A low-light image enhancement method combining attention mechanism and Retinex model

Abstract

The present invention discloses a low-light image enhancement method combining an attention mechanism and a Retinex model, and relates to the technical field of low-light images. The present invention proposes an information extraction network, which can learn to obtain the reflectivity, illumination and attention map of an image. First, we use Attention‑Net to obtain the attention map of the image, and then decouple the original space into two smaller subspaces through Decomposition‑Net. Secondly, the attention map is used to guide the reflectivity recovery in a regional adaptive manner. The present invention proposes a technology CA&R Net that combines the attention mechanism and the Retinex model to enhance low-light images. Under the guidance of the attention information, CA&R Net can successfully handle noise, color distortion and various types of degradation; the present invention formulates an attention map to guide the reflectivity recovery in a regional adaptive manner, so as to pay more attention to the underexposed areas during the enhancement process and avoid over-enhancing the normal exposure areas.

Description

Technical Field

[0001] The present invention belongs to the technical field of low-light images, and in particular relates to a low-light image enhancement method combining an attention mechanism and a Retinex model. Background Art

[0002] Low-light image enhancement has always been a hot topic in academic research, because these images not only have low visibility and cannot meet people's ideal visual effects and needs, but also bring great challenges to traditional computer vision tasks such as image segmentation, target detection and tracking. Early studies used frequency domain methods to enhance images. By enhancing wavelet coefficients, image details can be effectively highlighted, but at the same time, the noise hidden in the image will be amplified. Moreover, this type of algorithm requires a lot of calculations, and the selection of transformation parameters cannot be automatically selected; image fusion methods use image processing and computer vision technology to extract useful information from different images of the same target collected without the need for physical models, and synthesize high-quality images. Their enhancement effect is good, but when faced with real-time monitoring scenarios, it is difficult to achieve enhancement in a short time; defogging models have good performance and can enhance the visibility of images to a certain extent, but they lack the physical interpretation of the basic model and require some additional noise reduction processing to eliminate noise, which will inevitably lead to blurred details; methods based on traditional Retinex models have not only achieved good results in improving image contrast and brightness, but also have obvious advantages in color image enhancement, but such algorithms are only constructed under manual constraints and are not enough to adapt to the complex signal characteristics of various natural images.

[0003] In recent years, deep learning has achieved great success in the field of low-level image processing. While bringing great changes to the low-level image processing tasks, it has also brought significant improvements to the performance of low-light image enhancement. Lore et al. created a deep autoencoder to enhance low-light images to achieve adaptive brightness adjustment and denoising; Chen et al. introduced the short-exposure low-light image dataset SID in Raw format and proposed a method based on a fully convolutional network to process these images; Wei et al. constructed a paired dataset containing low-light images and normal-light images and proposed a deep network called Retinex-Net; Zhang et al. built a simple and effective Kind of Darkness (KinD) network, which consists of three subnetworks: layer decomposition, reflectivity restoration, and illumination adjustment, and tested it using paired datasets; Lv et al. proposed a new end-to-end attention-guided method based on a multi-branch convolutional neural network. The method simultaneously learns two attention maps to guide the brightness enhancement and denoising tasks respectively, which can achieve low-light enhancement and denoising at the same time; Xu et al. proposed a frequency-based decomposition and enhancement model, which first uses the attention context encoding module to restore the low-frequency layer image content and denoise, and then uses the restored image to enhance the high-frequency details; Fan et al. integrated the ideas of Retinex decomposition and semantic information perception, introduced semantic information into the Retinex model, and guided the enhancement of the two components; Jiang et al. applied unpaired learning to test the low-light enhancement model, which got rid of the construction of paired datasets and solved the domain transfer problem between the test dataset and the actual test application, but it is easy to cause underexposure and distortion in some areas of the image when used. Summary of the invention

[0004] The purpose of the present invention is to provide a low-light image enhancement method combining an attention mechanism and a Retinex model to solve the existing problem that the existing image enhancement technology is prone to underexposure and distortion of some areas of the image when used.

[0005] In order to solve the above technical problems, the present invention is achieved through the following technical solutions:

[0006] The present invention is a low-light image enhancement method combining an attention mechanism and a Retinex model, comprising the following steps:

[0007] Establish information extraction network, reflectance restoration network and light regulation network;

[0008] Create three subnets for f decom (·), f restore (·) and f adjust (·) Modeling;

[0009] Decomposition-Net is used to decompose the paired low-light image and normal-light image into reflectance R low / R high and Lighting I low / I high .

[0010] Furthermore, the information extraction network includes:

[0011] Image decomposition network: used to extract reflectance and illumination from images simultaneously;

[0012] Attention Network: Extract the image’s attention map A through Attention-Net map , A map to detect the degree of underexposure and use it to guide the reflectivity restoration network in the Restore-Net stage.

[0013] Further, wherein the image decomposition network includes a loss function for constraints;

[0014] The loss function consists of reflectivity similarity loss, illumination smoothness loss, mutual consistency loss, and reconstruction loss.

[0015] Furthermore, the calculation formula of the reflectivity similarity loss is:

[0016] L rs =||R low -R high ||2

[0017] Among them, L rs represents the reflectance similarity, and ||·||2 represents the 2-norm

[0018] The calculation formula of the illumination smoothness loss is:

[0019]

[0020] Among them, L is Indicates the lighting smoothness, represents the first-order derivative operator (including the first-order derivative in the x-direction and the y-direction), ||·||1 represents the 1-norm, and |·| represents the absolute value operator;

[0021] The calculation formula of the mutual consistency loss is:

[0022]

[0023]

[0024] Among them, L mc Indicates mutual consistency, c is a parameter

[0025] The calculation formula of the reconstruction loss is:

[0026]

[0027] Decomposition-Net decomposes the input image into reflectance and illumination. When reconstructing these two components, the input image should also be obtained. We use L rec To express;

[0028] Therefore, the total loss function of the image decomposition network is expressed as:

[0029] L decom =L rec +λ rs L rs +λ is L is +λ mc L mc

[0030] Among them, λ rs ,λ is ,λ mc They are the weight coefficients of reflectance similarity loss, illumination smoothness loss, and mutual consistency loss respectively.

[0031] Furthermore, the reflectivity recovery network includes:

[0032] Residual in Residual (RIR) module: The RIR module uses the deep residual to provide a powerful pixel adjustment capability for the reflectivity restoration network, which can effectively suppress noise and perform color correction;

[0033] Attention Residual in Residual (ARIR) module, which first uses two two-layer convolution from A map Extract features F1 and F2, and then use several convolutional layers to pass through the RIR module from R low Extract feature F3, and finally fuse the three features F1, F2, and F3 by multiplying F3 by F2 and then adding F1. After fusion, the output of the ARIR module is obtained through convolution operation;

[0034] The loss function can be expressed as follows:

[0035]

[0036] where λ i is the weight coefficient, MSE is the mean square error, SSIM is the structural similarity index, and Grad is the gradient loss.

[0037] Furthermore, the loss function of the illumination adjustment network stage can be expressed as:

[0038]

[0039] where λ i is the weight coefficient.

[0040] The present invention has the following beneficial effects:

[0041] 1. This paper proposes a technique called CA&R Net that combines the attention mechanism and the Retinex model to enhance low-light images. Under the guidance of attention information, CA&R Net can successfully handle noise, color distortion, and various types of degradation.

[0042] 2. The present invention develops an attention map to guide reflectance recovery in a region-adaptive manner so as to pay more attention to underexposed areas during the enhancement process and avoid over-enhancing normally exposed areas.

[0043] 3. The present invention uses the reconstructed reflectivity and low illumination for joint prediction. This joint prediction utilizes attention information and also introduces more contextual information into the light adjustment network, breaking the limitation of the single design of the light adjustment network and enabling the light adjustment to achieve better results. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0045] Figure 1 It is the overall network framework diagram of the present invention;

[0046] Figure 2 It is the network structure diagram module of the present invention;

[0047] Figure 3 The first visual comparison between the present invention and other low-light image enhancement methods;

[0048] Figure 4 A second visual comparison between the present invention and other low-light image enhancement methods;

[0049] Figure 5 A third visual comparison between the present invention and other low-light image enhancement methods;

[0050] Figure 6 This is the ablation experiment of the present invention. DETAILED DESCRIPTION

[0051] 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.

[0052] The present invention is a low-light image enhancement method combining an attention mechanism and a Retinex model, comprising the following steps:

[0053] Establish information extraction network, reflectance restoration network and light regulation network;

[0054] Create three subnets for f decom (·), f restore (·) and f adjust (·) Modeling;

[0055] In the decomposition stage, Decomposition-Net is used to decompose the paired low-light image and normal-light image into reflectivity R low / R high and Light I low / I high (expressed as f decom (·)), in which the low-light image and the normal-light image share the same reflectance and illumination smoothness constraints and are autonomously learned in a data-driven manner;

[0056] In the reflectivity recovery stage, Restore-Net is used to restore the reflectivity and suppress the hidden noise, and reflectivity recovery is achieved in a regional adaptive manner under the guidance of restore (·));

[0057] In the illumination adjustment stage, Adjust-Net is used to adjust the illumination, and the reconstructed reflectance and low illumination are used to jointly predict the illumination layer of the image. This joint prediction utilizes the attention information to achieve better illumination adjustment results (expressed as f adjust (·)).

[0058] The information extraction network includes:

[0059] 1.1. Image decomposition network: It is a challenging task to extract both reflectance and illumination from an image. Ground truth requires professionals with superb photography skills to shoot and undergo a lot of processing. Usually, such image data is difficult to obtain. Therefore, it is particularly important to design a reasonable loss function to constrain network testing without the guidance of ground truth. In the network, low-light images and normal-light images with different exposure levels are used [S low ,S high ] were tested, and based on the fact that low-light images and normal-light images share the same reflectivity, and the property that illumination maps are piecewise smooth and mutually consistent, a loss function for constraints was designed, which consists of four parts: reflectivity similarity loss, illumination smoothness loss, mutual consistency loss, and reconstruction loss.

[0060] Reflectance Similarity Loss:

[0061] L rs =||R low -R high ||2

[0062] Use L rs to normalize the reflectance similarity, where ||·||2 represents the 2-norm.

[0063] Lighting smoothness loss:

[0064]

[0065] Lighting smoothness is determined by L is To constrain, represents the first-order derivative operator (including the first-order derivative in the x-direction and the y-direction), ||·||1 represents the 1-norm, and |·| represents the absolute value operator. In addition, in order to avoid the denominator being 0, a very small positive constant ε is added to the denominator;

[0066] Here, the addition proposed here is not addition in the mathematical sense, but is implemented by the max() function, so that if it is not 0, the denominator is it. If the denominator is 0, it is replaced by a given very small normal number to avoid the denominator being 0 and ensure that the entire formula can be calculated normally.

[0067] Mutual consistency loss:

[0068] L mc =||X·exp(-c·X)||1

[0069]

[0070] Mutual consistency is expressed as L mcIndicates that c is a parameter.

[0071] Reconstruction loss:

[0072] L rec =||S low -R low I low ||1+||S high -R high I high ||1

[0073] Decomposition-Net decomposes the input image into reflectance and illumination. When reconstructing these two components, the input image should also be obtained. rec to constrain it.

[0074] Therefore, the total loss function of the image decomposition network is expressed as:

[0075] L decom =L rec +λ rs L rs +λ is L is +λ mc L mc

[0076] where λ rs ,λ is ,λ mc They are the weight coefficients of reflectance similarity loss, illumination smoothness loss, and mutual consistency loss respectively.

[0077] 1.2. Attention Network: Extract the image’s attention map A through Attention-Net map , and use it to guide reflectivity recovery in the Restore-Net stage. It is observed that the noise in the underexposed area of ​​the image is more serious than that in the normally exposed area. To effectively enhance the contrast of the image and suppress the noise, the key is to solve the problem in a region-aware adaptive way. To this end, the present invention designs an A map To detect the degree of underexposure and guide the enhancement in a region-adaptive manner, so as to better enhance the underexposed areas and avoid over-enhancing the normally exposed areas. U-Net is used to achieve this goal, with the aim of providing guidance for subsequent enhancement of underexposed areas and avoiding over-enhancing the normally exposed areas. The output is an attention map that represents the degree of underexposure. It can be expressed by the following formula:

[0078]

[0079] where max c(·) means returning the maximum value among the three color channels. high represents the normal light image, S low represents a low-light image. In order to obtain A map , constrained by the following formula:

[0080] L attention =||F a (S)-A||2

[0081] Where S represents the low-light image, F a (·) and A are the predicted and expected results, respectively.

[0082] Reflectivity Restoration Network:

[0083] Since the reflectivity R decomposed by Decomposition-Net low There is strong noise and color deviation. This paper uses Restore-Net to restore R under the guidance of attention information. low .

[0084] In the reflectivity restoration stage, inspired by Fan et al., the Residual in Residual (RIR) module is introduced. The RIR module uses deep residual learning to provide a powerful pixel adjustment capability for the reflectivity restoration network, which can effectively suppress noise and perform color correction. In order to introduce attention information, an Attention Residual in Residual (ARIR) module is designed. This module first uses two two-layer convolutions from A map Extract features F1 and F2, and then use several convolutional layers to pass through the RIR module from R low The feature F3 is extracted from the image, and finally the three features F1, F2, and F3 are fused by multiplying F3 by F2 and then adding F1. After fusion, the output of the ARIR module is obtained by convolution operation. In this stage, the convolution operation is first used to extract features, and then the features are fused through a series of RIR and ARIR modules, and finally the output result is obtained through convolution and Sigmoid layer. The loss function of this stage can be expressed as follows:

[0085]

[0086] where λ i is the weight coefficient, MSE is the mean square error, SSIM is the structural similarity index, and Grad is the gradient loss.

[0087] Lighting adjustment network:

[0088] Illumination component I decomposed by Decomposition-Netlow Usually visibility is low, in order to adjust I low , the present invention designs an Adjust-Net and uses the restored reflectivity and low light I low to jointly predict the illumination layer of the image.

[0089] First, use U-Net to Extract features from map The information will be introduced into Adjust-Net. Secondly, two layers of convolution are used to enhance the ratio α and I low Extract features from the network, fuse them with the features extracted by U-Net, and then perform convolution to get the adjusted illumination. Then use the Sigmoid layer to adjust the illumination of the network. is limited to [0,1]. Finally, the restored reflectivity is used and adjusted lighting Reconstruct according to element-by-element multiplication, the final prediction result

[0090] Since there is no real light level of an image, it is necessary to design a mechanism for converting light conditions to meet the needs of different scenes. In our dataset, we use pairs of low-light images and normal-light images, that is, we have pairs of light maps;

[0091] Even if we do not know the exact relationship between the paired illumination maps, we can use α = I high / I low to calculate their intensity ratio;

[0092] This ratio parameter can be used to guide the illumination adjustment process. During the testing phase, α can be provided by the user;

[0093] We use the illumination of the normal light image I high To guide the adjustment network, it introduces more contextual information for model testing. The loss function of this stage can be expressed as:

[0094]

[0095] where λ i is the weight coefficient.

[0096] The detection experiment for optical image enhancement technology includes the following steps:

[0097] 1. Quantitative measurement and detection

[0098] Since detecting the quality of enhanced images is not a simple task, in the present invention, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) commonly used in the field of image enhancement are used as the detection results (the larger the PSNR value, the better, and the closer the SSIM value is to 1, the better). As can be seen from Table 1, the present invention is superior to other advanced methods in both indicators, which proves the rationality and effectiveness of the present invention from a quantitative aspect.

[0099] Among them, Retinex-Net, Retinex-Net-TIP, and KinD involved in the comparison are all based on the Retinex model. In the subsequent enhancement network, Retinex-Net only designs a Relight-Net to re-estimate the brightness of the image, without considering the influence of the two components of reflectivity and illumination on the image at the same time, so it cannot achieve ideal enhancement. The image decomposition network of KinD is well designed, but in the subsequent enhancement network, the reflection recovery network only uses the U-Net structure, and the design of the illumination adjustment network is relatively simple, without introducing more context information for the subsequent enhancement of the two components. The present invention extracts an attention map as a guide for the reflection recovery stage. The attention map can be used to detect the degree of underexposure of the image, so that it can pay more attention to the underexposed area in the reflectivity recovery stage, thereby better enhancing the component. In addition, since the noise in the underexposed area of ​​the low-light image is more serious than that in the normally exposed area, if the entire component is directly enhanced, the noise hidden in the dark area will be amplified. Therefore, the present invention formulates an attention map as a guide for reflectivity recovery, which can effectively suppress the noise in the enhancement process. In the illumination adjustment stage, in addition to the enhancement rate, the restored reflectivity is also used as input to introduce attention information, so that the illumination adjustment can focus on more contextual information and achieve better results. The quantitative results of the experiment also prove the effectiveness and superiority of the present invention.

[0100] Table 1 Quantitative measurement results on the LOL dataset

[0101]

[0102] 2. Qualitative detection experiment

[0103] Figure 3 , Figure 4 , Figure 5The comparison of the visual effects of the present invention and the other methods is shown. The experimental results show that the four methods of BIMEF, SRIE, CRM, and Dong are insufficient in improving the brightness of the image. Even if the image is enhanced, the problem of dim light still exists. Although the four methods of GLAD, LIME, MF, and NPE improve the brightness of the image to a certain extent, there are obvious noise points in the enhanced image, and the color distortion is relatively serious. The image enhanced by Retinex-Net is relatively fuzzy. This is because the network structure only designs a Relight-Net to re-estimate the illumination of the image, without considering the influence of the reflectivity component and the illumination component on the image respectively. The enhancement effect of KinD is relatively good. Not only has the brightness been well improved, but the noise is also well handled, and the image fidelity is also considerable, but the enhanced image still has the problem of low brightness. The present invention can not only effectively improve the brightness of low-light images, but also the noise and color distortion of the enhanced images are very small. This once again proves that the attention map can be used to detect the degree of underexposure of the image and use it as a guide for the subsequent enhancement network, which can make it pay more attention to the underexposed areas during the enhancement process and avoid over-enhancing the normally exposed areas, so that the enhanced image can achieve better results.

[0104] 3. Ablation Experiment

[0105] The effectiveness of each component in the present invention is tested through ablation experiments. The settings of ablation experiments are shown in Table 2. The quantitative ablation experiment results of PSNR and SSIM are shown in Table 3. The present invention is based on Retinex-Net, KinD is the baseline, w / o AG w / o JP is the result of retraining KinD under our conditions;

[0106] The quantitative test results show that compared with Retinex-Net, the overall performance of the present invention has been greatly improved, with PSNR increased by 5.38 and SSIM increased by 26.86%. For the baseline, the injection of joint prediction (w / o AG, with JP) increased PSNR by 0.80 and SSIM by 1.62%; the injection of attention information (with AG, w / o JP) increased PSNR by 0.97 and SSIM by 4.15%; the injection of attention information and joint prediction (ours) increased PSNR by 1.95 and SSIM by 5.25%.

[0107] Qualitative test results such as Figure 6As shown. For (a2), the shadow area of ​​the enhanced photo (b2) has obvious noise. The injection of joint prediction (w / o AG, with JP) eliminates the noise in the shadow area, but there is visible color distortion in the background (c2). This is because the joint prediction can provide more contextual information for Adjust-Net. The injection of attention information (with AG, w / o JP) solves the problem of background color distortion, but the noise in the shadow part still exists (d2). This is because the attention information can distinguish between areas with different exposures and successfully process these areas. The enhancement effect of the present invention is quite considerable, and the existing noise and color distortion problems are successfully solved (e2). For (a3), the enhanced result suffers from severe artifacts and degradation (b3). Even the injection of joint prediction cannot solve the artifacts and degradation problems in severely underexposed areas (c3). However, after the injection of attention information, these defects are eliminated (d3). It is proved again that attention information can distinguish between areas with different exposures, so that it pays more attention to these areas during the enhancement process, thereby achieving better enhancement effect.

[0108] Table 2 The settings of the ablation experiment of the present invention

[0109]

[0110] Table 3 Ablation experiment results of the present invention

[0111]

[0112]

[0113] In summary, the following effects are achieved:

[0114] 1. This paper proposes a technique called CA&R Net that combines the attention mechanism and the Retinex model to enhance low-light images, which is used to extract the attention map of the image and serve as a guide for subsequent enhancement. Under the guidance of the attention information, CA&R Net can successfully handle noise, color distortion, and various types of degradation.

[0115] 2. The present invention formulates an attention map and guides reflectance recovery in a region-adaptive manner so that more attention is paid to underexposed areas during the enhancement process and over-enhancement of normally exposed areas is avoided.

[0116] 3. The present invention uses the reconstructed reflectivity and low illumination for joint prediction. This joint prediction utilizes attention information and also introduces more context information to the light adjustment network, breaking the limitation of the single design of the light adjustment network and enabling better light adjustment results.

[0117] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0118] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the present invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A low-light image enhancement method combining attention mechanism and Retinex model, characterized by: The following steps are involved: Establish information extraction network Decomposition-Net, reflectivity restoration network Restore-Net and light adjustment network Adjust-Net; In the decomposition stage, Decomposition-Net is used to decompose the paired low-light and normal-light images into reflectivity and light , expressed as ; In the reflectivity recovery stage, Restore-Net is used to restore the reflectivity and suppress the hidden noise, and reflectivity recovery is achieved in a regional adaptive manner under the guidance of, which is expressed as ; In the illumination adjustment stage, Adjust-Net is used to adjust the illumination, and the reconstructed reflectance and low illumination are used to jointly predict the illumination layer of the image. This joint prediction utilizes attention information to achieve better illumination adjustment results, which is expressed as ; Create three subnets for Conduct modeling; The information extraction network comprises: Image decomposition network: used to extract reflectance and illumination from images simultaneously; Attention Network: Extracting the Attention Map of an Image Using Attention-Net To detect the degree of underexposure and use it to guide the reflectivity recovery network in the Restore-Net stage; The reflectivity recovery network comprises: The Residual in Residual module,RIR module uses the deep residual to provide a powerful pixel adjustment capability for the reflectivity restoration network, effectively suppressing noise and performing color correction; The Attention Residual in Residual module first uses two two-layer convolutions from Extract features from and , and then use the convolutional layer through the RIR module from Extract features from , and finally through Multiply first Plus The way to The three features are fused and then convolution is performed to obtain the output of the ARIR module. The loss function is expressed as follows: ， in is the weight coefficient, is the mean square error, is the structural similarity index, is the gradient loss, is the reflectivity; The light adjustment network first uses U-Net to Extract features, and then use two layers of convolution to enhance the ratio and Extract features from the network, fuse them with the features extracted by U-Net, and then perform convolution to get the adjusted illumination. Then use the Sigmoid layer to adjust the illumination of the network. is limited to [0,1]; finally, the restored reflectivity is used and adjusted lighting Reconstruct according to element-by-element multiplication, the final prediction result .

2. According to claim 1, a low-light image enhancement method combining an attention mechanism and a Retinex model is characterized in that: in, The image decomposition network includes a loss function for constraints; The loss function consists of reflectance similarity loss, illumination smoothness loss, mutual consistency loss, and reconstruction loss.

3. The low-light image enhancement method combining the attention mechanism and the Retinex model according to claim 2, characterized in that: The calculation formula of the reflectivity similarity loss is: ， in, represents the reflectivity similarity, Represents the 2-norm The calculation formula of the illumination smoothness loss is: ， in, Indicates the lighting smoothness, represents the first-order derivative operator, represents the 1 norm, represents the absolute value operator, Represents low-light images, represents a normal light image, is a positive constant; The calculation formula of the mutual consistency loss is: ， in, Indicates mutual consistency, For parameters The calculation formula of the reconstruction loss is: ， Decomposition-Net decomposes the input image into reflectance and illumination. When reconstructing these two components, the input image should also be obtained. We use To express; Therefore, the total loss function of the image decomposition network is expressed as: ， in, They are the weight coefficients of reflectance similarity loss, illumination smoothness loss, and mutual consistency loss respectively.

4. The low-light image enhancement method combining the attention mechanism and the Retinex model according to claim 1, characterized in that: The loss function of the illumination adjustment network stage is expressed as: ， in is the weight coefficient, For illumination.

Images