to obtain the deep brightness feature G3.1 of the normal illumination image;

[0055] Step 9-1: Guide the attention mechanism GAB as Figure 3 As shown, IGAB is a GAB module guided by brightness features, where Guided_feature uses the brightness feature V2 in step 5 as a guide;

[0056] Step 9-2: The feature G2 in step 8 is subjected to normalization operations LN and G_MSA to obtain the high-frequency information focusing on the brightness feature, and then the high-frequency information is fused with the feature G2 in step 8 to obtain M1;

[0057] Step 9-3: The feature M1 in step 9-2 is subjected to the normalization operation LN and the feed-forward neural module FFN, and then fused with the feature M1 in step 9-2 to obtain the output feature;

[0058] Step 10: The feature S4 obtained in step 7 is subjected to a convolution process to obtain a shallow feature S5 with the same BatchSize, Channels, Height, and Weight as the saturation feature S3 in step 7;

[0059] Step 11: Take the shallow image feature S5 in step 10 as input, use the U-shaped structure network, and process it through the saturation guided attention SGAB to obtain the deep saturation feature S6.1 of the normal illumination image. Use a double branch connection, which is the same as step 9, except that SGAB is a GAB module guided by the saturation feature, and the Guided_feature uses the saturation feature S3 in step 7 as a guide;

[0060] Step 12: The encoder consists of three combination blocks. The output of step 9 is used as the input of step 9. The three combination blocks are IGAB iterated 2 times, 2 times, and 1 times; the decoder consists of two combination blocks. The second combination block and the third combination block in the encoder are fused as the input of the first combination block in the decoder. The output of the first combination block in the encoder and the first combination block of the decoder are fused as the input of the second combination block in the decoder. The two combination blocks are IGAB iterated 2 times and 2 times respectively. The intermediate feature G3.n output in step 9 is obtained, where n = 2, 3, 4, 5, 6, 7, 8, 9;

[0061] Step 13: The encoder consists of three combination blocks. The output of step 11 is used as the input of step 11. The three combination blocks are SGAB iterated 2 times, 2 times, and 1 times; the decoder consists of two combination blocks. The second combination block and the third combination block in the encoder are fused as the input of the first combination block in the decoder. The output of the first combination block in the encoder and the first combination block of the decoder are fused as the input of the second combination block in the decoder. The two combination blocks are SGAB iterated 2 times and 2 times respectively. The intermediate feature S6.n output in step 9 is obtained, where n = 2, 3, 4, 5, 6, 7, 8, 9;

[0062] Step 14: The G3.9 obtained in step 12 and the S6.9 obtained in step 13 are fused using a fusion module to obtain the final normal illumination map. The specific fusion module is as follows: Figure 4 As shown, the final G3.9 obtained in step 12 and the final S6.9 obtained in step 13 are put into the complementary attention PAM as input, and the obtained features are fused with the final G3.9 obtained in step 12 and the final S6.9 obtained in step 13 to obtain the output image.

[0063] Step 15: Input the training set samples in step 2 into the network from step 3 to step 12, set the network hyperparameters: learning rate is 2e-4, batch size is 8, optimizer is Adam, loss function is L1Loss, train the network to obtain the final low-light image enhancement pre-training model;

[0064] Step 16: Input the low-light image of the public test set into the pre-trained model obtained in step 13, and the network can restore the normal-light image.

[0065] Starting from the problem of insufficient brightness and color saturation deviation of low-light images, the present invention provides a low-light image enhancement method based on HSV and attention mechanism. First, the method adopts the fusion of HSV channel and brightness channel in brightness enhancement, and uses the Retinex variant through the shallow extraction module to solve the problem of insufficient brightness in low illumination; secondly, in order to solve the saturation human eye visual deviation caused by sudden brightness change, the method introduces the adaptive saturation enhancement module BSE; then, in order to better enhance the low-light image and solve the problem of noise in the enhanced image, the method introduces the guided attention mechanism, takes the brightness feature and the saturation feature as the guiding information respectively, and iterates multiple times to make the network pay more attention to the brightness feature and the saturation feature, so as to achieve better learning effect; finally, the method designs the fusion module Fusion to fuse multiple features to make the image information richer. The algorithm of the present invention is simple and highly operable, and is suitable for low-light images in actual scenes and simulated degraded low-light interference images.

[0066] Although the above describes an illustrative specific embodiment of the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiment. For ordinary technicians in this technical field, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A low-light image enhancement method based on HSV and attention mechanism, characterized in that: The HSV channel processing is combined with Retinex theory to enhance the low-light image, and the guided attention is combined with the complementary attention, including data set preprocessing, brightness and saturation enhancement and shallow feature extraction of low-light images, brightness attention guidance for enhanced images, saturation attention guidance for features after shallow feature extraction, iterative operation, image fusion operation, and network model training and testing. The first part consists of two steps: Step 1: Download the public dataset of low-light images and select data with low-light image pairs, which include indoor and outdoor scenes, as training data and test data; Step 2: Crop the low-light image and the normal-light image into image blocks of one-to-one corresponding size according to the sampling size of 160×160, perform data enhancement operation on the sampled data to expand the data samples and form the final training set samples; The second part consists of five steps; Step 3: Input the training sample image in step 2 into the network, and separate three channels by transforming the features from RGB space to HSV space, namely, hue channel feature H1, saturation channel feature S1, and lightness channel feature V1; Step 4: The saturation channel feature S1 separated in step 3 is subjected to the adaptive block saturation enhancement method BSE, that is, low saturation is linearly enhanced, and high saturation is not processed to avoid oversaturation, to obtain the enhanced saturation feature S2, which is specifically implemented as follows: (1) The linear enhancement function first performs mean processing on the saturation channel of the input image, then divides the entire feature into four equal parts, extracts the pixels contained in the four parts of the input image, and obtains the corresponding ratios μ1, μ2, μ3, and μ4; (2) By setting certain parameters, the corresponding enhancement function required for the saturation of each image is obtained. The specific formula is: α=-0.40μ1-0.12μ2+0.19μ3+0.83μ4 S en =(3.354-4.427α)·S Among them, μ1, μ2, μ3 and μ4 are the four block ratios of the input image features, and S is the saturation tensor of the low-saturation part of the input image; Step 5: The hue channel feature H1 in step 3 is fused with the lightness channel feature V1 and the saturation feature S1 to obtain the HSV image feature L1, which is then fused with the lightness channel feature V1 in step 3 to obtain L2, and a shallow feature extraction operation is performed to obtain the brightness feature V2, and the brightness feature V2 is subjected to a convolution operation to obtain the brightness feature map V3; Step 6: Based on the variation of Retinex theory, the brightness feature V3 obtained in step 5 is subjected to variation operation with the HSV low-light image feature L2 to obtain a preliminary normal-light image G1; Step 7: The HSV image feature L1 obtained in step 5 is fused with the enhanced saturation feature S2 in step 4 and a shallow feature extraction operation is performed to obtain a saturation feature S3. The saturation feature S3 is subjected to a convolution operation to obtain a saturation feature map S4; The third part consists of two steps: Step 8: The preliminary normal illumination image G1 obtained in step 6 is subjected to a convolution process to obtain a normal illumination image shallow feature G2 having the same BatchSize, Channels, Height, and Weight as the brightness feature V2 in step 5; Step 9: Take the shallow feature G2 of the normal illumination image in step 8 as input, use the U-shaped structure network, and process it through the brightness guided attention module IGAB to obtain the deep brightness feature G3.1 of the normal illumination image; (1) IGAB is a GAB module guided by brightness features, where Guided_feature uses the brightness feature V2 in step 5 as a guide; the feature G2 in step 8 is normalized by LN and G_MSA to obtain the high-frequency information focusing on the brightness feature, and then fused with the feature G2 in step 8 to obtain M1; (2) The feature M1 is subjected to the normalization operation LN and the feed-forward neural module FFN, and then fused with the feature M1 to obtain the output feature; Part 4 consists of two steps: Step 10: The feature S4 obtained in step 7 is subjected to a convolution process to obtain a shallow feature S5 with the same BatchSize, Channels, Height, and Weight as the saturation feature S3 in step 7; Step 11: Take the shallow image feature S5 in step 10 as input, use the U-shaped structure network, and process it through the saturation guided attention SGAB to obtain the deep saturation feature S6.1 of the normal illumination image. Use a double branch connection, which is the same as step 9, except that SGAB is a GAB module guided by the saturation feature, and the Guided_feature uses the saturation feature S3 in step 7 as a guide; Part 5 consists of two steps: Step 12: The encoder is composed of three combination blocks. The output in step 9 is used as the input of step 9. The three combination blocks are IGAB iterated 2 times, 2 times, and 1 times; the decoder is composed of two combination blocks. The second combination block and the third combination block in the encoder are fused as the input of the first combination block in the decoder. The output of the first combination block in the encoder and the first combination block of the decoder are fused as the input of the second combination block in the decoder. The two combination blocks are IGAB iterated 2 times and 2 times respectively; the intermediate feature G3.n output in step 9 is obtained, where n = 2, 3, 4, 5, 6, 7, 8, 9; Step 13: The encoder is composed of three combination blocks. The output in step 11 is used as the input of step 11. The three combination blocks are SGAB iterated 2 times, 2 times, and 1 times; the decoder is composed of two combination blocks. The second combination block and the third combination block in the encoder are fused as the input of the first combination block in the decoder, and the output of the first combination block in the encoder and the first combination block of the decoder are fused as the input of the second combination block in the decoder. The two combination blocks are SGAB iterated 2 times and 2 times respectively; the intermediate feature S6.n output in step 9 is obtained, where n = 2, 3, 4, 5, 6, 7, 8, 9; Part 6 includes a step: Step 14: Put the final G3.9 obtained in step 12 and the final S6.9 obtained in step 13 as input into the complementary attention PAM, and fuse the obtained features with the final G3.9 obtained in step 12 and the final S6.9 obtained in step 13 to obtain the output image; Part 7 consists of two steps: Step 15: Input the training set samples in step 2 into the network from step 3 to step 12, set the network hyperparameters: learning rate is 2e-4, batch size is 8, optimizer is Adam, loss function is L1Loss, train the network to obtain the final low-light image enhancement pre-training model; Step 16: Input the low-light image of the public test set into the pre-trained model obtained in step 13, and the network can restore the normal-light image.

2. According to claim 1, a low-light image enhancement method based on HSV and attention mechanism is characterized in that: In step 4, adaptive block saturation adjustment BSE is used for the S channel. The saturation of the image itself is used to divide the range into four areas. The saturation factor of each area is calculated to achieve the effect of enhancing the saturation.

3. The low-light image enhancement method based on HSV and attention mechanism according to claim 1, characterized in that: A variation of the Retinex theory in step 6 uses the brightness feature obtained by shallow feature extraction after HSV and brightness component fusion to perform dot multiplication with HSV to better enhance the brightness of low-light images.

4. The low-light image enhancement method based on HSV and attention mechanism according to claim 1, characterized in that: The dual-branch connection of the parallel structure of the third and fourth parts uses the same overall steps to guide different features, which effectively reduces the parameters and can focus on two features that need to be learned.

5. The low-light image enhancement method based on HSV and attention mechanism according to claim 1, characterized in that: The fusion module in step 14 introduces the complementary mechanism PAM to aggregate different features.

Images