Low-light image enhancement method based on zero-reference Retinex decomposition network

Abstract

The invention belongs to the technical field of digital image processing, and discloses a low-light image enhancement method based on a zero-reference Retinex decomposition network, which comprises the following steps of: constructing a low-light image enhancement data set; building and training a low-light noise extraction network; building and training a low-light decomposition network, and estimating a reflection component and an illumination component; enhancing the low-light image; and testing the low light enhancement network. According to the low-light image enhancement method based on the zero-reference Retinex decomposition network, the illumination component, the reflection component and the noise component of the low-light image are learned through the deep learning network, the denoised low-light image is obtained, the illumination component after illumination adjustment is obtained through the gamma correction technology, and the low-light image enhancement method based on the zero-reference Retinex decomposition network is obtained through the Retinex decomposition theory. And obtaining a final low-light enhancement result. According to the method, training data is not needed, the low-light image is directly and rapidly enhanced, the image visual quality is improved, and the visual task precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital image processing, and in particular to a low-light image enhancement method based on a zero-reference Retinex decomposition network. Background Art

[0002] Due to the technical limitations of the shooting environment and equipment, the visual quality of images deteriorates under low-light conditions, resulting in low contrast, unclear details, severe artifacts, and a large amount of sensor noise. These problems not only affect human visual perception, but also hinder the execution of advanced computer vision tasks such as autonomous driving and human behavior recognition. Therefore, improving image brightness and visual quality in low-light scenes is a challenging image processing problem.

[0003] There are some problems with low-light image enhancement methods in the past few decades. First, for traditional low-light image enhancement methods, there are mainly histogram equalization and Retinex-based methods. The former will over-enhance or compress local details of the image, causing the enhanced image to look unnatural or distorted as a whole; the latter decouples the image into illumination and reflection components based on the Retinex model, and enhances the image by adjusting the exposure of the illumination component based on the invariance of the reflection component. However, this method relies too much on artificial priors to achieve satisfactory enhancement; at the same time, these traditional enhancement methods are limited by artificial priors and artificial regularization.

[0004] In recent years, deep learning has developed rapidly, and learning-based methods have also been widely used in the field of low-light image enhancement. Among them, deep learning methods based on Retinex theory have better enhancement performance in most cases due to their physical interpretability, but most deep Retinex methods denoise the image reflection component, resulting in blurred images after enhancement. In addition, in supervised deep learning methods, some people propose to transform the decomposition problem into an implicit prior regularization model. Although this method can effectively suppress noise and retain image details, it relies on paired data sets, which limits its further application. In contrast, enhancement methods based on unsupervised learning and methods based on zero-reference deep learning do not require the preparation of paired data sets, and have received widespread attention in the field of low-light image enhancement and are highly competitive. Most of the existing low-light image enhancement methods currently use manual prior denoising in the reflection component; however, since the reflection component contains a large amount of image detail information and there is some pseudo-noise, the enhancement results obtained by these schemes do not remove the original noise, and the image details appear blurred. At the same time, zero-reference learning methods rely on prior constraints of the network due to the lack of guidance from reference images and training data; using inappropriate priors and denoising methods can also lead to blurred details in enhanced images. Therefore, studying how to design a reasonable prior for zero-reference low-light image enhancement has important theoretical research significance and practical application value. Summary of the invention

[0005] The purpose of the present invention is to provide a low-light image enhancement method based on a zero-reference Retinex decomposition network, which utilizes a deep learning network to perform Retinex decomposition on the original low-light image, and extracts the original sensor noise of the low-light image to obtain the illumination component, reflection component and noise component of the low-light image. A noise-free low-light enhanced image is obtained by exposing the illumination component corrected and the reflection component with noise removed. Without the need for additional pairing data, the image quality can be quickly and effectively improved under various lighting conditions, the visual perception and application effect of the image can be enhanced, and the image can be widely used in various subsequent advanced computer vision tasks.

[0006] To achieve the above object, the present invention provides a low-light image enhancement method based on a zero-reference Retinex decomposition network, comprising the following steps: Step S1, constructing a low-light image enhancement dataset based on the zero-reference Retinex decomposition network ZERRIN-Net; Step S2, building and training a low-noise optical noise extraction network; Step S3: building and training a low-light decomposition network to estimate the reflection component and the illumination component; Step S4, enhancing the low-light image, wherein the enhancement process is to adjust the illumination of the illumination component and generate a high-quality enhanced image; Step S5: Test the ZERRIN-Net low-light enhancement network.

[0007] Preferably, in step S1, a ZERRIN-Net low-light image enhancement dataset is constructed, and the specific process is as follows: Step S11, selecting low-light images of various scenes from an existing public low-light dataset; Step S12: collect a real low-light data set, remove some blurred images, and uniformly adjust the image size to (512, 512); the data set contains 5 low-light grayscale images and 20 low-light color images; Step S13: merge the above-selected public low-light dataset images and the real low-light images to obtain the ZERRIN-Net low-light image enhancement dataset.

[0008] Preferably, in step S2, building and training a low-noise optical noise extraction network comprises the following steps: Step S21, building a noise extraction network, which consists of multiple layers of convolution, including five CBPU layers and a convolution layer with a convolution kernel size of 3×3; wherein each CBPU layer consists of a convolution layer with a convolution kernel size of 3×3, a BN layer and a PRelu layer; Step S22: The noise extraction network is used to extract the original noise of the low-light image. First, the low-light image in the ZERRIN-Net low-light image enhancement dataset in step S1 is recorded as ,Will As input to the noise extraction network; Step S23, adjust the image size of the low-light image, adjust the image size of the ZERRIN-Net low-light image enhancement dataset to (512, 512); extract noise from the input image, and the output image is ;in, The noise estimated by the noise extraction network is the original noise of the low-light image; Step S24: denoise the low-light image. Subtract the low-light image from the original noise of the low-light image as input to obtain a denoised low-light image. The denoising formula is as follows: (1); in, For low-light images, is the original noise of the low-light image, This is a denoised low-light image.

[0009] Preferably, in step S3, building and training a low-light decomposition network comprises the following steps: Step S31, preprocessing the low-light image data in the ZERRIN-Net low-light image enhancement dataset constructed in step S1; First, the input low-light image is converted into HSV channels, and then the HSV channels are decoupled; represents the brightness of the image, Represents the hue of the image, represents the saturation of the image; finally, the original low-light image and The channels are fused, and the fused image is used as the input of the low-light decomposition network; Step S32, build a low-light decomposition network, which consists of an illumination estimation network and a reflection estimation network; wherein the illumination estimation network includes five CVPU layers and a convolution layer with a convolution kernel size of 1×1; each CVPU layer consists of a convolution layer and a PRelu layer; the reflection estimation network includes multiple CVPU layers and a convolution layer with a convolution kernel size of 3×3; the activation layers of the illumination estimation network and the reflection estimation network both use a sigmoid activation function; Obtain the illumination component and reflection component of the low-light image through the low-light decomposition network; Step S33: Design the total loss function of the ZERRIN-Net network Guide the network to enhance low-light images as follows: (2); in, For reconstruction loss; To smooth the loss; Enhance loss for texture; is the image noise loss.

[0010] Preferably, in step S33, the total loss function of the ZERRIN-Net network includes reconstruction loss, smoothing loss, texture enhancement loss and image noise loss, as follows: Step S331: Reconstruction loss : According to the Retinex principle, the product of the reflection component and the illumination component, plus the distance between the feature map of the noise component and the original low-light image in a certain feature space is calculated and expressed using the L1 distance, as shown below: (3); in, is the reconstruction loss, is the illumination component, is the reflection component; Step S332: Smoothing loss : Smooth the illumination component and the reflection component separately, as shown below: (4); in, To smooth the loss; is the illumination smoothing loss coefficient; is the reflection smoothing loss coefficient; is the piecewise smooth illumination loss; is a piecewise smooth reflection loss; the piecewise smooth illumination loss restores the detail information of the dark area, and the L1 loss is used to calculate the difference between the illumination component and the maximum value of the original low-light image, as shown below: (5); in, is the maximum channel of the original low-light image, are the horizontal and vertical gradients, is the weight function; The piecewise smooth reflection loss restores the image structure and detail information, expressed as an illumination loss, as shown below: (6); in, is the loss coefficient of the reflected component, is the weight parameter of the reflected component gradient; Step S333: Texture enhancement loss : Design texture enhancement loss to suppress the blurring of dark area details and color distortion in the process of low-light image reconstruction, as shown below: (7); in, represents color loss, Represents the loss in low light areas, stands for low-light attention loss; The color loss is designed to avoid color distortion in the enhanced image as follows: (8); in, For color loss; represents a pair of channels; Indicates the enhanced image The average intensity value of the channel; Indicates the enhanced image The average intensity value of the channel; Adding low light area loss improves the quality of low light areas, as shown below: (9); in, Loss for low light areas; for the low-light areas of the reflected image; is the low-light area of ​​the original low-light image, is the rest of the reflected image, for the rest of the original low-light image; is the smoothing weight for low-light areas; is the smoothing weight for low-light areas; and Pixels for low-light areas; and are pixels in the highlight area; On the basis of the original area loss, low-light attention loss is added to avoid over-enhancement in image restoration, as shown below: (10); in, Represents the maximum value among the three color channels; Step S334: Image noise loss :The main design is based on the noise extraction module, using the illumination component of the image as the weight to extract low-light image noise, as shown below: (11); in, represents the Frobenius norm of the matrix X.

[0011] Preferably, in step S4, the low-light image is enhanced, wherein the enhancement process is to adjust the illumination of the illumination component and generate a high-quality enhanced image, and the specific process is as follows: Step S41: The single-channel illumination component obtained in step S32 Use cat to fuse into three channels, then perform gamma transformation to enhance the brightness and obtain the gamma transformation illumination component , as shown below: (12); in, is the coefficient of gamma transform; Step S42: subtract the noise component from the original low-light image, thereby eliminating the sensor noise of the image and obtaining a noise-free low-light image, as shown below: (13); in, For low-light images, is the original noise of the low-light image, This is the low-light image after denoising; Step S43: In order to suppress noise amplification, the noise-free low-light image Divide it by the three-channel illumination component to get the noise-free reflection component , as shown below: (14); Step S44: The illumination component corrected in step S41 The de-noised reflection component obtained in step S43 Combined to generate the final enhanced image , as shown below: (15); in, The generated enhanced image.

[0012] Preferably, in step S5, the low-light enhancement network is tested, comprising the following steps: Step S51: low light image As the input of the noise extraction network, it is converted into HSV channels and extracted Channels and low-light images The components are fused and sent to the decomposition network for decomposition to obtain reflection component, illumination component and noise component. Step S52, obtaining the illumination component and the reflection component of the low-light image through the decomposition network, and obtaining the original noise of the low-light image through the noise extraction network, and obtaining the denoised reflection component, and finally multiplying the gamma-corrected illumination component and the denoised reflection component to obtain the final enhanced image; Step S53, visualizing the detected result to obtain an enhanced image; Step S54: Apply the algorithm to high-level visual tasks such as grayscale image, face detection, object detection and instance segmentation to verify the generalization performance of the proposed method.

[0013] Therefore, the present invention adopts the above-mentioned low-light image enhancement method based on the zero-reference Retinex decomposition network, uses a deep learning network to learn the illumination component, reflection component and noise component of the low-light image, obtains the denoised low-light image, and uses gamma correction technology to obtain the illumination component after illumination adjustment, and uses Retinex decomposition theory to obtain the final low-light enhancement result; this method does not require training data, can directly and quickly enhance the low-light image, improve the visual quality of the image, and improve the accuracy of visual tasks.

[0014] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a workflow diagram of a low-light image enhancement method based on a zero-reference Retinex decomposition network in the present invention; Figure 2 : is a diagram of the N-Net network structure in an embodiment of the present invention; Figure 3This is a diagram of the RI-Net network structure in an embodiment of the present invention; Figure 4 is a result diagram of the low-light image enhancement network in an embodiment of the present invention; wherein (a) is the original image of the real scene; (b) is the result diagram after image enhancement; Figure 5 1 is a comparison diagram showing the results of face detection using the enhancement method in an embodiment of the present invention; wherein (a) is an original image of a real scene; and (b) is a result image after image enhancement. Figure 6 1 is a comparison diagram showing the results of the enhancement method in target detection in an embodiment of the present invention; wherein (a) is the original image of the real scene; (b) is the result image after image enhancement; Figure 7 1 is a comparison diagram showing the results of the enhancement method in grayscale images in an embodiment of the present invention; wherein (a) is the original image of the real scene; (b) is the result image after image enhancement; Figure 8 1 is a comparison diagram showing the results of the enhancement method in the instance segmentation task in an embodiment of the present invention; wherein (a) is the original image of the real scene; and (b) is the result image after image enhancement. DETAILED DESCRIPTION

[0016] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0017] like Figure 1 As shown, the present invention adopts the above-mentioned low-light image enhancement method based on the zero-reference Retinex decomposition network, comprising the following steps: Step S1, constructing a low-light image enhancement dataset based on a zero-reference Retinex decomposition network (ZERRIN-Net); Step S2, building and training a low-noise optical noise extraction network N-Net; Step S3, building and training a low-light decomposition network RI-Net to estimate the reflection component and the illumination component; Step S4, enhancing the low-light image, wherein the enhancement process is to adjust the illumination of the illumination component and generate a high-quality enhanced image; Step S5: Test the ZERRIN-Net low-light enhancement network.

[0018] Example Step S1: construct a low-light image enhancement dataset based on the zero-reference Retinex decomposition network ZERRIN-Net.

[0019] Step S11: Select low-light images of various scenes from existing public low-light datasets such as DICM, ExDark, Fusion, LIME, MEF, NPE and VV.

[0020] Step S12: Collect a real low-light dataset, use the Huawei P30Pro camera to collect low-light images of various scenes, remove some blurred images, and uniformly adjust the image size to (512, 512); the dataset contains 5 low-light grayscale images and 20 low-light color images. The collection of the above dataset is shown in Table 1.

[0021] Table 1 Dataset collection ;

[0022] Step S13: merge the above-selected public low-light dataset images and the real low-light images to obtain the ZERRIN-Net low-light image enhancement dataset.

[0023] Step S2: Build and train a low-noise optical noise extraction network N-Net.

[0024] Step S21: Figure 2 As shown in Figure 2, a noise extraction network N-Net is built, which consists of multiple layers of convolution, including five CBPU layers and a convolution layer with a convolution kernel size of 3×3; each CBPU layer consists of a convolution layer with a convolution kernel size of 3×3, a BN layer and a PRelu layer. The detailed structure of the noise extraction module N-Net network is shown in Table 2.

[0025] Table 2 Detailed structure of the noise extraction module N-Net network ;

[0026] Step S22, the noise extraction network N-Net is used to extract the original noise of the low-light image. First, the low-light image in the ZERRIN-Net low-light image enhancement dataset in step S1 is recorded as ,Will As input to the noise extraction network.

[0027] Step S23, then, adjust the image size of the low-light image, and adjust the image size of the ZERRIN-Net low-light image enhancement dataset to (512, 512); extract noise from the input image, and the output image is ;in, The noise estimated by the noise extraction network is the original noise of the low-light image.

[0028] Step S24: denoise the low-light image. Subtract the low-light image from the original noise of the low-light image as input to obtain a denoised low-light image. The denoising formula is as follows: (1); in, For low-light images, is the original noise of the low-light image, This is a denoised low-light image.

[0029] Step S3: Build and train the low-light decomposition network RI-Net to estimate the reflection component and illumination component.

[0030] Step S31, preprocessing the low-light image data in the ZERRIN-Net low-light image enhancement dataset constructed in step S1.

[0031] First, the input low-light image is converted into HSV channels, and then the HSV channels are decoupled; represents the brightness of the image, Represents the hue of the image, Represents the saturation of the image. The channel contains the detailed information of the image, which combines the original low-light image and The channels are fused to enhance the details of the decomposed reflection component and illumination component, and the fused image is used as the input of the low-light decomposition network RI-Net.

[0032] Step S32: Figure 3 As shown in the figure, a low-light decomposition network RI-Net is built, which consists of an illumination estimation network I-Net and a reflection estimation network R-Net. Among them, I-Net contains five CVPU layers and a convolution layer with a convolution kernel size of 1×1; each CVPU layer consists of a convolution layer and a PRelu layer. The R-Net network contains multiple CVPU layers and a convolution layer with a convolution kernel size of 3×3. The R-Net network is deeper than the I-Net network and contains more information; the R-Net network uses the same sigmoid activation function as the I-Net activation layer.

[0033] Obtaining the illumination component of low-light images through the low-light decomposition network RI-Net and the reflected component .

[0034] Table 3 Decomposition module RI-Net network structure ;

[0035] Step S33: In order to better guide the network to enhance low-light images, the total loss function of the ZERRIN-Net network is designed , as shown below: (2); in, For reconstruction loss; To smooth the loss; Enhance loss for texture; is the image noise loss.

[0036] Step S331: Reconstruction loss : According to the Retinex principle, the product of the reflection component and the illumination component, plus the distance between the feature map of the noise component and the original low-light image in a certain feature space is calculated and expressed using the L1 distance, as shown below: (3); in, is the reconstruction loss, is the illumination component, is the reflection component.

[0037] Step S332: Smoothing loss : Smooth the illumination component and the reflection component separately, as shown below: (4); in, To smooth the loss; is the illumination smoothing loss coefficient; is the reflection smoothing loss coefficient; is the piecewise smooth illumination loss; is the piecewise smooth reflection loss. The piecewise smooth illumination loss amplifies the contrast of adjacent pixels at the same illumination value when they have similar intensities, so the detail information of dark areas can be better restored. The L1 loss is used to calculate the difference between the illumination component and the maximum value of the original low-light image, as shown below: (5); in, is the maximum channel of the original low-light image, are the horizontal and vertical gradients, is the weight function.

[0038] Similarly, in order to restore image structure and detail information, the reflection loss is piecewise smoothed A similar representation is used as the illumination loss, but with a greater focus on points with small horizontal and vertical gradients, as shown below: (6); in, is the loss coefficient of the reflected component, is the weight parameter of the reflected component gradient.

[0039] Step S333: Texture enhancement loss : In order to further suppress the problem of blurred dark area details and color distortion in the process of low-light image reconstruction, a texture enhancement loss is designed as follows: (7); in, represents color loss, Represents the loss in low light areas, Stands for low-light attention loss.

[0040] To avoid color distortion in enhanced images, a color loss , as shown below: (8); in, For color loss; represents a pair of channels; Indicates the enhanced image The average intensity value of the channel; Indicates the enhanced image The average intensity value of the channel.

[0041] At the same time, in order to improve the quality of low-light areas, low-light area loss is added , as shown below: (9); in, Loss for low light areas; for the low-light areas of the reflected image; is the low-light area of ​​the original low-light image, is the rest of the reflected image, for the rest of the original low-light image; is the smoothing weight for low-light areas; is the smoothing weight for low-light areas; and Pixels for low-light areas; and is the pixel in the highlight area. This method recommends coefficient and Set to 4 and 1 respectively.

[0042] In order to avoid over-enhancement, low-light attention loss is added to the original area loss. This loss limits the value of the reflective component, allowing to correctly enhance underexposed areas and avoid overenhancing normally exposed areas, resulting in low-light attention loss. , as shown below: (10); in, Represents the maximum value among the three color channels.

[0043] Step S334: Image noise loss :The main design is based on the noise extraction module, using the illumination component of the image as the weight to extract low-light image noise, as shown below: (11); in, represents the Frobenius norm of the matrix X.

[0044] Step S4: enhance the low-light image, wherein the enhancement process is to adjust the illumination of the illumination component and generate a high-quality enhanced image.

[0045] Step S41: The single-channel illumination component obtained in step S32 Use cat to fuse into three channels, then perform gamma transformation to enhance the brightness and obtain the gamma transformation illumination component , as shown below: (12); in, are the gamma transform coefficients.

[0046] Step S42: subtract the noise component extracted in step S22 from the original low-light image, thereby eliminating the sensor noise of the image and obtaining a noise-free low-light image, as shown below: (13); in, For low-light images, is the original noise of the low-light image, This is a denoised low-light image.

[0047] Step S43: In order to suppress noise amplification, the noise-free low-light image Divide it by the three-channel illumination component to get the noise-free reflection component , as shown below: (14); Step S44: The illumination component corrected in step S41 The de-noised reflection component obtained in step S43 Combined to generate the final enhanced image , as shown below: (15); in, The generated enhanced image.

[0048] Step S5: testing the low-light enhancement network.

[0049] Step S51: The low-light image in step S1 As the input of the noise extraction network, it is converted into HSV channels and extracted Channels and low-light images The fusion is then sent to the RI-Net decomposition network for decomposition to obtain the reflection component, illumination component, and noise component.

[0050] Step S52, obtaining the illumination component and the reflection component of the low-light image through the decomposition network, and obtaining the original noise of the low-light image through the noise extraction network, and obtaining the denoised reflection component, and finally multiplying the gamma-corrected illumination component and the denoised reflection component to obtain the final enhanced image; Step S53: Visualize the detected results to obtain an enhanced image, such as Figure 4 As shown; Step S54: Apply the algorithm to face detection. Figure 5 As shown, target detection is Figure 6 As shown, the grayscale image is Figure 7 As shown and instance segmentation as Figure 8 The generalization performance of the proposed method is verified on high-level visual tasks such as those shown in the figure.

[0051] Therefore, the present invention adopts the above-mentioned low-light image enhancement method based on the zero-reference Retinex decomposition network, utilizes the deep learning network to perform Retinex decomposition on the original low-light image, and extracts the original sensor noise of the low-light image at the same time, to obtain the illumination component, reflection component and noise component of the low-light image, and obtains the noise-free low-light enhanced image through the exposure-corrected illumination component and the noise-removed reflection component. Without the need for additional pairing data, the image quality can be improved quickly and effectively under various lighting conditions, the visual perception and application effect of the image can be improved, and it is widely used in various subsequent advanced computer vision tasks.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A low-light image enhancement method based on a zero-reference Retinex decomposition network, characterized in that: The following steps are involved: Step S1: Construct a low-light image enhancement dataset based on the zero-reference Retinex decomposition network ZERRIN-Net. The specific process is as follows: Step S11, selecting low-light images of various scenes from an existing public low-light dataset; Step S12: collect a real low-light data set, remove some blurred images, and uniformly adjust the image size to (512, 512); the data set contains 5 low-light grayscale images and 20 low-light color images; Step S13, merging the above-selected public low-light dataset image and the real low-light image to obtain the ZERRIN-Net low-light image enhancement dataset; Step S2: Building and training a low-noise optical noise extraction network, including the following steps: Step S21, building a noise extraction network, which consists of multiple layers of convolution, including five CBPU layers and a convolution layer with a convolution kernel size of 3×3; wherein each CBPU layer consists of a convolution layer with a convolution kernel size of 3×3, a BN layer and a PRelu layer; Step S22: The noise extraction network is used to extract the original noise of the low-light image. First, the low-light image in the ZERRIN-Net low-light image enhancement dataset in step S1 is recorded as ,Will As input to the noise extraction network; Step S23, adjust the image size of the low-light image, adjust the image size of the ZERRIN-Net low-light image enhancement dataset to (512, 512); extract noise from the input image, and the output image is ;in, The noise estimated by the noise extraction network is the original noise of the low-light image; Step S24: denoise the low-light image. Subtract the low-light image from the original noise of the low-light image as input to obtain a denoised low-light image. The denoising formula is as follows: （1）； in, For low-light images, is the original noise of the low-light image, This is the low-light image after denoising; Step S3: Building and training a low-light decomposition network to estimate the reflection component and the illumination component, including the following steps: Step S31, preprocessing the low-light image data in the ZERRIN-Net low-light image enhancement dataset constructed in step S1; First, the input low-light image is converted into HSV channels, and then the HSV channels are decoupled; represents the brightness of the image, Represents the hue of the image, represents the saturation of the image; finally, the original low-light image and The channels are fused, and the fused image is used as the input of the low-light decomposition network; Step S32, build a low-light decomposition network, which consists of an illumination estimation network and a reflection estimation network; wherein the illumination estimation network includes five CVPU layers and a convolution layer with a convolution kernel size of 1×1; each CVPU layer consists of a convolution layer and a PRelu layer; the reflection estimation network includes multiple CVPU layers and a convolution layer with a convolution kernel size of 3×3; the activation layers of the illumination estimation network and the reflection estimation network both use a sigmoid activation function; Obtain the illumination component and reflection component of the low-light image through the low-light decomposition network; Step S33: Design the total loss function of the ZERRIN-Net network Guide the network to enhance low-light images as follows: （2）； in, For reconstruction loss; To smooth the loss; Enhance loss for texture; is the image noise loss; Step S4: enhance the low-light image, wherein the enhancement process is to adjust the illumination of the illumination component and generate a high-quality enhanced image. The specific process is as follows: Step S41: The single-channel illumination component obtained in step S32 Use cat to fuse into three channels, then perform gamma transformation to enhance the brightness and obtain the gamma transformation illumination component , as shown below: （12）； in, is the coefficient of gamma transform; Step S42: subtract the noise component from the original low-light image, thereby eliminating the sensor noise of the image and obtaining a noise-free low-light image, as shown below: （13）； in, For low-light images, is the original noise of the low-light image, This is the low-light image after denoising; Step S43: In order to suppress noise amplification, the noise-free low-light image Divide it by the three-channel illumination component to get the noise-free reflection component , as shown below: （14）； Step S44: The illumination component corrected in step S41 The denoised reflection component obtained in step S43 Combined to generate the final enhanced image , as shown below: （15）； in, For the generated enhanced image; Step S5, testing the ZERRIN-Net low-light enhancement network, including the following steps: Step S51: low light image As the input of the noise extraction network, it is converted into HSV channels and extracted Channels and low-light images The components are fused and sent to the decomposition network for decomposition to obtain reflection component, illumination component and noise component. Step S52, obtaining the illumination component and the reflection component of the low-light image through the decomposition network, and obtaining the original noise of the low-light image through the noise extraction network, and obtaining the denoised reflection component, and finally multiplying the gamma-corrected illumination component and the denoised reflection component to obtain the final enhanced image; Step S53, visualizing the detected result to obtain an enhanced image; Step S54: Apply the algorithm to high-level visual tasks such as grayscale image, face detection, object detection and instance segmentation to verify the generalization performance of the proposed method.

2. The low-light image enhancement method based on zero-reference Retinex decomposition network according to claim 1, characterized in that: In step S33, the total loss function of the ZERRIN-Net network includes reconstruction loss, smoothness loss, texture enhancement loss and image noise loss, as follows: Step S331: Reconstruction loss : According to the Retinex principle, the product of the reflection component and the illumination component, plus the distance between the feature map of the noise component and the original low-light image in a certain feature space is calculated and expressed using the L1 distance, as shown below: （3）； in, is the reconstruction loss, is the illumination component, is the reflection component; Step S332: Smoothing loss : Smooth the illumination component and the reflection component separately, as shown below: （4）； in, To smooth the loss; is the illumination smoothing loss coefficient; is the reflection smoothing loss coefficient; is the piecewise smooth illumination loss; is a piecewise smooth reflection loss; the piecewise smooth illumination loss restores the detail information of the dark area, and the L1 loss is used to calculate the difference between the illumination component and the maximum value of the original low-light image, as shown below: （5）； in, is the maximum channel of the original low-light image, are the horizontal and vertical gradients, is the weight function; The piecewise smooth reflection loss restores the image structure and detail information, expressed as an illumination loss, as shown below: （6）； in, is the loss coefficient of the reflected component, is the weight parameter of the reflected component gradient; Step S333: Texture enhancement loss : Design texture enhancement loss to suppress the blurring of dark area details and color distortion in the process of low-light image reconstruction, as shown below: （7）； in, represents color loss, Represents the loss in low light areas, stands for low-light attention loss; The color loss is designed to avoid color distortion in the enhanced image as follows: （8）； in, For color loss; represents a pair of channels; Indicates the enhanced image The average intensity value of the channel; Indicates the enhanced image The average intensity value of the channel; Adding low light area loss improves the quality of low light areas, as shown below: （9）； in, Loss for low light areas; for the low-light areas of the reflected image; is the low-light area of ​​the original low-light image, is the rest of the reflected image, for the rest of the original low-light image; is the smoothing weight for low-light areas; is the smoothing weight for low-light areas; and Pixels for low-light areas; and are pixels in the highlight area; On the basis of the original area loss, low-light attention loss is added to avoid over-enhancement in image restoration, as shown below: （10）； in, Represents the maximum value among the three color channels; Step S334: Image noise loss :The main design is based on the noise extraction module, using the illumination component of the image as the weight to extract low-light image noise, as shown below: （11）； in, represents the Frobenius norm of the matrix X.

Images