Double-layer graph reasoning method for low-light image enhancement

Abstract

The invention provides a double-layer graph reasoning method for low-light image enhancement, which comprises the following steps of: generating images with different exposure levels by using gamma correction, and calculating corresponding weight matrixes; after weighted fusion is carried out on low-light images with different exposures, a graph convolution enhancement optimization module is constructed by utilizing the characteristics of a graph convolution neural network, and graph node information is continuously updated. And the obtained de-noised image is enhanced from two dimensions of space and channel. According to the invention, gamma correction is carried out on an original low-light image to generate a series of enhanced images which highlight different regional features. In order to avoid noise caused by overexposure of the image, an image fusion weight matrix is constructed according to features of the generated image, and the image fusion weight matrix and the original image are subjected to weighted fusion.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a two-layer graph reasoning method for low-light image enhancement. Background Art

[0002] The light distribution of images collected under low illumination conditions is uneven, and most of the information of the image will be hidden when the signal-to-noise ratio (SNR) is low. Obviously, according to the characteristics of low-light images, it cannot provide useful assistance for subsequent advanced visual tasks, but will increase unnecessary difficulty, such as night target detection, action recognition, etc. Therefore, it is necessary to restore the details of low-quality images to prepare for subsequent advanced image processing tasks.

[0003] Low-light images have low contrast and are mixed with a lot of noise, so a lot of image details will be lost. The process of restoring high-quality images from low-quality images is full of uncertainty. In response to this situation, a large number of researchers have started to restore the quality of low-light images from different angles. Early work used the method of stretching pixel values ​​to restore the overall brightness of the image. Since this is a global operation, many local details will be lost. Another method is to use Retinex theory to decompose the original low-light image into illumination component and reflection component, and perform subsequent operations on each component separately. This method works well in restoring image details, but it will amplify noise and cause overexposure. Image fusion with different exposures has great advantages in restoring contrast, but it also has limitations. On the one hand, images with different exposures in the same scene are not easy to collect, and on the other hand, a lot of artificial prior knowledge is required as guidance.

[0004] In recent years, convolutional neural networks (CNNs) have been favored by a large number of scholars in the field of low-light image and video restoration due to their advantages in processing data diversity. In particular, scholars have conducted research from different perspectives in the field of image fusion with different exposures. This type of method effectively solves the illumination problem of the image, but in order to restore the illumination, the basic information of the image is lost invisibly, especially under low-light conditions, most of the information is hidden in the low signal-to-noise ratio area. Summary of the invention

[0005] According to the technical problems mentioned in the above background technology, a two-layer graph reasoning method for low-light image enhancement is provided. The present invention proposes a new image weighted fusion method, which uses gamma correction to generate images with different exposure levels and calculates the corresponding weight matrix. After weighted fusion of low-light images with different exposures, a graph convolution enhancement optimization module is constructed using the characteristics of the graph convolutional neural network to continuously update the graph node information. The denoised image obtained is enhanced from two dimensions: spatial and channel.

[0006] The technical means adopted by the present invention are as follows:

[0007] A two-layer graph reasoning method for low-light image enhancement includes the following steps:

[0008] S01: Get the original RGB low-light image I low , the given input low-light image I low Input to the Gamma-Net and WF-Net networks of the denoising module, the Gamma-Net network generates a series of parameters [λ1γ1,λ2γ2,…,λ n γ n ] and input it into the WF-Net network to perform weighted fusion on images with different exposure ratios to generate a clear image I at this stage clear ;

[0009] S02: Clear image I clear Input into the spatial information interaction module to obtain I clear The relationship between spatial nodes with similar image content and similar regions with similar features is constructed, and information is exchanged. The graph convolutional neural network is used to update the graph nodes to obtain the final result image I at this stage. s ;

[0010] S03: Clear image I clear Input into the channel information interaction module to obtain the clear image I at this stage clear The relationship between node features with similar channel dimension content is captured, the relationship between similar texture information of related channel node features is captured and information is exchanged, and the graph convolutional neural network is used to update the graph nodes to obtain the final result image I of this stage. c ;

[0011] S04: The results of the spatial information interaction module I s and channel information interaction module I c Input into the feature fusion module, use the squeeze operation and excitation operation to obtain the feature information between channels, splice the features, and input into the decoder to generate the final enhancement result.

[0012] Furthermore, the denoising module includes: a generation stage and a weighted fusion stage;

[0013] In the generation stage, the low-light image I low As input, a series of parameters [λ1γ 1 ,λ1γ 1 ,…,λ n γ n] and add these parameters to the low-light image I low Generate images with different exposure levels [I1,I2,…,I n ];

[0014] In the weighted fusion stage, the network WF-Net generates the corresponding weight matrix, combines the feature information of the different exposure images with the original image, and realizes denoising. The operation of the denoising module is:

[0015] I clear =DNM(I)=[I low , I1,…,I i ,…,I n ];

[0016] Among them, DNM(·) represents the denoising module operation, I i represents the generated i-th exposure intensity image, which has a low-light image I low and the exposure parameter [λ1γ 1 ,λ1γ 1 ,…,λ n γ n ] combined to obtain, 1≤i≤n; I clear It represents the final result image generated; the output of the Gamma-Net network is a set of corresponding pixel-level parameters.

[0017] Furthermore, the Gamma-Net network includes: two convolutional layers and a ReLU activation function, an average pooling function, and a sigmoid activation function; in order to highlight the feature information between different regions, an iterative method is used to generate different exposure images;

[0018]

[0019] Among them, I n-1 represents the exposure image generated for the n-1th time; I0 represents the original low-light image; λ and γ are the parameters output by Gama-Net; the generated γ is the result of the three channels of RGB in the image, compared with directly generating a unified γ value to restore the overall color difference.

[0020] Furthermore, the weighted fusion stage adopts a weighted fusion method to fuse feature information between different regions;

[0021] Two 3x3 convolutions are used for feature extraction and normalized with the sigmoid function. The input is the fused image obtained by the i-th exposure and the i-1-th exposure, and the output is two different weight matrices. The weight matrix represents the health status of each pixel, which is expressed as:

[0022] ω i-1 ,ω i=WF(I i-1 , I i );

[0023]

[0024] Among them, WF() represents the network WF-Net that generates the weight matrix; ω i Indicates that i The corresponding weight graph, ω i-1 Indicates that i-1 The corresponding weight map; c∈{R,G,B} represents the red (R), green (G) and blue (B) channels in the low-light image.

[0025] Furthermore, in step S02, the spatial information interaction module is composed of three convolutions and a graph convolution structure, wherein two convolutions are used to restore and expand the input feature X and the output Y, and one convolution is used to generate a double projection B and between the coordinate and the potential interaction space. Graph convolutional networks are used to infer global information based on the interaction space and the adjacency matrix of channel nodes.

[0026] Furthermore, the spatial information interaction module uses spatial information interaction to obtain the relationship between similar regions of related spatial node features and to perform information interaction. Given an image X s ∈R H×W×C , where H×W represents the size of the feature map and C represents the feature dimension; the input image first undergoes two 1x1 convolutions φ(x) and θ(x), and then φ(x) is used to reduce the dimension of X to obtain X∈R L×C φ(x) in the transpose of X generates a N×L The graph projection matrix φ(x) of the same dimension uses N nodes to represent the constructed graph information:

[0027] V = g(X) = B × X;

[0028] Where V∈R N×C represents the new reconstructed graph node features, B is the graph node projection matrix, and g(X) is the linear projection operation;

[0029] In order to realize the information interaction of spatial graph nodes, it is necessary to construct an adjacency matrix A s ∈N×N, each node stores the updated node feature state; by continuously updating the state of each node, the context information between different nodes is inferred.

[0030] Furthermore, the graph node interactive update operation is:

[0031]

[0032] Among them, W is the graph convolution weight, A s ∈N×N is the constructed adjacency matrix, V∈R N×C Represents the new reconstructed graph node features; at this time, the graph node dimension is N×C, and the graph convolution is:

[0033] SCI(S)=Conv1D(Conv1D(S) T ) T ;

[0034] Where S represents graph node information, Conv1D(·) represents 1x1 convolution, and SCI(·) represents the spatial information interaction module;

[0035] Restore the feature information from the graph node to the feature domain, use θ(x) to transpose X, and generate a new projection matrix And project the graph space back to the original feature space; the operation of projecting the graph space back to the original feature space is as follows:

[0036]

[0037] The obtained feature information Y is in r L×C In the above example, we reshape Y into r H×W×C , at this time, Y is the residual image of the input image X; add Y to the original image X to get the final feature map I s .

[0038] Furthermore, the channel information interaction module uses the channel information interaction to capture the relationship between similar texture information of related channel node features and to perform information interaction; an image x is given to the channel information interaction module. c ∈r H ×W×C , first perform a 1x1 convolution φ(x) on the image, φ(x) uses a convolution operation to change the feature dimension X of the image c ∈R H ×W×N , N represents the number of nodes; by introducing the average pooling operation to reduce X c The size of the feature;

[0039] X' c =Avgpool(X c );

[0040] Among them, x' c ∈r Q×Q×N ,Q×Q represents the recovered feature size;

[0041] Then for x′ c Reposition and transform to get X c ∈R N×S , then we get Xc Represents the node features projected on the channel dimension. The feature information of different nodes interacts through GCN to obtain context information; then create a c ∈R N×S The corresponding adjacency matrix a c ∈N×N is used to store the updated node feature state for information update between different nodes; the node update principle is as follows:

[0042]

[0043] Among them, A c ∈N×N represents the created adjacency matrix, V represents the new reconstructed graph node features, and W represents the graph convolution weights; by transposing X with θ(x), a new projection matrix B∈R is generated L×N , and then project the graph space back to the original feature space to obtain Y∈R N×L ;

[0044] Use a convolution operation on Y∈R H×W×N Reshape, change the feature dimension of Y from N to C, and sum it with the original image X to get the final feature map I c .

[0045] Furthermore, in S03, the channel attention mechanism uses a squeeze operation and an excitation operation to obtain feature information between channels; the fused feature map is put into the decoder D(.) in the image domain and restored back to the image domain, thereby obtaining the final enhancement result. I s and I c They represent the processed feature information respectively;

[0046] I s and I c The spatial self-attention operation and channel self-attention operation are used to fuse the acquired feature maps; then the acquired feature information is converted to the image domain to obtain the final enhancement result:

[0047]

[0048] Among them, F(·) represents a comprehensive operation, F(·) includes the feature map I s and I c The fusion operation of , as well as the spatial and channel attention operations; D(·) represents the decoder operation, which converts the obtained feature information into the image domain, is the enhanced image obtained.

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

[0050] 1. The present invention performs gamma correction on the original low-light image to generate a series of enhanced images that highlight the features of different regions. In order to avoid the noise caused by overexposure of the image, an image fusion weight matrix is ​​constructed according to the features of the generated image, and it is weightedly fused with the original image.

[0051] 2. This paper proposes a graph convolution-guided structured texture enhancement network that can enhance the global structural features and local texture features of low-light images in both spatial and channel dimensions. In the spatial dimension, multiple anchor points are defined, and the similarity of pixel features is used to aggregate graph nodes to reduce the propagation of noise. In the channel dimension, a feature map with good performance is used to guide the enhancement of fuzzy feature information between the same scene. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0053] Figure 1 Schematic diagram of the architecture of the WFGNN model of the present invention.

[0054] Figure 2 The principle diagram of images with different exposure levels is generated using Gama-Net for the present invention.

[0055] Figure 3 It is a schematic diagram of the structure of spatial information interaction and channel information interaction of the present invention.

[0056] Figure 4 The figures are qualitative comparison and contrast diagrams of the present invention on the LOL synthetic data set, wherein Figure (a) 1, Figure (a) 2, and Figure (a) 3 are original low-light images; Figure (b) 1, Figure (b) 2, and Figure (b) 3 are schematic diagrams after the method LIME is enhanced; Figure (c) 1, Figure (c) 2, and Figure (c) 3 are schematic diagrams after the method Retinex-Net is enhanced; Figure (d) 1, Figure (d) 2, and Figure (d) 3 are schematic diagrams after the method MBLLEN is enhanced; Figure (e) 1, Figure (e) 2, and Figure (e) 3 are schematic diagrams after the method SCI is enhanced; Figure (f) 1, Figure (f) 2, and Figure (f) 3 are schematic diagrams after the method URetinex is enhanced; Figure (g) 1, Figure (g) 2, and Figure (g) 3 are schematic diagrams after the method Semantic is enhanced; Figure (h) 1, Figure (h) 2, and Figure (h) 3 are schematic diagrams after the present invention is enhanced. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0058] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0059] like Figure 1-4 As shown, the present invention provides a two-layer graph reasoning method for low-light image enhancement, comprising the following steps:

[0060] S01: Get the original RGB low-light image I low , the given input low-light image I low Input to the Gamma-Net and WF-Net networks of the denoising module, the Gamma-Net network generates a series of parameters [λ1γ1,λ2γ2,…,λ n γ n ] and input it into the WF-Net network to perform weighted fusion on images with different exposure ratios to generate a clear image I at this stage clear ;

[0061] S02: Clear image I clear Input into the spatial information interaction module to obtain I clear The relationship between spatial nodes with similar image content and similar regions with similar features is constructed, and information is exchanged. The graph convolutional neural network is used to update the graph nodes to obtain the final result image I at this stage. s ;

[0062] S03: Clear image I clear Input into the channel information interaction module to obtain the clear image I at this stage clearThe relationship between node features with similar channel dimension content is captured, the relationship between similar texture information of related channel node features is captured and information is exchanged, and the graph convolutional neural network is used to update the graph nodes to obtain the final result image I of this stage. c ;

[0063] S04: The results of the spatial information interaction module I s and channel information interaction module I c Input into the feature fusion module, use the squeeze operation and excitation operation to obtain the feature information between channels, splice the features, and input into the decoder to generate the final enhancement result.

[0064] In the present application, the denoising module includes: a generation stage and a weighted fusion stage;

[0065] In the generation stage, the low-light image I low As input, a series of parameters [λ1γ 1 ,λ1γ 1 ,…,λ n γ n ] and add these parameters to the low-light image I low Generate images with different exposure levels [I1,I2,…,I n ];

[0066] In the weighted fusion stage, the network WF-Net generates the corresponding weight matrix, combines the feature information of the different exposure images with the original image, and realizes denoising. The operation of the denoising module is:

[0067] I clear =DNM(I)=[I low , I1,…,I i ,…,I n ];

[0068] Among them, DNM(·) represents the denoising module operation, I i represents the generated i-th exposure intensity image, which has a low-light image I low and the exposure parameter [λ1γ 1 ,λ1γ 1 ,…,λ n γ n ] combined to obtain, 1≤i≤n; I clear Represents the final result image generated; the output of the Gamma-Net network is a set of corresponding pixel-level parameters.

[0069] Preferably, in the present application, the Gamma-Net network includes: two convolutional layers and a ReLU activation function, an average pooling function, and a sigmoid activation function; in order to highlight the feature information between different regions, an iterative method is used to generate different exposure images;

[0070]

[0071] Among them, I n-1 represents the exposure image generated for the n-1th time; I0 represents the original low-light image; λ and γ are the parameters output by Gamma-Net; the generated γ is the result of the three channels of RGB in the image, compared with directly generating a unified γ value to restore the overall color difference.

[0072] As a preferred implementation, the weighted fusion stage uses a weighted fusion method to fuse feature information between different regions;

[0073] Two 3x3 convolutions are used for feature extraction and normalized with the sigmoid function. The input is the fused image obtained by the i-th exposure and the i-1-th exposure, and the output is two different weight matrices. The weight matrix represents the health status of each pixel, which is expressed as:

[0074] ω i-1 ,ω i =WF(I i-1 , I i );

[0075]

[0076] Among them, WF() represents the network WF-Net that generates the weight matrix; ω i Indicates that i The corresponding weight graph, ω i-1 Indicates that i-1 The corresponding weight map; c∈{R,G,B} represents the red (R), green (G) and blue (B) channels in the low-light image.

[0077] As a preferred implementation, in step S02, the spatial information interaction module is composed of three convolutions and one graph convolution structure, wherein two convolutions are used to restore and expand the input feature X and the output Y, and one convolution is used to generate a double projection B and between the coordinates and the potential interaction space. Graph convolutional networks are used to infer global information based on the interaction space and the adjacency matrix of channel nodes.

[0078] In this application, as a preferred implementation, the spatial information interaction module uses spatial information interaction to obtain the relationship between similar regions of related spatial node features and to perform information interaction. Given an image X s ∈R H ×W×C , where H×W represents the size of the feature map and C represents the feature dimension; the input image first undergoes two 1x1 convolutions φ(x) and θ(x), and then φ(x) is used to reduce the dimension of X to obtain X∈R L×C φ(x) in the transpose of X generates a N×L The graph projection matrix φ(x) of the same dimension uses N nodes to represent the constructed graph information:

[0079] V = g(X) = B × X;

[0080] Where V∈R N×C represents the new reconstructed graph node features, B is the graph node projection matrix, and g(X) is the linear projection operation;

[0081] In order to realize the information interaction of spatial graph nodes, it is necessary to construct an adjacency matrix A s ∈N×N, each node stores the updated node feature state; by continuously updating the state of each node, the context information between different nodes is inferred.

[0082] Preferably, the graph node interactive update operation is:

[0083]

[0084] Among them, W is the graph convolution weight, A s ∈N×N is the constructed adjacency matrix, V∈R N×C Represents the new reconstructed graph node features; at this time, the graph node dimension is N×C, and the graph convolution is:

[0085] SCI(S)=Conv1D(Conv1D(S) T ) T ;

[0086] Where S represents graph node information, Conv1D(·) represents 1x1 convolution, and SCI(·) represents the spatial information interaction module;

[0087] Restore the feature information from the graph node to the feature domain, use θ(x) to transpose X, and generate a new projection matrix And project the graph space back to the original feature space; the operation of projecting the graph space back to the original feature space is as follows:

[0088]

[0089] The obtained feature information Y is in r L×C In the above example, we reshape Y into r H×W×C , at this time, Y is the residual image of the input image X; add Y to the original image X to get the final feature map I s .

[0090] Preferably, the channel information interaction module uses the channel information interaction to capture the relationship between similar texture information of related channel node features and to perform information interaction; an image X is given to the channel information interaction module. c ∈R H ×W×C , first perform a 1x1 convolution φ(x) on the image, φ(x) uses a convolution operation to change the feature dimension X of the image c ∈R H ×W×N , N represents the number of nodes; by introducing the average pooling operation to reduce X c The size of the feature;

[0091] X' c =Avgpool(X c );

[0092] Among them, X' c ∈R Q×Q×N ,Q×Q represents the recovered feature size;

[0093] Then for X′ c Reposition and transform to get X c ∈R N×S , then we get X c Represents the node features projected on the channel dimension. The feature information of different nodes interacts through GCN to obtain context information; then create a c ∈R N×S The corresponding adjacency matrix A c ∈N×N is used to store the updated node feature state for information update between different nodes; the node update principle is as follows:

[0094]

[0095] Among them, A c ∈N×N represents the created adjacency matrix, V represents the new reconstructed graph node features, and W represents the graph convolution weights; by transposing X with θ(x), a new projection matrix B∈R is generated L×N , and then project the graph space back to the original feature space to obtain Y∈R N×L ;

[0096] Use a convolution operation on Y∈R H×W×NReshape, change the feature dimension of Y from N to C, and sum it with the original image X to get the final feature map I c .

[0097] As a preferred implementation, in S03, the channel attention mechanism uses a squeeze operation and an excitation operation to obtain feature information between channels; the fused feature map is put into the decoder D(.) in the image domain and restored back to the image domain, thereby obtaining the final enhancement result. I s and I c They represent the processed feature information respectively;

[0098] I s and I c The spatial self-attention operation and channel self-attention operation are used to fuse the acquired feature maps; then the acquired feature information is converted to the image domain to obtain the final enhancement result:

[0099]

[0100] Among them, F(·) represents a comprehensive operation, F(·) includes the feature map I s and I c The fusion operation of , as well as the spatial and channel attention operations; D(·) represents the decoder operation, which converts the obtained feature information into the image domain, is the enhanced image obtained.

[0101] Example:

[0102] In the present invention, the training data is cropped into 224x224 pixel blocks to obtain paired training sample pairs, and the batch size is set to 8, and the initial learning rate is set to 0.0001. The initial learning rate is optimized using the Adam optimizer, where β1 and β2 are taken as 0.9 and 0.999 respectively. When training to 400 epochs, the learning rate drops to 0.5 times the original. All experiments are performed on an NVIDIA 2080 GPU under the PyTorch framework.

[0103] Quantitative comparison on LOL synthetic dataset

[0104] LIME Retinex-net MBLLEN SCI URetinex Semantic Ours PSNR 16.769 18.530 17.902 22.617 21.328 19.367 23.475 SSIM 0.5696 0.8764 0.7152 0.8971 0.8342 0.7673 0.913

[0105] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A two-layer graph reasoning method for low-light image enhancement, characterized in that: The following steps are involved: S01: Get the original RGB low-light image I low , the given input low-light image I low Input to the Gamma-Net and WF-Net networks of the denoising module, the Gamma-Net network generates a series of parameters [λ1γ1,λ2γ2,…,λ n γ n ] and input it into the WF-Net network to perform weighted fusion on images with different exposure ratios to generate a clear image I at this stage clear ; S02: Clear image I clear Input into the spatial information interaction module to obtain I clear The relationship between spatial nodes with similar image content and similar regions with similar features is constructed, and information is exchanged. The graph convolutional neural network is used to update the graph nodes to obtain the final result image I at this stage. s ; S03: Clear image I clear Input into the channel information interaction module to obtain the clear image I at this stage clear The relationship between node features with similar channel dimension content is captured, the relationship between similar texture information of related channel node features is captured and information is exchanged, and the graph convolutional neural network is used to update the graph nodes to obtain the final result image I of this stage. c ; S04: The results of the spatial information interaction module I s and channel information interaction module I c Input into the feature fusion module, use the squeeze operation and excitation operation to obtain the feature information between channels, splice the features, and input into the decoder to generate the final enhancement result.

2. The two-layer graph reasoning method for low-light image enhancement according to claim 1, characterized in that: The denoising module includes: a generation stage and a weighted fusion stage; In the generation stage, the low-light image I low As input, a series of parameters [λ1γ 1 ,λ1γ 1 ,…,λ n γ n ] and add these parameters to the low-light image I low Generate images with different exposure levels [I1,I2,…,I n ]; In the weighted fusion stage, the network WF-Net generates the corresponding weight matrix, combines the feature information of the different exposure images with the original image, and realizes denoising. The operation of the denoising module is: I clear =DNM(I)=[I low ,I1,…,I i ,…,I n ]; Among them, DNM(·) represents the denoising module operation, I i represents the generated i-th exposure intensity image, which has a low-light image I low and the exposure parameter [λ1γ 1 ,λ1γ 1 ,…,λ n γ n ] combined to obtain, 1≤i≤n; I clear Represents the final result image generated; the output of the Gamma-Net network is a set of corresponding pixel-level parameters.

3. The two-layer graph reasoning method for low-light image enhancement according to claim 2, characterized in that: The Gamma-Net network includes: two convolutional layers and ReLU activation function, average pooling function and sigmoid activation function; in order to highlight the feature information between different regions, an iterative method is used to generate different exposure images; Among them, I n-1 represents the exposure image generated for the n-1th time; I0 represents the original low-light image; λ and γ are the parameters output by Gama-Net; the generated γ is the result of the three channels of RGB in the image, compared with directly generating a unified γ value to restore the overall color difference.

4. The two-layer graph reasoning method for low-light image enhancement according to claim 2, characterized in that: The weighted fusion stage adopts a weighted fusion method to fuse feature information between different regions; Use two 3x3 convolutions for feature extraction and combine them with the sigmoid function for normalization. The input is the fused image obtained by the i-th exposure and the i-1-th exposure, and the output is two different weight matrices. For the weight matrix, which represents the health status of each pixel, it is expressed as: oh i-1 ,oh i =WF(I i-1 ,I i ); Where WF(·) represents the network WF-Net that generates the weight matrix; ω i Indicates that i The corresponding weight graph, ω i-1 Indicates that i-1 The corresponding weight map; c∈{R,G,B} represents the red (R), green (G) and blue (B) channels in the low-light image.

5. The two-layer graph reasoning method for low-light image enhancement according to claim 2, characterized in that: In step S02, the spatial information interaction module consists of three convolutions and a graph convolution structure, two of which are used to restore and expand the input feature X and output Y, and one convolution is used to generate a double projection B and between the coordinates and the potential interaction space. Graph convolutional networks are used to infer global information based on the interaction space and the adjacency matrix of channel nodes.

6. The two-layer graph reasoning method for low-light image enhancement according to claim 1, characterized in that: The spatial information interaction module uses spatial information interaction to obtain the relationship between similar regions of related spatial node features and to perform information interaction. Given an image X s ∈R H×W×C , where H×W represents the size of the feature map and C represents the feature dimension; The input image first undergoes two 1x1 convolutions φ(x) and θ(x), and then φ(x) is used to reduce the dimension of X to obtain X∈R L×C φ(x) in the transpose of X generates a N×L The graph projection matrix φ(x) of the same dimension uses N nodes to represent the constructed graph information: V = g(X) = B × X; Where V∈R N×C represents the new reconstructed graph node features, B is the graph node projection matrix, and g(X) is the linear projection operation; In order to realize the information interaction of spatial graph nodes, it is necessary to construct an adjacency matrix A s ∈N×N, each node stores the updated node feature state; By continuously updating the state of each node, the contextual information between different nodes is inferred.

7. The two-layer graph reasoning method for low-light image enhancement according to claim 1, characterized in that: The graph node interactive update operation is: Among them, W is the graph convolution weight, A s ∈N×N is the constructed adjacency matrix, V∈R N×C Represents the new reconstructed graph node features; at this time, the graph node dimension is N×C, and the graph convolution is: SCI(S)=Conv1D(Conv1D(S) T ) T ; Where S represents graph node information, Conv1D(·) represents 1x1 convolution, and SCI(·) represents the spatial information interaction module; Restore the feature information from the graph node to the feature domain, use θ(x) to transpose X, and generate a new projection matrix And project the graph space back to the original feature space; the operation of projecting the graph space back to the original feature space is as follows: The obtained feature information Y is in R L×C In the above example, we reshape Y into r H×W×C , at this time, Y is the residual image of the input image X; add Y to the original image X to get the final feature map I s .

8. The two-layer graph reasoning method for low-light image enhancement according to claim 1, characterized in that: The channel information interaction module uses the channel information interaction to capture the relationship between similar texture information of related channel node features and to perform information interaction. Give the channel information interaction module an image X c ∈R H×W×C , first perform a 1x1 convolution φ(x) on the image, φ(x) uses a convolution operation to change the feature dimension X of the image c ∈R H×W×N , N represents the number of nodes; by introducing the average pooling operation to reduce X c The size of the feature; X' c =Avgpool(X c ); Among them, X' c ∈R Q×Q×N ,Q×Q represents the recovered feature size; Then for X′ c Reposition and transform to get X c ∈R N×S , then we get X c Represents the node features projected on the channel dimension. The feature information of different nodes interacts through GCN to obtain context information; then create a c ∈R N ×S The corresponding adjacency matrix a c ∈N×N is used to store the updated node feature state for information update between different nodes; the node update principle is as follows: Among them, A c ∈N×N represents the created adjacency matrix, V represents the new reconstructed graph node features, and W represents the graph convolution weights; by transposing X with θ(x), a new projection matrix B∈R is generated L×N , and then project the graph space back to the original feature space to obtain Y∈R N×L ; Use a convolution operation on Y∈R H×W×N Reshape, change the feature dimension of Y from N to C, and sum it with the original image X to get the final feature map I c .

9. The two-layer graph reasoning method for low-light image enhancement according to claim 1, characterized in that: In S03, the channel attention mechanism uses squeezing operation and excitation operation to obtain feature information between channels; the fused feature map is put into the decoder D(.) in the image domain and restored back to the image domain to obtain the final enhancement result. I s and I c They represent the processed feature information respectively; I s and I c The spatial self-attention operation and channel self-attention operation are used to fuse the acquired feature maps; then the acquired feature information is converted to the image domain to obtain the final enhancement result: Among them, F(·) represents a comprehensive operation, F(·) includes the feature map I s and I c The fusion operation of , as well as the spatial and channel attention operations; D(·) represents the decoder operation, which converts the obtained feature information into the image domain, is the enhanced image obtained.

Images