A low-light image unsupervised training method based on binocular cameras

Abstract

The application discloses a kind of low-light image unsupervised training methods based on binocular camera, comprising the following steps: by binocular camera, the low-light image pair of real world is captured synchronously, and data set is constructed;Lightweight optimization network is built and used, and optimization image pair is generated;Based on retina theory, build double branch network;Build feature extraction network TB-FEN, and carry out feature consistency constraint;Total loss function is constructed, and network training is guided.The beneficial effects of the application: solve the problem that high-quality paired data set is more difficult, monocular camera can only obtain image information from single view.The method is by jointly optimizing Retinex decomposition model and feature consistency constraint strategy, while improving image brightness and detail definition, maintaining original three-dimensional geometric structure information, suitable for image enhancement and reconstruction task in low-illumination scene.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and image processing, and in particular relates to an unsupervised training method for low-light images based on a binocular camera. Background Art

[0002] Low-light image enhancement technology has broad application value in scenarios such as video surveillance, autonomous driving, medical imaging, and robotic perception. In recent years, deep learning-based image enhancement methods have made significant progress, and these methods can be roughly categorized as supervised and unsupervised. Supervised methods rely on a large number of paired low-light and normal-light images as training data. While they can achieve good results under ideal conditions, obtaining a realistic and rich paired dataset presents significant challenges in practical applications.

[0003] To reduce reliance on labeled data, unsupervised image enhancement methods are gaining attention. Retinex is a classic image decomposition model that posits that an image can be represented as the pixel-by-pixel product of illumination and reflectance. In low-light image enhancement, this theory enhances the illumination component to improve image brightness while preserving the reflectance component to preserve image structure.

[0004] While most current methods utilize retinex-based structural modeling, they still generally rely on monocular images. Due to the limited information captured by monocular cameras, it is difficult to accurately recover geometric structures in complex scenes. Especially in low-light conditions, image noise, contrast degradation, and texture loss further reduce the accuracy of depth estimation and image restoration. Summary of the Invention

[0005] To address the shortcomings of the aforementioned technologies, this paper aims to provide an unsupervised training method for low-light images based on a binocular camera. This method addresses the difficulties in obtaining high-quality paired datasets and the fact that monocular cameras can only capture image information from a single perspective. By jointly optimizing a Retinex decomposition model and a feature consistency constraint strategy, this method improves image brightness and detail clarity while preserving the original 3D geometric structure, making it suitable for image enhancement and reconstruction tasks in low-light scenarios.

[0006] To achieve the above object, the present invention adopts the following technical solution: an unsupervised training method for low-light images based on a binocular camera, comprising the following steps:

[0007] Step S1, using a binocular camera to synchronously capture a pair of real-world low-light images, remove blurry images, and construct a dataset;

[0008] Step S2: Build and use a lightweight optimization network to remove sensor noise and compression artifacts from the dataset in step S1 to generate optimized image pairs.

[0009] Step S3: Building a two-branch network based on retinal theory, wherein the left branch network decomposes the left image of the optimized image pair into a left reflected image and a left illuminated image, and the right branch network decomposes the right image of the optimized image pair into a right reflected image and a right illuminated image;

[0010] Step S4: Building a feature extraction network TB-FEN to extract features of the left image of the optimized image pair and the left reflected image in step S3, as well as the right image of the optimized image pair and the right reflected image in step S3, and perform feature consistency constraints;

[0011] In step S5, a total loss function is constructed based on the feature consistency constraints in steps S2, S3, and S4 to guide network training.

[0012] Furthermore, in step S2, the lightweight optimization network constructed is specifically:

[0013] The constructed lightweight optimization network consists of 4 convolutional activation layers and 1 convolutional layer with a convolution kernel size of 3x3. Each convolutional activation layer contains a convolution layer with a convolution kernel size of 3x3 and a ReLU activation function. The final output channel number of the convolution layer with a convolution kernel size of 3x3 is 3, which is used to generate the optimized image.

[0014] The left and right images of the low-light image pair are passed through a lightweight optimization network to obtain the optimized image, see the expression;

[0015] ;

[0016] The projection consistency loss function is: ;

[0017] Among them, I L is the left image of the low-light image pair, I R is the right image of the low-light image pair; i L To optimize the left image of the image pair, i R To optimize the right image of the image pair, To optimize the network, represents the square of the Euclidean norm, is the value of the projection consistency loss function.

[0018] Furthermore, in step S3, a two-branch network is built based on retinal theory, wherein the left branch network decomposes the left image of the optimized image pair to obtain a left reflected image and a left illuminated image, and the right branch network decomposes the right image of the optimized image pair to obtain a right reflected image and a right illuminated image; specifically, as follows:

[0019] Step S31: A dual-branch network processes the left image and the right image of the optimized image pair respectively. The left branch and the right branch of the dual-branch network are symmetrically arranged and have the same network structure, consisting of a reflection decomposition network and an illumination decomposition network.

[0020] Step S32: extracting the reflection image using a reflection decomposition network, which consists of five convolutional activation layers. The first four convolutional activation layers include a convolutional layer with a convolution kernel size of 3x3 and a ReLU activation function. The last convolutional activation layer includes a convolutional layer with a convolution kernel size of 3x3 and a Sigmoid activation function. The number of output channels is 3.

[0021] The illumination decomposition network extracts the illumination image and consists of 5 convolutional activation layers. The first 4 convolutional activation layers include a convolution layer with a convolution kernel size of 3x3 and a ReLU activation function. The last convolutional activation layer includes a convolution layer with a convolution kernel size of 3x3 and a Sigmoid activation function. The number of output channels is 1.

[0022] Step S33, the image decomposition process is expressed as:

[0023] ;

[0024] in, are the left and right images of the optimized image pair, and Represent the left and right images of the reflected image and the left and right images of the illuminated image, Represents the reflection decomposition network, represents the illumination decomposition network;

[0025] Step S34, introduce the illumination smoothing loss function based on total variation regularization, expressed as:

[0026] ;

[0027] in, , Represents the gradient of pixel (i, j) on the horizontal x-axis, , Represents the gradient of pixel (i, j) on the vertical y-axis, Indicates summation, Indicates the absolute value, Represents the value of the illumination smoothness loss function;

[0028] Step S35, introduce reflection reconstruction error loss, expressed as:

[0029] ;

[0030] in represents the value of the reflection reconstruction error loss function, represents the square of the L2 norm, Represents dot product.

[0031] Furthermore, in step S4: a feature extraction network TB-FEN is constructed to extract features of the left image of the optimized image pair and the left reflected image in step S3, and the right image of the optimized image pair and the right reflected image in step S3, respectively, and perform feature consistency constraints; specifically:

[0032] Step S41, the feature extraction network TB-FEN extracts the left and right images of the optimized image pair respectively Left and right images with reflected image Extract image semantic information at different levels and achieve feature alignment;

[0033] Step S42: the feature extraction network TB-FEN extracts the left and right images of the input optimized image pair. Left and right images with reflected image A parameter-sharing weight encoder is used for feature extraction to ensure the consistency of extracted features;

[0034] Step S43: The parameter-shared weight encoder includes three levels of feature extraction modules, corresponding to a shallow feature extraction module, a middle feature extraction module, and a deep feature extraction module, respectively. Each level of feature extraction module is composed of a residual connection module.

[0035] The residual connection module consists of a convolution layer with a convolution kernel size of 3x3, a batch normalization layer, and a ReLU activation function. The input and output are added together to form a residual connection. Each residual connection module uses the same network structure configuration.

[0036] Step S44, shallow feature extraction module: shallow feature consistency loss, defined as:

[0037] ;

[0038] in, Represents the L1 norm, which is used to measure the absolute difference between features. represents the shallow feature extraction module, Represents shallow feature consistency loss;

[0039] The middle-level feature extraction module measures the similarity between the optimized image and the reflected image in terms of local structural distribution: the middle-level feature consistency loss is calculated based on the structural similarity index and is defined as:

[0040] ;

[0041] in, represents the structural similarity index, represents the middle-level feature extraction module, represents the consistency loss of mid-level features;

[0042] Deep feature extraction module: The loss function of the deep feature extraction module is based on cosine similarity and is defined as:

[0043] ;

[0044] in, represents the deep feature extraction module, represents the deep feature consistency loss;

[0045] In step S45, a weighted total loss function is used to optimize the feature alignment effect of multiple levels. The total feature consistency loss function is defined as follows:

[0046] ;

[0047] in, is the weight hyperparameter of the shallow, middle and deep loss terms, which is used to control the contribution ratio of different level features to the total loss. represents the total feature consistency loss function.

[0048] Furthermore, in step S5, a total loss function is constructed to guide network training; specifically:

[0049] The total loss function is: .

[0050] The present invention adopts another technical solution: a low-light image enhancement method based on a binocular camera, which applies the unsupervised training method for low-light images based on a binocular camera, including the following steps:

[0051] Step T1: synchronously capture the left image I of the low-light image pair through the binocular camera L Right image I of the low-light image pair R ;

[0052] Step T2, loading the parameters of the lightweight optimization network and the dual-branch network trained in steps S2 and S3;

[0053] Step T3: the left image I of the image pair L and right image I R Pass the lightweight optimization network to get the left optimized image i of the optimized image pair L , the right optimized image i of the optimized image pair R , the left branch network optimizes the left optimized image i of the image pair through the reflection decomposition network and the illumination decomposition networkL Decomposed into left reflected image R L and the left illumination image L L ,

[0054] The right branch network optimizes the right optimized image i of the image pair through the reflection decomposition network and the illumination decomposition network R Decomposed into right reflected image R R and the right illumination image L R ;

[0055] Step T4: the left and right images L of the illumination image x Enhance the image and fuse the enhanced illumination image with the reflection image to obtain the enhanced illumination image pair. .

[0056] Furthermore, in step T4, the left and right images L of the illumination image are x Enhance the image and fuse the enhanced illumination image with the reflection image to obtain the enhanced illumination image pair. Specifically:

[0057] Step T41: the left and right images L of the illumination image x , the enhanced illumination image L is obtained by gamma transformation Hx ,formula: ;

[0058] in, is the left and right image L of the illumination image x Perform gamma transformation;

[0059] Step T42: Enhance the illumination image L Hx Combined with the reflected image, we get the image pair after illumination enhancement, formula: L Hx ;

[0060] Among them, H x is the image pair after illumination enhancement, Represents the left and right images of the illumination image.

[0061] Compared with the prior art, the advantages of the present invention are:

[0062] (1) The image enhancement method proposed in this paper does not rely on paired low-light and normal-light image data during training, which solves the problem of difficulty in obtaining high-quality paired data in supervised methods and significantly improves the adaptability and practicality of the method.

[0063] (2) The present invention uses a binocular camera as input and utilizes the structural redundant information between the left and right views for image enhancement and feature alignment, effectively maintaining the three-dimensional geometric consistency of the original scene and avoiding the common structural distortion problem in the monocular image enhancement process;

[0064] (3) In the image decomposition stage, the illumination and reflection modeling based on Retinex theory is introduced, and the consistency constraint functions of shallow, middle and deep features are designed. The optimized image and the reflected image are aligned from the details, structure to semantic level, which significantly improves the realism and visual quality of the enhanced image. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0066] Figure 1 This is a flow chart of the unsupervised training method for low-light images based on a binocular camera of the present invention;

[0067] Figure 2 This is a flow chart of the low-light image enhancement method based on a binocular camera of the present invention. DETAILED DESCRIPTION

[0068] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0069] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0070] like Figure 1 As shown in FIG, a low-light image unsupervised training method based on a binocular camera includes the following steps:

[0071] Step S1, using a binocular camera to synchronously capture a pair of real-world low-light images, remove blurry images, and construct a dataset;

[0072] Step S2: Build and use a lightweight optimization network to remove sensor noise and compression artifacts from the dataset in step S1 to generate optimized image pairs.

[0073] Step S3: Building a two-branch network based on retinal theory, wherein the left branch network decomposes the left image of the optimized image pair into a left reflected image and a left illuminated image, and the right branch network decomposes the right image of the optimized image pair into a right reflected image and a right illuminated image;

[0074] Step S4: Building a feature extraction network TB-FEN to extract features of the left image of the optimized image pair and the left reflected image in step S3, as well as the right image of the optimized image pair and the right reflected image in step S3, and perform feature consistency constraints;

[0075] In step S5, a total loss function is constructed based on the feature consistency constraints in steps S2, S3, and S4 to guide network training.

[0076] Furthermore, in step S2, the lightweight optimization network constructed is specifically:

[0077] The constructed lightweight optimization network consists of 4 convolutional activation layers and 1 convolutional layer with a convolution kernel size of 3x3. Each convolutional activation layer contains a convolution layer with a convolution kernel size of 3x3 and a ReLU activation function. The final output channel number of the convolution layer with a convolution kernel size of 3x3 is 3, which is used to generate the optimized image.

[0078] The left and right images of the low-light image pair are passed through a lightweight optimization network to obtain the optimized image, see the expression;

[0079] ;

[0080] The projection consistency loss function is: ;

[0081] Among them, I L is the left image of the low-light image pair, I R is the right image of the low-light image pair; i L To optimize the left image of the image pair, i R To optimize the right image of the image pair, To optimize the network, represents the square of the Euclidean norm, is the value of the projection consistency loss function.

[0082] Furthermore, in step S3, a two-branch network is built based on retinal theory, wherein the left branch network decomposes the left image of the optimized image pair to obtain a left reflected image and a left illuminated image, and the right branch network decomposes the right image of the optimized image pair to obtain a right reflected image and a right illuminated image; specifically, as follows:

[0083] Step S31: A dual-branch network processes the left image and the right image of the optimized image pair respectively. The left branch and the right branch of the dual-branch network are symmetrically arranged and have the same network structure, consisting of a reflection decomposition network and an illumination decomposition network.

[0084] Step S32: extracting the reflection image using a reflection decomposition network, which consists of five convolutional activation layers. The first four convolutional activation layers include a convolutional layer with a convolution kernel size of 3x3 and a ReLU activation function. The last convolutional activation layer includes a convolutional layer with a convolution kernel size of 3x3 and a Sigmoid activation function. The number of output channels is 3.

[0085] The illumination decomposition network extracts the illumination image and consists of 5 convolutional activation layers. The first 4 convolutional activation layers include a convolution layer with a convolution kernel size of 3x3 and a ReLU activation function. The last convolutional activation layer includes a convolution layer with a convolution kernel size of 3x3 and a Sigmoid activation function. The number of output channels is 1.

[0086] Step S33, the image decomposition process is expressed as:

[0087] ;

[0088] in, are the left and right images of the optimized image pair, and Represent the left and right images of the reflected image and the left and right images of the illuminated image, Represents the reflection decomposition network, represents the illumination decomposition network;

[0089] Step S34, introduce the illumination smoothing loss function based on total variation regularization, expressed as:

[0090] ;

[0091] in, , Represents the gradient of pixel (i, j) on the horizontal x-axis, , Represents the gradient of pixel (i, j) on the vertical y-axis, Indicates summation, Indicates the absolute value, Represents the value of the illumination smoothness loss function;

[0092] Step S35, introduce reflection reconstruction error loss, expressed as:

[0093] ;

[0094] in represents the value of the reflection reconstruction error loss function, represents the square of the L2 norm, Represents dot product.

[0095] Furthermore, in step S4: a feature extraction network TB-FEN is constructed to extract features of the left image of the optimized image pair and the left reflected image in step S3, and the right image of the optimized image pair and the right reflected image in step S3, respectively, and perform feature consistency constraints; specifically:

[0096] Step S41, the feature extraction network TB-FEN extracts the left and right images of the optimized image pair respectively Left and right images with reflected image Extract image semantic information at different levels and achieve feature alignment;

[0097] Step S42: the feature extraction network TB-FEN extracts the left and right images of the input optimized image pair. Left and right images with reflected image A parameter-sharing weight encoder is used for feature extraction to ensure the consistency of extracted features;

[0098] Step S43: The parameter-shared weight encoder includes three levels of feature extraction modules, corresponding to a shallow feature extraction module, a middle feature extraction module, and a deep feature extraction module, respectively. Each level of feature extraction module is composed of a residual connection module.

[0099] The residual connection module consists of a convolution layer with a convolution kernel size of 3x3, a batch normalization layer, and a ReLU activation function. The input and output are added together to form a residual connection. Each residual connection module uses the same network structure configuration.

[0100] Step S44, shallow feature extraction module: shallow feature consistency loss, defined as:

[0101] ;

[0102] in, Represents the L1 norm, which is used to measure the absolute difference between features. represents the shallow feature extraction module, Represents shallow feature consistency loss;

[0103] The middle-level feature extraction module measures the similarity between the optimized image and the reflected image in terms of local structural distribution: the middle-level feature consistency loss is calculated based on the structural similarity index and is defined as:

[0104] ;

[0105] in, represents the structural similarity index, represents the middle-level feature extraction module, represents the consistency loss of mid-level features;

[0106] Deep feature extraction module: The loss function of the deep feature extraction module is based on cosine similarity and is defined as:

[0107] ;

[0108] in, represents the deep feature extraction module, represents the deep feature consistency loss;

[0109] In step S45, a weighted total loss function is used to optimize the feature alignment effect of multiple levels. The total feature consistency loss function is defined as follows:

[0110] ;

[0111] in, is the weight hyperparameter of the shallow, middle and deep loss terms, which is used to control the contribution ratio of different level features to the total loss. represents the total feature consistency loss function.

[0112] Furthermore, in step S5, a total loss function is constructed to guide network training; specifically:

[0113] The total loss function is: .

[0114] like Figure 2 As shown, the present invention adopts another technical solution: a low-light image enhancement method based on a binocular camera, applying the low-light image unsupervised training method based on a binocular camera, comprising the following steps:

[0115] Step T1: synchronously capture the left image I of the low-light image pair through the binocular camera L Right image I of the low-light image pair R ;

[0116] Step T2, loading the parameters of the lightweight optimization network and the dual-branch network trained in steps S2 and S3;

[0117] Step T3: the left image I of the image pair L and right image I R Pass the lightweight optimization network to get the left optimized image i of the optimized image pair L , the right optimized image i of the optimized image pair R , the left branch network optimizes the left optimized image i of the image pair through the reflection decomposition network and the illumination decomposition networkL Decomposed into left reflected image R L and the left illumination image L L ,

[0118] The right branch network optimizes the right optimized image i of the image pair through the reflection decomposition network and the illumination decomposition network R Decomposed into right reflected image R R and the right illumination image L R ;

[0119] Step T4: the left and right images L of the illumination image x Enhance the image and fuse the enhanced illumination image with the reflection image to obtain the enhanced illumination image pair. .

[0120] Furthermore, in step T4, the left and right images L of the illumination image are x Enhance the image and fuse the enhanced illumination image with the reflection image to obtain the enhanced illumination image pair. Specifically:

[0121] Step T41: the left and right images L of the illumination image x , the enhanced illumination image L is obtained by gamma transformation Hx ,formula: ;

[0122] in, is the left and right image L of the illumination image x Perform gamma transformation;

[0123] Step T42: Enhance the illumination image L Hx Combined with the reflected image, we get the image pair after illumination enhancement, formula: L Hx ;

[0124] Among them, H x is the image pair after illumination enhancement, Represents the left and right images of the illumination image.

[0125] The foregoing is merely a list of specific embodiments of the present application, intended to enable those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the broadest scope consistent with the principles and novel features of the present application.

Claims

1. An unsupervised training method for low-light images based on a binocular camera, characterized in that: The following steps are involved: Step S1, using a binocular camera to synchronously capture a pair of real-world low-light images, remove blurry images, and construct a dataset; Step S2: Build and use a lightweight optimization network to remove sensor noise and compression artifacts from the dataset in step S1 to generate optimized image pairs. Step S3: Building a two-branch network based on retinal theory, wherein the left branch network decomposes the left image of the optimized image pair into a left reflected image and a left illuminated image, and the right branch network decomposes the right image of the optimized image pair into a right reflected image and a right illuminated image; Step S4: Building a feature extraction network TB-FEN to extract features of the left image of the optimized image pair and the left reflected image in step S3, as well as the right image of the optimized image pair and the right reflected image in step S3, and perform feature consistency constraints; Step S5, constructing a total loss function based on the feature consistency constraints in steps S2, S3, and S4 to guide network training; In step S4: a feature extraction network TB-FEN is constructed to extract features of the left image of the optimized image pair and the left reflected image in step S3, and the right image of the optimized image pair and the right reflected image in step S3, respectively, and perform feature consistency constraints; Specifically: Step S41, the feature extraction network TB-FEN extracts the left and right images of the optimized image pair respectively Left and right images with reflected image Extract image semantic information at different levels and achieve feature alignment; Step S42: the feature extraction network TB-FEN extracts the left and right images of the input optimized image pair. Left and right images with reflected image A parameter-sharing weight encoder is used for feature extraction to ensure the consistency of extracted features; Step S43: The parameter-shared weight encoder includes three levels of feature extraction modules, corresponding to a shallow feature extraction module, a middle feature extraction module, and a deep feature extraction module, respectively. Each level of feature extraction module is composed of a residual connection module. The residual connection module consists of a convolution layer with a convolution kernel size of 3x3, a batch normalization layer, and a ReLU activation function. The input and output are added together to form a residual connection. Each residual connection module uses the same network structure configuration. Step S44, shallow feature extraction module: shallow feature consistency loss, defined as: ； in, Represents the L1 norm, which is used to measure the absolute difference between features. represents the shallow feature extraction module, Represents shallow feature consistency loss; The middle-level feature extraction module measures the similarity between the optimized image and the reflected image in terms of local structural distribution: the middle-level feature consistency loss is calculated based on the structural similarity index and is defined as: ； in, represents the structural similarity index, represents the middle-level feature extraction module, Represents the consistency loss of mid-level features; Deep feature extraction module: The loss function of the deep feature extraction module is based on cosine similarity and is defined as: ； in, represents the deep feature extraction module, represents the deep feature consistency loss; In step S45, a weighted total loss function is used to optimize the feature alignment effect of multiple levels. The total feature consistency loss function is defined as follows: ； in, is the weight hyperparameter of the shallow, middle and deep loss terms, which is used to control the contribution ratio of different level features to the total loss. represents the total feature consistency loss function.

2. The unsupervised training method for low-light images based on a binocular camera according to claim 1, characterized in that: In step S2, the lightweight optimization network is constructed as follows: The constructed lightweight optimization network consists of 4 convolutional activation layers and 1 convolutional layer with a convolution kernel size of 3x3. Each convolutional activation layer contains a convolution layer with a convolution kernel size of 3x3 and a ReLU activation function. The final output channel number of the convolution layer with a convolution kernel size of 3x3 is 3, which is used to generate the optimized image. The left and right images of the low-light image pair are passed through a lightweight optimization network to obtain the optimized image, see the expression; ； The projection consistency loss function is: ; Among them, I L is the left image of the low-light image pair, I R is the right image of the low-light image pair; i L To optimize the left image of the image pair, i R To optimize the right image of the image pair, To optimize the network, represents the square of the Euclidean norm, is the value of the projection consistency loss function.

3. The unsupervised training method for low-light images based on a binocular camera according to claim 2, characterized in that: Step S3: Building a dual-branch network based on retinal theory, the left branch network decomposes the left image of the optimized image pair to obtain a left reflected image and a left illuminated image, and the right branch network decomposes the right image of the optimized image pair to obtain a right reflected image and a right illuminated image; specifically: Step S31: A dual-branch network processes the left image and the right image of the optimized image pair respectively. The left branch and the right branch of the dual-branch network are symmetrically arranged and have the same network structure, consisting of a reflection decomposition network and an illumination decomposition network. Step S32: extracting the reflection image using a reflection decomposition network, which consists of five convolutional activation layers. The first four convolutional activation layers include a convolutional layer with a convolution kernel size of 3x3 and a ReLU activation function. The last convolutional activation layer includes a convolutional layer with a convolution kernel size of 3x3 and a Sigmoid activation function. The number of output channels is 3. The illumination decomposition network extracts the illumination image and consists of 5 convolutional activation layers. The first 4 convolutional activation layers include a convolution layer with a convolution kernel size of 3x3 and a ReLU activation function. The last convolutional activation layer includes a convolution layer with a convolution kernel size of 3x3 and a Sigmoid activation function. The number of output channels is 1. Step S33, the image decomposition process is expressed as: ； in, are the left and right images of the optimized image pair, and Represent the left and right images of the reflected image and the left and right images of the illuminated image, Reflection decomposition network, represents the illumination decomposition network; Step S34, introduce the illumination smoothing loss function based on total variation regularization, expressed as: ； in, , Represents the gradient of pixel (i, j) on the horizontal x-axis, , Represents the gradient of pixel (i, j) on the vertical y-axis, Indicates summation, Indicates the absolute value, Represents the value of the illumination smoothness loss function; Step S35, introduce reflection reconstruction error loss, expressed as: ； in represents the value of the reflection reconstruction error loss function, represents the square of the L2 norm, Represents dot product.

4. The unsupervised training method for low-light images based on a binocular camera according to claim 3, characterized in that: In step S5, a total loss function is constructed to guide network training; specifically: The total loss function is: .

5. A low-light image enhancement method based on a binocular camera, applying the low-light image unsupervised training method based on a binocular camera according to claim 4, characterized in that: The following steps are involved: Step T1: synchronously capture the left image I of the low-light image pair through the binocular camera L and the right image I of the low-light image pair R ; Step T2, loading the parameters of the lightweight optimization network and the dual-branch network trained in steps S2 and S3; Step T3: the left image I of the image pair L and right image I R Pass the lightweight optimization network to get the left optimized image i of the optimized image pair L , the right optimized image i of the optimized image pair R , the left branch network optimizes the left optimized image i of the image pair through the reflection decomposition network and the illumination decomposition network L Decomposed into left reflected image R L and the left illumination image L L ; The right branch network optimizes the right optimized image i of the image pair through the reflection decomposition network and the illumination decomposition network R Decomposed into right reflected image R R and the right illumination image L R ; Step T4: The left and right images L of the illumination image are x Enhance the image and fuse the enhanced illumination image with the reflection image to obtain the enhanced illumination image pair. .

6. The low-light image enhancement method based on a binocular camera according to claim 5, characterized in that: Step T4: The left and right images L of the illumination image are x Enhance the image and fuse the enhanced illumination image with the reflection image to obtain the enhanced illumination image pair. Specifically: Step T41: the left and right images L of the illumination image x , the enhanced illumination image L is obtained by gamma transformation Hx ,formula: ; in, is the left and right image L of the illumination image x Perform gamma transformation; Step T42: Enhance the illumination image L Hx Combined with the reflected image, we get the image pair after illumination enhancement, formula: L Hx ; Among them, H x is the image pair after illumination enhancement, Represents the left and right images of the illumination image.

Images