Self-supervised low-light image enhancement method and system

Abstract

The application provides a self-supervised low-illumination image enhancement method and system, and belongs to the technical fields of computer vision and image processing, and the method comprises the following steps: performing histogram equalization processing on a low-illumination image, and then extracting histogram equalization prior information by using a prior mapping function; combining the maximum channel value with the low-illumination image to obtain a feature image in a connection graph by using a convolution layer; taking the histogram equalization prior information and the feature image in the connection graph as inputs of a spatial feature conversion layer, learning a modulation parameter of the histogram equalization prior information to adaptively perform a spatial biomimetic transformation on the feature image in the connection graph, and obtaining a feature image into which the prior information is fused; inputting the feature image into which the prior information is fused into an adaptive deep neural network model to obtain a reflection map and a brightness map; and the reflection map is a recovered image. The low-illumination image is enhanced by using the trained adaptive deep neural network, and the enhancement effect is greatly improved.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision technology and image processing technology, and more specifically, relates to a self-supervised low-light image enhancement method and system. Background Technology

[0002] Low-light image enhancement is one of the core issues in image processing technology and can be widely applied to high-level computer vision tasks such as semantic segmentation, object recognition, and object detection. Low-light images are caused by limitations in existing visual imaging technologies. In low-light conditions such as nighttime, insufficient light illumination significantly reduces horizontal visibility, resulting in severe degradation of the images acquired by the imaging sensor.

[0003] Traditional low-light enhancement methods include histogram equalization (HUE)-based methods and methods based on human visual models. The latter has received relatively more attention. A typical method based on human visual models decomposes a low-light image into reflectance and illumination components through some prior or regularization. The estimated reflectance is then considered as the enhancement effect. This method has some limitations: 1) The assumption that the reflectance component is the ideal result of the enhancement does not always hold, especially considering various lighting characteristics, which may lead to unrealistic enhancements such as loss of detail and color distortion; 2) Noise is usually ignored in HUE-based methods, so noise is either preserved or amplified in the enhancement result; 3) Finding an effective prior or regularization is challenging, and inaccurate priors or regularizations may lead to artifacts and color deviations in the enhanced result, and the running time is relatively long due to its complex optimization process.

[0004] Deep convolutional neural networks (CNNs) have achieved state-of-the-art performance in low-light image enhancement. Compared to traditional methods, CNNs can learn better feature representations, resulting in enhancements with superior visual quality, thanks to large datasets and powerful computational capabilities. However, current CNN-based methods require paired training instances, and capturing both low-light and normal-light images of the same visual scene is extremely challenging in practice. Therefore, current techniques are severely limited by the collection of datasets and do not represent a fundamental solution. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a self-supervised low-light image enhancement method and system. This invention addresses the problem that existing low-light image enhancement methods, which utilize deep convolutional neural networks, are severely limited by the collection of datasets. In practice, capturing low-light and normal-light images of the same visual scene is very challenging, resulting in poor enhancement effects of trained deep convolutional neural networks on low-light images.

[0006] To achieve the above objectives, the present invention provides a self-supervised low-light image enhancement method, comprising the following steps:

[0007] Histogram equalization is performed on low-light images, and then a mapping function is used to extract high-order feature information of the low-light images after histogram equalization. The high-order feature information is used as the prior information of histogram equalization. Here, low-light images refer to images taken when the ambient light intensity is less than 1 Lux.

[0008] The maximum channel value of the low-light image is calculated, and the maximum channel image is combined with the low-light image and then input into the first combination layer to obtain the feature image in the connection graph; wherein, the first combination layer includes a 9×9 convolution and ReLU layer;

[0009] Using histogram equalization prior information and feature images in the connection graph as inputs to the spatial feature transformation layer, the modulation parameter pair of histogram equalization prior information is learned to adaptively perform spatial affine transformation on the feature images in the connection graph to obtain feature images incorporating prior information.

[0010] The feature image incorporating prior information is input into an adaptive deep neural network model constructed based on the theory of retinal cortex, and a reflection map and a brightness map are obtained through several convolutional layers and sigmoid layers; wherein, the reflection map is a restored image.

[0011] More preferably, the prior mapping function is a deep convolutional neural network, which includes a five-layer structure: the first layer is a second combined layer, the middle three layers are a third combined layer, and the last layer is a 1×1 convolutional layer; wherein, the second combined layer includes a 3×3 convolution and a LeakyReLU layer; and the third combined layer includes a 1×1 convolution and a LeakyReLU layer.

[0012] More preferably, the method for the spatial feature transformation layer to obtain feature images incorporating prior information is as follows:

[0013] Two 1×1 convolutional layers are used to extract the intermediate parameters of the histogram equalization prior information, which are then multiplied with the feature image in the connection graph to obtain the first modulation parameters of the histogram equalization prior information.

[0014] Then, two 1×1 convolutional layers are used to extract the second modulation parameters of the histogram equalization prior information;

[0015] The first modulation parameter and the second modulation parameter are added together to obtain a feature image incorporating prior information.

[0016] More preferably, the method for obtaining the reflectance map and brightness map by incorporating prior information into the feature image is as follows:

[0017] The feature image incorporating prior information is input into the U-Net network to extract a deep feature image containing deeper feature information; wherein, the deeper feature information includes structural information and texture information of the low-light image;

[0018] A fourth combination layer is used to extract the feature image of the input low-light image; the fourth combination layer includes a 3×3 convolution and a ReLU layer.

[0019] Then, the feature image of the input image is combined with the deep feature image through a concatenation method;

[0020] Two 3×3 convolutional layers are used to fuse the combined feature images again;

[0021] Based on the re-fused feature images, a sigmoid layer is used to obtain the final reflection map and brightness map, where the reflection map is the restored image.

[0022] More preferably, the method for obtaining deep feature images containing deeper feature information is as follows:

[0023] The fifth combination layer is used to downsample the feature image incorporating prior information, thereby reducing the size of the feature image incorporating prior information and extracting a deep feature image containing deeper feature information; the fifth combination layer includes two 3×3 convolutional layers and a ReLU layer;

[0024] The sixth combination layer is used to upsample the deep feature image to restore it to the size of the feature image that incorporates prior information, while retaining deeper feature information; the sixth combination layer includes two 3×3 deconvolution layers and a ReLU layer.

[0025] On the other hand, the present invention provides a self-supervised low-light image enhancement system, comprising:

[0026] The histogram equalization processing module is used to perform histogram equalization processing on low-light images; wherein, the low-light image is an image taken when the ambient light intensity is less than 1 Lux;

[0027] The prior mapping module is used to extract high-order feature information of the low-light image after histogram equalization using a mapping function, and use the high-order feature information as the prior information of histogram equalization.

[0028] The maximum channel value calculation module is used to calculate the maximum channel value of low-light images;

[0029] The image connection module is used to combine the maximum channel image and the low-light image and input them into the first combination layer to obtain the feature image in the connection graph; wherein, the first combination layer includes a 9×9 convolution and ReLU layer;

[0030] The spatial feature transformation module is used to take the histogram equalization prior information and the feature image in the connection graph as input to the spatial feature transformation layer, learn the modulation parameter pair of the histogram equalization prior information to adaptively perform spatial affine transformation on the feature image in the connection graph, and obtain the feature image incorporating the prior information.

[0031] An adaptive neural network module, which contains an adaptive deep neural network model, is used to input feature images incorporating prior information into the adaptive deep neural network model constructed based on the theory of retinal cortex. The model obtains a reflection map and a brightness map through several convolutional layers and sigmoid layers; wherein the reflection map is a restored image.

[0032] More preferably, the prior mapping function in the prior mapping module is a deep convolutional neural network, which includes a five-layer structure: the first layer is a second combined layer, the middle three layers are a third combined layer, and the last layer is a 1×1 convolutional layer; wherein, the second combined layer includes a 3×3 convolution and a LeakyReLU layer; and the third combined layer includes a 1×1 convolution and a LeakyReLU layer.

[0033] More preferably, the method for the spatial feature transformation layer to obtain feature images incorporating prior information is as follows:

[0034] Two 1×1 convolutional layers are used to extract the intermediate parameters of the histogram equalization prior information, which are then multiplied with the features in the connection graph to obtain the first modulation parameters of the histogram equalization prior information.

[0035] Then, two 1×1 convolutional layers are used to extract the second modulation parameters of the histogram equalization prior information;

[0036] The first modulation parameter and the second modulation parameter are added together to obtain a feature image incorporating prior information.

[0037] More preferably, the adaptive deep neural network model in the adaptive neural network module includes a U-Net network, a fourth combined layer, two 3×3 convolutional layers, a sigmoid layer, and an adder; wherein the fourth combined layer includes a 3×3 convolutional layer and a ReLU layer;

[0038] The U-Net network receives feature images incorporating prior information to extract deep feature images containing deeper feature information; wherein, the deeper feature information includes structural and texture information of low-light images;

[0039] The fourth combination layer is used to extract the feature image of the input image from the input low-light image;

[0040] The adder is used to combine the feature image of the input image with the deep feature image through a concatenation mechanism;

[0041] The remaining two 3×3 convolutional layers fuse the combined feature images again;

[0042] A sigmoid layer is used to obtain the final reflection map and brightness map based on the re-fused feature image, where the reflection map is the restored image.

[0043] More preferably, the U-Net network includes a fifth combined layer and a sixth combined layer; wherein, the fifth combined layer includes two 3×3 convolutional and ReLU layers; and the sixth combined layer includes two 3×3 deconvolutional and ReLU layers.

[0044] The fifth combination layer is used to downsample the feature image incorporating prior information, reduce the size of the feature image incorporating prior information, and extract the deep feature image containing deeper feature information.

[0045] The sixth layer is used to upsample the deep feature image, restoring it to the size of the feature image that incorporates prior information, while retaining deeper feature information.

[0046] In summary, compared with the prior art, the above-described technical solutions conceived by this invention have the following advantages:

[0047] Beneficial effects:

[0048] Compared to existing deep neural networks that require enormous computational resources, this invention provides a self-supervised low-light image enhancement method and system. It offers a deep neural network structure for low-light image enhancement based on the characteristics of human vision. By fusing feature images through spatial feature layers, the depth of existing network structures is reduced. While reducing the number of convolutional layers, the same performance is maintained, enabling the new model to achieve more effective training and faster convergence speed with limited computational resources.

[0049] Currently, there is a lack of effective prior information to drive self-supervised learning of low-light images. This invention provides a method for obtaining histogram equalization prior information. By performing histogram equalization on the low-light image, the histogram equalization prior information is extracted from the histogram equalized image through a mapping function as guidance information for the network, thus eliminating the dependence on paired data information. Therefore, compared with existing self-supervised methods, the novel prior information provides highly effective information, enabling the use of a trained adaptive deep neural network to enhance low-light images, resulting in a significant improvement in enhancement performance. Attached Figure Description

[0050] Figure 1This is a schematic diagram of the architecture of the self-supervised low-light image enhancement method provided in an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of a deep neural network based on human visual characteristics provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the prior mapping function provided in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the spatial feature conversion layer provided in an embodiment of the present invention;

[0054] Figure 5 This is a low-light image enhancement effect diagram provided in an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] The architecture of the self-supervised low-light image enhancement method and system provided by the present invention includes three parts: (1) a prior mapping function; extracting histogram equalization prior information from the histogram equalized image through the prior mapping function; (2) a spatial feature transformation layer; fusing the histogram equalization prior information with the feature image of the connection graph through the spatial feature transformation layer to obtain a feature image incorporating prior information; (3) an adaptive deep neural network based on human visual characteristics; combining the low-light image and the feature image incorporating prior information as input, and decomposing it into a reflectance map and a brightness map.

[0057] On the one hand, the present invention provides a self-supervised low-light image enhancement method, specifically including the following steps:

[0058] S1: Histogram equalization processing for low-light images:

[0059] Histogram equalization is performed on the low-light image, and then a priori mapping function is used to extract the high-order feature information of the low-light image after histogram equalization.

[0060] Among them, the low-light image was brightened after histogram equalization, but it had overexposure or color cast problems; the high-order feature information of the low-light image was used as the prior information for histogram equalization.

[0061] S2: Calculate the maximum channel value of the low-light image, combine the maximum channel value with the low-light image, and use a convolutional layer to obtain the feature image in the connection graph;

[0062] S3: Using the prior information of histogram equalization and the features in the connection graph in S2 as inputs to the spatial feature transformation layer, the modulation parameter pair of the prior information of histogram equalization is learned, and the feature image in the connection graph is adaptively transformed in space based on the modulation parameter pair to obtain the feature image incorporating the prior information; the affine transformation is specifically: scaling and moving the feature image in the connection graph to perform feature transformation;

[0063] S4: As Figure 1 As shown, based on the theory of retinal cortex, an adaptive deep neural network model that conforms to the visual characteristics of the human eye is constructed; the feature image incorporating prior information is input into the adaptive deep neural network model, and the reflection map and brightness map are obtained through several convolutional layers and sigmoid layers; among them, the reflection map is used as the restored image.

[0064] More preferably, the prior mapping function is a deep convolutional neural network, which includes a five-layer structure: the first layer is a second combined layer, the middle three layers are a third combined layer, and the last layer is a 1×1 convolutional layer; wherein, the second combined layer includes a 3×3 convolution and a LeakyReLU layer; and the third combined layer includes a 1×1 convolution and a LeakyReLU layer.

[0065] More preferably, the method for the spatial feature transformation layer to obtain feature images incorporating prior information is as follows:

[0066] Two 1×1 convolutional layers are used to extract the intermediate parameters of the histogram equalization prior information, which are then multiplied with the feature image in the connection graph to obtain the first modulation parameters of the histogram equalization prior information.

[0067] Then, two 1×1 convolutional layers are used to extract the second modulation parameters of the histogram equalization prior information;

[0068] The first modulation parameter and the second modulation parameter are added together to obtain a feature image incorporating prior information.

[0069] More preferably, the method for obtaining the reflectance map and brightness map by incorporating prior information into the feature image is as follows:

[0070] The feature image incorporating prior information is input into the U-Net network to extract a deep feature image containing deeper feature information; among which, the deeper feature information includes structural and texture information of the low-light image;

[0071] A fourth combination layer is used to extract the feature image of the input low-light image; the fourth combination layer includes a 3×3 convolution and a ReLU layer.

[0072] Then, the feature image of the input image is combined with the deep feature image through a concatenation method;

[0073] Two 3×3 convolutional layers are used to further fuse the combined feature images;

[0074] A sigmoid layer is used to obtain the final reflection and brightness maps, where the reflection map is the restored image.

[0075] More preferably, the method for obtaining deep feature images containing deeper feature information is as follows:

[0076] The fifth combination layer is used to downsample the feature image that incorporates prior information, reduce the size of the feature image, and extract the deep feature image containing deeper feature information; the fifth combination layer includes two 3×3 convolutional layers and a ReLU layer;

[0077] The sixth combination layer is used to upsample the deep feature image to restore it to the size of the feature image that incorporates prior information, while retaining deeper feature information; the sixth combination layer includes two 3×3 deconvolution layers and a ReLU layer.

[0078] On the other hand, the present invention provides a self-supervised low-light image enhancement system, comprising:

[0079] The histogram equalization processing module is used to perform histogram equalization processing on low-light images; wherein, the low-light image is an image taken when the ambient light intensity is less than 1 Lux;

[0080] The prior mapping module is used to extract high-order feature information of the low-light image after histogram equalization using a mapping function, and use the high-order feature information as the prior information of histogram equalization.

[0081] The maximum channel value calculation module is used to calculate the maximum channel value of low-light images;

[0082] The image connection module is used to combine the maximum channel image and the low-light image and input them into the first combination layer to obtain the feature image in the connection graph; wherein, the first combination layer includes a 9×9 convolution and ReLU layer;

[0083] The spatial feature transformation module is used to take the histogram equalization prior information and the feature image in the connection graph as input to the spatial feature transformation layer, learn the modulation parameters of the histogram equalization prior information to adaptively perform spatial affine transformation on the feature image in the connection graph, and obtain the feature image incorporating the prior information.

[0084] An adaptive neural network module, which contains an adaptive deep neural network model, is used to input feature images incorporating prior information into the adaptive deep neural network model constructed based on the theory of retinal cortex. The model obtains a reflection map and a brightness map through several convolutional layers and sigmoid layers; wherein the reflection map is a restored image.

[0085] More preferably, the prior mapping function in the prior mapping module is a deep convolutional neural network, which includes a five-layer structure: the first layer is a second combined layer, the middle three layers are a third combined layer, and the last layer is a 1×1 convolutional layer; wherein, the second combined layer includes a 3×3 convolution and a LeakyReLU layer; and the third combined layer includes a 1×1 convolution and a LeakyReLU layer.

[0086] More preferably, the method for the spatial feature transformation layer to obtain feature images incorporating prior information is as follows:

[0087] Two 1×1 convolutional layers are used to extract the intermediate parameters of the histogram equalization prior information, which are then multiplied with the features in the connection graph to obtain the first modulation parameters of the histogram equalization prior information.

[0088] Then, two 1×1 convolutional layers are used to extract the second modulation parameters of the histogram equalization prior information;

[0089] The first modulation parameter and the second modulation parameter are added together to obtain a feature image incorporating prior information.

[0090] More preferably, the adaptive deep neural network model in the adaptive neural network module includes a U-Net network, a fourth combined layer, two 3×3 convolutional layers, a sigmoid layer, and an adder; wherein, the fourth combined layer includes a 3×3 convolutional layer and a ReLU layer;

[0091] The U-Net network receives feature images incorporating prior information to extract deep feature images containing deeper feature information; wherein, the deeper feature information includes structural and texture information of low-light images;

[0092] The fourth combination layer is used to extract the feature image of the input image from the input low-light image;

[0093] The adder is used to combine the feature image of the input image with the deep feature image through a concatenation mechanism;

[0094] The remaining two 3×3 convolutional layers further fuse the combined feature images;

[0095] A sigmoid layer is used to obtain the final reflection and brightness maps, where the reflection map is the restored image.

[0096] More preferably, the U-Net network includes a fifth combined layer and a sixth combined layer; wherein, the fifth combined layer includes two 3×3 convolutional and ReLU layers; and the sixth combined layer includes two 3×3 deconvolutional and ReLU layers.

[0097] The fifth combination layer is used to downsample the feature image incorporating prior information, reduce the size of the feature image, and extract a deep feature image containing deeper feature information.

[0098] The sixth layer is used to upsample the deep feature image, restoring it to the size of the feature image that incorporates prior information, while retaining deeper feature information.

[0099] Example

[0100] The neural network structure used in the self-supervised low-light image enhancement method of this invention is as follows: Figure 2 As shown, a 9×9 convolutional layer is first used to extract feature images from the input low-light image and its connection graph with the maximum channel value. The spatial feature transformation layer uses histogram equalization prior information as input to learn the modulation parameter pair of histogram equalization prior information, and applies affine transformation to adjust the fused feature transformation. The feature image with prior information is fed into the U-Net network, which includes two 3×3 convolutional layers, two 3×3 deconvolutional layers, and four ReLU layers. At the same time, initial features are extracted from the connection graph after feature transformation based on the 3×3 convolutional layers and ReLU layers. Then, the feature image of the input image is combined with the deep feature image from U-Net and fed into two 3×3 convolutional layers. Finally, a sigmoid layer is used to extract the reflectance map and brightness map from the feature space. The sigmoid function is used to constrain the value range of the reflectance map and brightness map to be in the range of [0,1].

[0101] Methods for extracting histogram equalization prior information from low-light images after histogram equalization processing include... Figure 3 As shown, the method of utilizing histogram equalization prior information and incorporating it into the feature image of the connectivity graph is as follows: Figure 4 As shown; to better map the histogram balance prior information, this invention uses a deep neural network as a learnable prior mapping function. The deep neural network used for the prior mapping function contains 5 convolutional layers, where the first layer is a 3×3 convolutional layer and a LeakyReLU layer, the middle three layers are 1×1 convolutional layers and LeakyReLU layers, and the last layer is a 1×1 convolutional layer. The extracted histogram balance prior information can be directly incorporated into the adaptive deep neural network model by combining it with the feature image in the connection graph. However, to better represent the histogram balance prior information, such as... Figure 4As shown, two 1×1 convolutional layers extract learnable parameters, which are then multiplied with features in the connection graph. Histogram-balanced prior information is then passed through two more 1×1 convolutional layers to obtain another learnable parameter, which is added to the multiplied features. This method allows the spatial feature transformation layer to better learn the histogram-balanced prior information. Finally, an adaptive deep neural network model outputs a reflection map as the final enhanced image. Figure 5 As shown, Figure 5 The image on the left is a low-light image. Figure 5 The image on the right is the enhanced version.

[0102] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-supervised low-light image enhancement method, characterized in that, Includes the following steps: Histogram equalization is performed on low-light images, and then a prior mapping function is used to extract high-order feature information of the low-light images after histogram equalization. The high-order feature information is used as the prior information of histogram equalization. Here, low-light images refer to images taken when the ambient light intensity is less than 1 Lux. The maximum channel value of the low-light image is calculated, and the maximum channel image is combined with the low-light image and then input into the first combination layer to obtain the feature image in the connection graph; wherein, the first combination layer includes a 9×9 convolution and ReLU layer; Using histogram equalization prior information and feature images in the connection graph as inputs to the spatial feature transformation layer, the modulation parameter pair of histogram equalization prior information is learned to adaptively perform spatial affine transformation on the feature images in the connection graph to obtain feature images incorporating prior information. The feature image incorporating prior information is input into an adaptive deep neural network model constructed based on the theory of retinal cortex, and a reflectance map and a brightness map are obtained through several convolutional layers and sigmoid layers; wherein, the reflectance map is a restored image; The method for obtaining reflectance and brightness maps by incorporating prior information into feature images is as follows: The feature image incorporating prior information is input into the U-Net network to extract a deep feature image containing deeper feature information; wherein, the deeper feature information includes structural information and texture information of the low-light image; A fourth combination layer is used to extract the feature image of the input low-light image; the fourth combination layer includes a 3×3 convolution and a ReLU layer. Then, the feature image of the input image is combined with the deep feature image through a concatenation method; Two 3×3 convolutional layers are used to fuse the combined feature images again; Based on the re-fused feature images, a sigmoid layer is used to obtain the final reflection map and brightness map, where the reflection map is the restored image; The method for obtaining deep feature images containing more advanced feature information is as follows: The fifth combination layer is used to downsample the feature image incorporating prior information, thereby reducing the size of the feature image incorporating prior information and extracting a deep feature image containing deeper feature information; the fifth combination layer includes two 3×3 convolutional layers and a ReLU layer; The sixth combination layer is used to upsample the deep feature image to restore it to the size of the feature image that incorporates prior information, while retaining deeper feature information; the sixth combination layer includes two 3×3 deconvolution layers and a ReLU layer.

2. The self-supervised low-light image enhancement method according to claim 1, characterized in that, The prior mapping function is a deep convolutional neural network, which includes a five-layer structure. The first layer is a second combined layer, the middle three layers are a third combined layer, and the last layer is a 1×1 convolutional layer. The second combined layer includes a 3×3 convolution and a LeakyReLU layer, and the third combined layer includes a 1×1 convolution and a LeakyReLU layer.

3. The self-supervised low-light image enhancement method according to claim 1 or 2, characterized in that, The method for the spatial feature transformation layer to acquire feature images incorporating prior information is as follows: Two 1×1 convolutional layers are used to extract the intermediate parameters of the histogram equalization prior information, which are then multiplied with the feature image in the connection graph to obtain the first modulation parameters of the histogram equalization prior information. Then, two 1×1 convolutional layers are used to extract the second modulation parameters for histogram equalization prior information; The first modulation parameter and the second modulation parameter are added together to obtain a feature image incorporating prior information.

4. A self-supervised low-light image enhancement system, characterized in that, include: The histogram equalization processing module is used to perform histogram equalization processing on low-light images; wherein, the low-light image is an image taken when the ambient light intensity is less than 1 Lux; The prior mapping module is used to extract high-order feature information of the low-light image after histogram equalization using a prior mapping function, and to use the high-order feature information as the prior information of histogram equalization. The maximum channel value calculation module is used to calculate the maximum channel value of low-light images; The image connection module is used to combine the maximum channel image and the low-light image and input them into the first combination layer to obtain the feature image in the connection graph; wherein, the first combination layer includes a 9×9 convolution and ReLU layer; The spatial feature transformation module is used to take the histogram equalization prior information and the feature image in the connection graph as input to the spatial feature transformation layer, learn the modulation parameter pair of the histogram equalization prior information to adaptively perform spatial affine transformation on the feature image in the connection graph, and obtain the feature image incorporating the prior information. An adaptive neural network module, which contains an adaptive deep neural network model, is used to input feature images incorporating prior information into the adaptive deep neural network model constructed based on the theory of retinal cortex. The model obtains a reflection map and a brightness map through several convolutional layers and sigmoid layers; wherein, the reflection map is a restored image. The adaptive deep neural network model in the adaptive neural network module includes a U-Net network, a fourth combined layer, two 3×3 convolutional layers, a sigmoid layer, and an adder; the fourth combined layer includes a 3×3 convolution and a ReLU layer. The U-Net network receives feature images incorporating prior information to extract deep feature images containing deeper feature information; wherein, the deeper feature information includes structural and texture information of low-light images; The fourth combination layer is used to extract the feature image of the input image from the input low-light image; The adder is used to combine the feature image of the input image with the deep feature image through a concatenation mechanism; The remaining two 3×3 convolutional layers fuse the combined feature images again; A sigmoid layer is used to obtain the final reflection map and brightness map based on the re-fused feature image, where the reflection map is the restored image; The U-Net network includes a fifth combined layer and a sixth combined layer; wherein, the fifth combined layer includes two 3×3 convolutional layers and a ReLU layer; the sixth combined layer includes two 3×3 deconvolutional layers and a ReLU layer; The fifth combination layer is used to downsample the feature image incorporating prior information, reduce the size of the feature image incorporating prior information, and extract the deep feature image containing deeper feature information. The sixth layer is used to upsample the deep feature image, restoring it to the size of the feature image that incorporates prior information, while retaining deeper feature information.

5. The self-supervised low-light image enhancement system according to claim 4, characterized in that, The prior mapping function in the prior mapping module is a deep convolutional neural network, which includes a five-layer structure. The first layer is the second combined layer, the middle three layers are the third combined layer, and the last layer is a 1×1 convolutional layer. The second combined layer includes 3×3 convolutions and LeakyReLU layers, and the third combined layer includes 1×1 convolutions and LeakyReLU layers.

6. The self-supervised low-light image enhancement system according to claim 4 or 5, characterized in that, The method for obtaining feature images incorporating prior information using a spatial feature transformation layer is as follows: Two 1×1 convolutional layers are used to extract the intermediate parameters of the histogram equalization prior information, which are then multiplied with the features in the connection graph to obtain the first modulation parameters of the histogram equalization prior information. Then, two 1×1 convolutional layers are used to extract the second modulation parameters for histogram equalization prior information; The first modulation parameter and the second modulation parameter are added together to obtain a feature image incorporating prior information.

Images