An unsupervised low-light image enhancement method and device based on MRBNet

Abstract

The application discloses an unsupervised low-light image enhancement method and device based on MRBNet, and the method comprises the following steps: constructing a data set, wherein only low-light images are included; training MRBNet by using low-light pictures in the data set; presetting two-stage image enhancement stages and a denoising stage of MRBNet; wherein the image enhancement stage comprises two branches: the first branch is composed of five convolutional layers and a local perception module (LAN), and the second branch is composed of an encoder-decoder structure and a jump connection; in the test stage, only the first branch is used for low-light picture enhancement after training; and the enhanced picture is denoised by a blind spot network (BSN). This method no longer needs paired low-light and normal-light pictures, only uses low-light pictures for training, and can better solve the problems of insufficient image enhancement in the unsupervised training mode and large noise of the enhanced image affecting the picture quality.

Description

Technical Field

[0001] This invention relates to the field of low-light image enhancement, specifically a low-light image enhancement method and apparatus based on MRBNet (Mutual referencing and blind denoising network) for mutual referencing low-light enhancement and blind spot denoising. Background Technology

[0002] Low-light image enhancement is an important research direction in computer vision, aiming to improve the quality and visualization of images acquired under low-light conditions. In uncontrolled environments, the imaging quality of capturing devices is often difficult to guarantee, especially under dim or extremely low lighting conditions, making low-quality images inevitable. These low-light images not only impair the aesthetic quality of the images but also severely impact the performance of advanced computer vision tasks that rely on images as information input, such as object tracking, object recognition, and object detection. The development of low-light image enhancement techniques provides an effective solution for improving image quality. This not only improves the human visual experience but also significantly enhances the performance of computer vision tasks that rely on image processing and analysis. Through the application of these techniques, high-quality images can be acquired in a variety of complex environments, thereby promoting the development and application of related fields.

[0003] The main purpose of low-light image enhancement is to make originally unclear or low-brightness images clearer or brighter by highlighting key information and minimizing interference. With continuous technological advancements, deep learning has demonstrated strong application potential in various fields and is increasingly being applied to low-light image enhancement, achieving significant progress. However, several important problems remain to be solved in low-light image enhancement: supervised low-light image enhancement methods show strong enhancement effects on dataset images but have low generalization ability, perform poorly when processing different scenes from the training dataset, and low-light and normal-light image pairs are difficult to capture in reality, requiring a large number of synthetic images for training. Unsupervised low-light image enhancement methods suffer from low performance due to the lack of supervision information. Furthermore, both supervised and unsupervised methods face noise problems. The degree of noise suppression after enhancement directly determines the image quality. This invention employs a deep learning method, utilizing a dual-branch approach to extract reflection components and form mutual references, using a local detail perception module to preserve detail information, and then using a blind spot network to denoise the enhanced image, thus completing the low-light image enhancement. Summary of the Invention

[0004] To address the issues of low image enhancement performance and poor noise suppression in unsupervised image enhancement methods, this invention proposes an unsupervised low-light image enhancement method and apparatus based on MRBNet.

[0005] This invention employs two image enhancement branches to extract the reflection component, using the results as a reference. This method improves the overall performance of the image enhancement stage. Furthermore, the two branches utilize different network structures, aiming to preserve more texture detail, color information, and extract deeper semantic information. The method employs a blind spot network for noise suppression in the enhanced image, significantly reducing noise. This represents a bold application of deep learning denoising techniques to low-light image enhancement, achieving excellent results. This method improves the network's image enhancement performance, suppresses noise in the enhanced image, and exhibits good generalization ability.

[0006] This invention proposes an unsupervised low-light image enhancement method based on MRBNet, comprising:

[0007] S1. Construct the dataset: The dataset contains only low-light images;

[0008] S2. Randomly crop the input low-light image to be enhanced into a 128×128 pixel image;

[0009] S3. Input the low-light image into the preset MRBNet; the preset MRBNet includes two stages: image enhancement stage and denoising stage; the image enhancement stage includes: image detail enhancement branch and encoder-decoder enhancement branch; the denoising stage includes: blind spot network (BSN) denoising module;

[0010] S4. Input the low-light image into the image detail enhancement branch, and separate the illumination component through convolution operation; calculate the reflection component according to Retinex theory; input the reflection component into the local detail perception module to further extract image detail information;

[0011] S5. Input the low-light image into the encoder-decoder enhancement branch to separate the illumination component, and calculate the reflection component based on Retinex theory;

[0012] S6. During training, the two branches of the image enhancement stage simultaneously decompose the same input to obtain the illumination component, and then calculate the reflection component and use them as a reference; during testing, only the image detail enhancement branch is used to obtain the reflection component.

[0013] S7. The reflection component obtained from the image detail enhancement branch is used as the input image to the blind spot network for denoising, and finally the low-light image enhancement result is obtained.

[0014] The image enhancement stage described in S3 includes two image enhancement branches: an image detail enhancement branch and an encoder-decoder enhancement branch. The image detail enhancement branch uses only a simple five-layer convolutional layer to extract the image's feature information, and then uses a local detail awareness module to refine the texture details and color information. Compared to the image detail enhancement branch, the encoder-decoder enhancement branch aims to extract multi-scale feature information of the image and adds skip connections to the network, while preserving low-level edge information and high-level semantic information. Both networks ultimately calculate the reflection components based on Retinex theory. According to the consistency of the reflection components, the reflection components obtained by the two branches form a reference for each other, removing noise information and complementing each other's missing feature information.

[0015] Furthermore, the image detail enhancement branch uses a concise five-layer convolutional layer to separate the illumination components, with each convolutional kernel being 3×3 in size; reflection fill is used before each convolutional layer, which effectively reduces edge artifacts while preserving the spatial structure of the image.

[0016] The output layer uses the Sigmoid activation function, while other layers use the ReLU activation function. The reflection component is calculated using the illumination component based on Retinex theory, and the corresponding calculation formula is as follows:

[0017]

[0018] Where R is the reflection component, T is the illumination component, I is the input low-light image, and the operator is element-wise division;

[0019] The image detail enhancement branch uses a local detail perception module to refine the texture details and color information of the reflection component; the local detail perception module creates a local receptive field by stacking 3×3 convolutional layers; in order to refine the color information, a channel attention mechanism is also added to the network; the reflection component is passed through the local perception network to obtain the enhancement result of the image enhancement stage.

[0020] Furthermore, the encoder-decoder enhancement branch consists of an encoder, a decoder, and a skip connection structure;

[0021] The encoder structure has five convolutional blocks, each consisting of a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function. Downsampling is performed using convolution operations with a stride of 2, allowing the network to learn richer features while performing downsampling.

[0022] The decoder also consists of five convolutional blocks, each containing a transposed convolutional layer for upsampling. Skip connections are used to concatenate the feature maps of the encoder to the corresponding layers of the decoder. The encoder-decoder network structure can extract multi-scale features from the input image, from low-level edge information to high-level semantic information. Skip connections are used to directly pass the features of each layer of the encoder to the corresponding decoder layer. This connection method can retain more detailed information.

[0023] The illumination component is obtained by separating it using an encoder-decoder network, and the reflection component is then calculated based on Retinex theory.

[0024] Specifically, a hybrid loss function is used to train the network for the image detail enhancement branch and the encoder-decoder enhancement branch in the image enhancement stage; this includes illumination separation loss, reflection consistency loss, Retinex loss, and local detail perception loss.

[0025] Among them, the illumination separation loss can effectively preserve more detailed information, so as to avoid losing too much information when separating illumination components, which would prevent the subsequent local detail perception module from functioning; its expression is as follows:

[0026]

[0027] Where I is the input low-light image, and T1 is the illumination component separated by the image detail enhancement branch. Represented as the squared L2 norm;

[0028] According to Retinex theory, the reflectance components of an image are independent of external factors, and the reflectance components obtained from the two branches should be identical. The reason for differences in the reflectance components obtained from the two branches is the presence of significant noise in low-light images and the varying performance of the two branches' networks. Using reflectance consistency loss allows both reflectance components to retain the same information while removing the differing information, which is mostly noise. Its expression is:

[0029]

[0030] Where R1 is the reflection component calculated by the image detail enhancement branch, and R2 is the reflection component calculated by the encoder-decoder enhancement branch;

[0031] Since Retinex theory is used to enhance low-light images, Retinex loss is introduced as a constraint, and its expression is:

[0032]

[0033] Where I0 is calculated from the maximum values ​​of the R, G, and B channels, the expression is:

[0034]

[0035] The expression for the local detail-aware loss is:

[0036] L lan =||R1-R e ||1 (6)

[0037] Where R e For the output of the local detail perception module, ||·||1 represents the L1 norm;

[0038] The loss function L used during the final image enhancement training stage is L t L c L r L lan The weighted sum, i.e., L = λ1L t +λ2L c +λ3L r +λ4L lan Where λ1, λ2, λ3, and λ4 are loss weight coefficients.

[0039] In step S3, the denoising stage takes the output of the image detail enhancement branch in the image enhancement stage as input and uses a blind spot network for denoising. The characteristic of using a blind spot network for denoising is that the network can be trained for denoising using only a single noisy image.

[0040] Blind spot networks make the following assumptions about noisy images: signals at different pixel locations are correlated, noise at different pixel locations is uncorrelated, and the mean of the noise is 0.

[0041] Based on this assumption, the blind spot network generates blind spot pixels from the input image pixels and predicts the blind spot pixels based on the surrounding pixels. According to the assumption, during training, since the signals of the pixels at the blind spot position are correlated with the signals of the surrounding pixels, the network can predict the pixel signal at the blind spot position. However, since the noise information contained in the pixels at the blind spot position is uncorrelated with the noise information contained in the surrounding pixels, the network cannot predict the noise information of the pixels at the blind spot position, thus achieving the effect of noise reduction.

[0042] The superiority of blind spot network denoising has been proven by scholars in the field of denoising. However, in reality, noisy images are difficult to meet the assumptions of blind spot networks regarding noisy images. Therefore, asymmetric PD (Pixel-shuffle) downsampling is used to minimize the spatial correlation between adjacent noisy pixels, so that the noisy image can meet the assumptions of blind spot networks regarding noisy images. Therefore, in the denoising stage, asymmetric PD downsampling is first used to process the image, then blind spot network is used to denoise the image, and finally the inverse operation of asymmetric PD downsampling is used to restore the original size of the image, thus completing the denoising.

[0043] Furthermore, the loss function used during the blind spot network training in the denoising stage is expressed as follows:

[0044] L BSN =||PD -1 (BSN(PD(R1)))-R1||1 (7)

[0045] Where PD represents the PD downsampling operation, PD -1 R1 is the inverse operation of asymmetric PD downsampling, BSN is the blind spot network denoising operation, and R1 is the output of the image enhancement stage.

[0046] A second aspect of the present invention relates to an unsupervised low-light image enhancement apparatus based on MRBNet, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the unsupervised low-light image enhancement method based on MRBNet of the present invention.

[0047] A third aspect of the invention relates to a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the MRBNet-based unsupervised low-light image enhancement method of the present invention.

[0048] This invention eliminates the need for paired low-light and normal-light images, using only low-light images for training. It also better addresses the issues of insufficient image enhancement in unsupervised training methods and the problem of high noise levels in enhanced images affecting image quality.

[0049] Compared with the prior art, the present invention has the following advantages and technical effects:

[0050] The method proposed in this invention trains the network using only low-light images in an unsupervised manner. It proposes a dual-branch approach to extract reflection components and cross-reference them, enhancing image enhancement performance. One branch uses simple stacked convolutional layers to extract image features, incorporating a local detail perception module to enhance the extraction of texture and color information. The other branch employs an encoder-decoder network structure with skip connections, allowing it to simultaneously preserve low-level edge information and high-level semantic information. The reflection components extracted by the two branches cross-reference each other, significantly improving low-light enhancement performance and achieving satisfactory results. Furthermore, this method proposes using a blind spot network to denoise the low-light image enhancement results. The blind spot network exhibits excellent noise suppression capabilities for enhanced images, effectively addressing the issue of noise affecting image quality. The proposed method improves low-light image enhancement performance and demonstrates good generalization ability even when trained using only low-light images. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method of the present invention;

[0052] Figure 2 This is a network structure diagram of the method of the present invention;

[0053] Figure 3 This is a network structure diagram of the LAN module in the network structure diagram of the present invention. Detailed Implementation

[0054] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings:

[0055] Example 1

[0056] like Figure 1 As shown, this invention proposes an unsupervised low-light image enhancement method based on MRBNet, comprising the following steps:

[0057] Step 1: Preprocess the image by randomly cropping the input image to a size of 128×128 pixels;

[0058] Step 2: Input the image into the image detail enhancement branch, extract features through convolution operation and LAN module, separate the illumination component, and then calculate the reflection component according to Retinex theory;

[0059] Step 3: Input the image into the encoder-decoder branch, separate the illumination component through the encoder-decoder network, and then calculate the reflection component according to Retinex theory;

[0060] Step 4: The reflection components obtained from the two branches are cross-referenced to obtain the final output result of the image enhancement stage;

[0061] Step 5: Input the output of the image enhancement stage into the blind spot network for image denoising;

[0062] Step 6: Output the final low-light image enhancement result.

[0063] The following section provides a detailed explanation of each step and the network structure.

[0064] like Figure 2 As shown, MRBNet uses two branches to separate and obtain two illumination components during the training phase, and calculates the illumination components according to Retinex theory. During the testing phase, only the image detail enhancement branch is used to complete the low-light image enhancement task.

[0065] During the training phase, the initial learning rate is set to 1×10. -4 The network was optimized using the Adam optimizer; the number of training iterations was set to 200. The low-light images in the dataset are from the LOL dataset.

[0066] After random cropping preprocessing of the image, the image is input into the image detail enhancement branch and the encoder-decoder enhancement branch;

[0067] The image detail enhancement branch aims to extract more image detail and color information. To simplify the network structure, it uses 5 convolutional layers to extract feature information. The kernel size used is 3×3. Reflection padding is used before each convolutional layer. Reflection padding effectively preserves the features at the edges by adding a ring of reflected pixel values ​​to the image edges, reducing the loss of edge information during the convolution process.

[0068] The output layer uses the Sigmoid activation function, while other layers use the ReLU activation function. The reflection component is calculated using the illumination component based on Retinex theory, and the corresponding calculation formula is as follows:

[0069]

[0070] Where R is the reflection component, T is the illumination component, I is the input low-light image, and the operator is element-wise division;

[0071] The image detail branch uses a local detail-aware module to refine the texture details and color information of the reflection component; the local detail-aware module, such as... Figure 3As shown, the local detail perception module consists of convolutional layers and multiple residual blocks (RBs). The use of residual structures and nested residual blocks within the module aims to enhance the model's expressive and learning capabilities, thereby improving the performance of image enhancement tasks. Nesting multiple residual blocks allows for increasing network depth without significantly increasing computational complexity. Deeper networks typically possess stronger expressive power, enabling them to better capture and process detailed information in images, especially in tasks requiring high precision, such as low-light image enhancement.

[0072] A Channel Attention Layer (CA) is added to the RB structure. This CA adaptively allocates weights, enhancing the module's overall focus on key features. During feature extraction, certain feature channels may contain more important brightness or texture details. The channel attention mechanism dynamically increases the weights of these important channels, making the network pay more attention to this key information when processing the image, thereby improving the enhancement effect. In summary, the Local Detail Awareness module enables the network to extract more texture details, color, and brightness information, enhancing the final low-light image enhancement effect.

[0073] The reflection component, after passing through the local detail module, finally yields the output reflection component R1 of the image detail branch.

[0074] like Figure 2 As shown, the encoder-decoder enhancement branch consists of an encoder-decoder network and a skip connection structure.

[0075] The encoder structure has five convolutional blocks, each consisting of a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function. Downsampling is performed using convolution operations with a stride of 2, allowing the network to learn richer features while performing downsampling.

[0076] The decoder also consists of five convolutional blocks, each containing a transposed convolutional layer for upsampling. Skip connections are used to concatenate the feature maps of the encoder to the corresponding layers of the decoder. The encoder-decoder network structure can extract multi-scale features from the input image, from low-level edge information to high-level semantic information. Skip connections are used to directly pass the features of each layer of the encoder to the corresponding decoder layer. This connection method can retain more detailed information.

[0077] The advantage of the above-described structure in extracting features lies in its ability to extract features at different scales. The encoder progressively downsamples to extract global information of the image, while the decoder progressively upsamples to restore image details. This achieves efficient fusion of features at different scales. This fusion helps improve the image enhancement effect, enabling the network to better integrate information at different levels when processing low-light images, enhancing details and contrast.

[0078] The illumination component is obtained by separating the light components through an encoder-decoder network, and the reflection component R2 is calculated according to Retinex theory.

[0079] Two reflection components, R1 and R2, are obtained through the image detail enhancement branch and the encoder-decoder enhancement branch, respectively. These two reflection components form a mutual reference. This is done to provide more information to both branches during training, thus improving the training of the enhancement network. The theoretical basis for this mutual reference is Retinex theory, which posits that image perception is the result of separating illumination and reflection. Illumination represents the lighting conditions in a scene, while reflection represents the inherent color and texture of an object's surface. Therefore, the reflection components obtained from the two branches should be consistent. The difference in reflection components between the two branches is due to the significant noise in low-light images and the different performance characteristics of the two branches' networks. Therefore, the reflection components extracted by the two branches can form a mutual reference.

[0080] The network is trained using a hybrid loss function for the image detail enhancement branch and the encoder-decoder enhancement branch in the image enhancement stage; this includes illumination separation loss, reflection consistency loss, Retinex loss, and local detail perception loss.

[0081] Illumination separation loss can effectively preserve more detail information, preventing excessive information loss during illumination component separation, which could render subsequent local detail-aware modules ineffective; its expression is as follows:

[0082]

[0083] Where I is the input low-light image, and T1 is the illumination component separated by the image detail enhancement branch. It is represented as the squared L2 norm.

[0084] Using reflection consistency loss allows two reflection components to retain the same information while removing the differing information, which is mostly noise. Its expression is:

[0085]

[0086] Where R1 is the reflection component calculated by the image detail enhancement branch, and R2 is the reflection component calculated by the encoder-decoder enhancement branch.

[0087] Since Retinex theory is used to enhance low-light images, Retinex loss is introduced as a constraint, and its expression is:

[0088]

[0089] Where I0 is calculated from the maximum values ​​of the R, G, and B channels, the expression is:

[0090]

[0091] The expression for the local detail-aware loss is:

[0092] L lan =||R1-R e ||1 (6)

[0093] Where R e For the output of the local detail perception module, ||·||1 represents the L1 norm;

[0094] The loss function L used during the final image enhancement training stage is L t L c L r L lan The weighted sum, i.e., L = λ1L t +λ2L c +λ3L r +λ4L lan Where λ1, λ2, λ3, and λ4 are loss weight coefficients.

[0095] Finally, R1 is used as the final output in the image enhancement stage and as the input image in the denoising stage.

[0096] like Figure 2 As shown, a blind spot network is used for denoising. Blind spot network denoising only requires one noisy image. The input image after the image enhancement stage has a brightness similar to that of a normal light image, but contains a lot of noise, and can be directly used as the input image for the blind spot network.

[0097] Blind spot networks make the following assumptions about noisy images: signals at different pixel locations are correlated, noise at different pixel locations is uncorrelated, and the mean of the noise is 0.

[0098] Based on this assumption, the blind spot network generates blind spot pixels for the input image pixels and predicts the blind spot pixels based on the surrounding pixels. According to the assumption, during training, since the signals of the pixels at the blind spot position are correlated with the signals of the surrounding pixels, the network can predict the pixel signal at the blind spot position. However, since the noise information contained in the pixels at the blind spot position is uncorrelated with the noise information contained in the surrounding pixels, the network cannot predict the noise information of the pixels at the blind spot position, thus achieving the effect of noise reduction.

[0099] The superiority of blind spot network denoising has been proven by scholars in the field of denoising. However, in reality, noisy images are difficult to meet the assumptions of blind spot networks regarding noisy images. Therefore, asymmetric PD (Pixel-shuffle) downsampling is used to minimize the spatial correlation between adjacent noisy pixels, so that the noisy image can meet the assumptions of blind spot networks regarding noisy images. Therefore, in the denoising stage, asymmetric PD downsampling is first used to process the image, then blind spot network is used to denoise the image, and finally the inverse operation of asymmetric PD downsampling is used to restore the original size of the image, thus completing the denoising.

[0100] The loss function used during the blind spot network training in the denoising stage is expressed as follows:

[0101] L BSN =||PD -1 (BSN(PD(R1)))-R1||1 (7)

[0102] Where PD represents the PD downsampling operation, PD -1 R1 is the inverse operation of asymmetric PD downsampling, BSN is the blind spot network denoising operation, and R1 is the output of the image enhancement stage.

[0103] The proposed method, which involves extracting reflection components from two branches to form a mutual reference and using a blind spot network to enhance image denoising, significantly improves the performance of low-light image enhancement and achieves satisfactory low-light enhancement results.

[0104] Example 2

[0105] This embodiment relates to an unsupervised low-light image enhancement device based on MRBNet, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the unsupervised low-light image enhancement method based on MRBNet described in Embodiment 1.

[0106] Example 3

[0107] This embodiment relates to a computer-readable storage medium storing a program that, when executed by a processor, implements the MRBNet-based unsupervised low-light image enhancement method of Embodiment 1.

[0108] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. An unsupervised low-light image enhancement method based on MRBNet, characterized in that, Includes the following steps: S1. Construct the dataset: The dataset contains only low-light images; S2. Randomly crop the input low-light image to be enhanced into a 128×128 pixel image; S3. Input the low-light image into the preset MRBNet; The default MRBNet consists of two stages: an image enhancement stage and a denoising stage; The image enhancement stage includes an image detail enhancement branch and an encoder-decoder enhancement branch; the denoising stage includes a blind spot network (BSN) denoising module. The image detail branch uses only a simple five-layer convolutional layer to extract image feature information, and then uses a local detail awareness module to refine texture details and color information. Compared to the image detail branch, the encoder-decoder enhancement branch aims to extract multi-scale feature information of the image and adds skip connections to the network, while preserving low-level edge information and high-level semantic information. Both networks ultimately calculate the reflection components based on Retinex theory. According to the consistency of the reflection components, the reflection components obtained by the two branches form a reference for each other, removing noise information and complementing each other's missing feature information. S4. Input the low-light image into the image detail enhancement branch, and separate the illumination component through convolution operation; calculate the reflection component according to Retinex theory; input the reflection component into the local detail perception module to further extract image detail information; S5. Input the low-light image into the encoder-decoder enhancement branch to separate the illumination component, and calculate the reflection component based on Retinex theory; S6. During training, the two branches of the image enhancement stage simultaneously decompose the same input to obtain the illumination component, and then calculate the reflection component and use them as a reference; during testing, only the image detail enhancement branch is used to obtain the reflection component. S7. The reflection component obtained from the image detail enhancement branch is used as the input image to the blind spot network for denoising, and finally the low-light image enhancement result is obtained.

2. The unsupervised low-light image enhancement method based on MRBNet according to claim 1, characterized in that, The image detail enhancement branch uses a concise five-layer convolutional layer to separate the illumination components, with each convolutional kernel being 3×3. Reflection padding is used before each convolutional layer. Reflection padding can effectively reduce edge artifacts while preserving the spatial structure of the image. The output layer uses the Sigmoid activation function, while other layers use the ReLU activation function. The reflection component is calculated using the illumination component based on Retinex theory, and the corresponding calculation formula is: in, For the reflection component, For light component, Given a low-light image as input, the operator is element-wise division; The image detail branch uses a local detail perception module to refine the texture details and color information of the reflection component; the local detail perception module creates a local receptive field by stacking 3×3 convolutional layers; in order to refine the color information, a channel attention mechanism is also added to the network; the reflection component is processed by the local perception network to obtain the enhancement result of the image enhancement stage.

3. The unsupervised low-light image enhancement method based on MRBNet according to claim 1, characterized in that, The encoder-decoder enhancement branch consists of an encoder, a decoder, and a skip connection structure; The encoder structure has five convolutional blocks, each consisting of a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function. Downsampling is performed using convolution operations with a stride of 2, allowing the network to learn richer features while performing downsampling. The decoder also consists of five convolutional blocks, each containing a transposed convolutional layer for upsampling. Skip connections are used to concatenate the feature maps of the encoder to the corresponding layers of the decoder. The encoder-decoder network structure can extract multi-scale features from the input image, from low-level edge information to high-level semantic information. Skip connections are used to directly pass the features of each layer of the encoder to the corresponding decoder layer. This connection method can retain more detailed information. The illumination component is obtained by separating it using an encoder-decoder network, and the reflection component is then calculated based on Retinex theory.

4. The unsupervised low-light image enhancement method based on MRBNet according to claim 1, characterized in that, The image detail enhancement branch and encoder-decoder enhancement branch in the image enhancement stage described in step S3 use a hybrid loss function to train the network; These include illumination separation loss, reflection consistency loss, Retinex loss, and local detail perception loss.

5. The unsupervised low-light image enhancement method based on MRBNet according to claim 4, characterized in that, Illumination separation loss can effectively preserve more detail information, preventing excessive information loss during illumination component separation, which could render subsequent local detail-aware modules ineffective; its expression is as follows: in, For the input low-light image, The illumination component separated for image detail enhancement branches. Represented as the squared L2 norm; According to Retinex theory, the reflectance components of an image are independent of external factors, and the reflectance components obtained from the two branches should be identical. The reason for differences in the reflectance components obtained from the two branches is the presence of significant noise in low-light images and the varying performance of the two branches' networks. Using reflectance consistency loss allows both reflectance components to retain the same information while removing the differing information, which is mostly noise. Its expression is: in, The reflection component is calculated for the image detail enhancement branch. The reflection component is calculated for the encoder-decoder enhancement branch; Since Retinex theory is used to enhance low-light images, Retinex loss is introduced as a constraint, and its expression is: in It is calculated using the maximum values ​​of the R, G, and B channels, and the expression is: The expression for the local detail-aware loss is: in Output results for the local detail perception module. Represented as the L1 norm; The loss function used during the final image enhancement stage training for , , , The weighted sum, i.e. ;in , , , This is the loss weighting coefficient.

6. The unsupervised low-light image enhancement method based on MRBNet according to claim 1, characterized in that, The denoising stage described in step S3 takes the output of the image detail enhancement branch in the image enhancement stage as input and uses a blind spot network for denoising. The characteristic of using a blind spot network for denoising is that the network can be trained for denoising using only a single noisy image. Blind spot networks make the following assumptions about noisy images: signals at different pixel locations are correlated, noise at different pixel locations is uncorrelated, and the mean of the noise is 0. Based on this assumption, the blind spot network generates blind spot pixels from the input image pixels and predicts the blind spot pixels based on the surrounding pixels. According to the assumption, during training, since the signals of the pixels at the blind spot position are correlated with the signals of the surrounding pixels, the network can predict the pixel signal at the blind spot position. However, since the noise information contained in the pixels at the blind spot position is uncorrelated with the noise information contained in the surrounding pixels, the network cannot predict the noise information of the pixels at the blind spot position, thus achieving the effect of noise reduction. Asymmetric PD (Pixel-shuffle) downsampling is used to minimize the spatial correlation between adjacent noisy pixels, so that the noisy image satisfies the assumptions of the blind spot network about the noisy image. Therefore, in the denoising stage, asymmetric PD downsampling is first used to process the image, then the blind spot network is used to denoise the image, and finally the inverse operation of asymmetric PD downsampling is used to restore the original size of the image, thus completing the denoising.

7. The unsupervised low-light image enhancement method based on MRBNet according to claim 6, characterized in that, The loss function used during the blind spot network training in the denoising stage is expressed as follows: in For PD downsampling operation, This is the inverse operation of asymmetric PD downsampling. Denoising operations for blind spot networks, This is the output of the image enhancement stage.

8. An unsupervised low-light image enhancement device based on MRBNet, characterized in that, The method includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the unsupervised low-light image enhancement method based on MRBNet as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the unsupervised low-light image enhancement method based on MRBNet as described in any one of claims 1-7.

Images