A two-layer graph reasoning method for low-light image enhancement

Abstract

The application provides a double-layer graph reasoning method for low-light image enhancement, utilizes gamma correction to generate images of different exposure levels, and calculates a corresponding weight matrix. After the low-light images of different exposures are weighted and fused, a graph convolution enhancement optimization module is constructed by utilizing the characteristics of the graph convolution neural network, and the graph node information is continuously updated. The obtained denoised image is enhanced from two dimensions of space and channel. The original low-light image is subjected to gamma correction in the application, and a series of enhanced images highlighting different regional features are generated. In order to avoid noise caused by overexposure of the image, an image fusion weight matrix is constructed according to the features of the generated image, and the original image is weighted and fused.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular, relates to a double-layer graph reasoning method for low-light image enhancement. BACKGROUND

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

[0003] The low-illumination image has low contrast, and a large number of noise points are mixed in the image, so that a lot of image details are lost. The process of restoring high-quality images from low-quality images is full of uncertainty. In view of this situation, a large number of researchers start from different angles to restore the quality of low-light images. The early work adopts the method of stretching pixel values to restore the overall brightness of the image, and since this is a global operation, a lot of local details will be lost. Another method is to use the Retinex theory to decompose the original low-illumination image into illumination component and reflection component, and to perform subsequent operations on each component respectively. This method has good effect in restoring image details, but it will amplify noise and cause overexposure. Image fusion of different exposure has great advantage in restoring contrast, but also has limitations. On the one hand, it is not easy to collect images of different exposure in the same scene, and on the other hand, a large amount of manual prior knowledge is needed as guidance.

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

[0005] According to the technical problems mentioned in the above background art, a double-layer graph reasoning method for low-light image enhancement is provided. The present application proposes a new image weighted fusion method, which uses gamma correction to generate images of different exposure levels and calculates the corresponding weight matrix. After weighted fusion of low-light images of different exposure, a graph convolution neural network is used to construct a graph convolution enhancement optimization module to continuously update the graph node information. The denoising image obtained is enhanced from two dimensions of space and channel.

[0006] The technical means employed in this invention are as follows:

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

[0008] S01: Acquire the original RGB low-light image I low Given a low-light image I low The input is fed into 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 The images are then input into the WF-Net network, where they are weighted and fused to generate the clear image I for this stage. clear ;

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

[0010] S03: Clear the image at this stage I clear The image I is input into the channel information interaction module to obtain the clear image at this stage. clear This involves identifying relationships between similar node features across different channel dimensions, capturing the connections between similar texture information of related channel node features, and enabling information exchange. A graph convolutional neural network is then used to update the graph nodes, resulting in the final image I for this stage. c ;

[0011] S04: Transfer the results from the spatial information interaction module I s Channel information interaction module I c The input is fed into the feature fusion module, where squeezing and excitation operations are used to obtain inter-channel feature information. The features are then concatenated and fed into the decoder to generate the final enhanced result.

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

[0013] During the generation stage, low-light image I low As input, a series of parameters [λ1γ] are generated through the Gamma-Net network. 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 WF-Net network generates a corresponding weight matrix, combining the feature information of images with different exposures with the original image to achieve denoising. The denoising module operates as follows:

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

[0016] Where DNM(·) represents the denoising module operation, I i This represents the generated exposure intensity image I, which has a low-light image. low and exposure parameters [λ1γ] 1 ,λ1γ 1 ,…,λ n γ n Combining these, we get 1 ≤ i ≤ n; I clear This represents the final 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 In-1 represents the exposure image generated; 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 RGB channels of the image, compared with directly generating a uniform γ value to restore the overall color difference.

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

[0021] Feature extraction is performed using two 3x3 convolutions, followed by sigmoid normalization. The input is the fused image obtained from the i-th and (i-1)-th exposures, and the output is two distinct weight matrices. The weight matrix represents the health state of each pixel, denoted as:

[0022] ω i-1 ω i= WF(I i-1 , I i );

[0023]

[0024] where WF() denotes the network WF-Net that generates the weight matrix; ω i denotes the weight map corresponding to I i , ω i-1 denotes the weight map corresponding to I i-1 , and c ∈ {R, G, B} denotes the red (R) green (G) blue (B) three channels in the low-light image.

[0025] Further, the spatial information interaction module in the step S02 is composed of three convolutions and one graph convolution structure, wherein two convolutions are used for restoring and expanding the input feature X and the output Y, and one convolution is used for generating double projections B and The graph convolution network is used for global information inference based on the adjacency matrix of the interaction space and the channel node.

[0026] Further, the purpose of using the spatial information interaction module is to obtain the relationship of the similar regions of the related spatial node features and to perform information interaction; given an image X s ∈ R H×W×C , wherein HxW denotes the size of the feature map, and C denotes the feature dimension; the input image is first subjected to two 1x1 convolutions φ(x) and θ(x), then φ(x) is used to reduce and reshape X to obtain φ(x) in R L×C , and φ(x) is used to transpose X to generate a graph projection matrix φ(x) with the same dimension as B ∈ R N×L , and N nodes are used to represent the constructed graph information:

[0027] V = g(X) = BxX;

[0028] where V ∈ R N×C denotes the new reconstructed graph node feature, B is the graph node projection matrix, and g(X) is a linear projection operation.

[0029] In order to realize the information interaction of the spatial graph nodes, an adjacency matrix A s ∈ NxN is constructed, and 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] Further, the graph node interaction update operation is:

[0031]

[0032] where W is the graph convolution weight, A s ∈N×N is the adjacency matrix constructed, V∈R N×C represents the new reconstructed graph node feature; 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 the graph node information, Conv1D(·) represents the 1x1 convolution, and SCI(·) represents the spatial information interaction module;

[0035] The feature information is restored from the graph node to the feature domain, and θ(x) is used to transpose X to 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 , and Y is reshaped to r H×W×C At this time, Y is the residual graph of the input image X; adding Y and the original image X obtains the final feature map I s .

[0038] Further, the purpose of using the channel information interaction module to interact channel information is to capture the relationship between similar texture information of related channel node features and interact information; the channel information interaction module is given an image x c ∈r H ×W×C First, 1x1 convolution φ(x) is performed on the image, and φ(x) uses convolution operation to change the feature dimension X c ∈R H ×W×N of the image; by introducing an average pooling operation, the size of X c feature is reduced;

[0039] X’ c =Avgpool(X c );

[0040] where x’ c ∈r Q×Q×N , and QxQ represents the newly obtained feature size;

[0041] Then, x′ c is repositioned and transformed to obtain X c ∈R N×S At this time, the obtained Xc represent the node features projected on the channel dimension, and the feature information of different nodes is interacted through the GCN to obtain context information; then a and X c ∈R N×S The corresponding adjacency matrix a c ∈N×N is used to store the updated node feature state for information updating between different nodes; the node updating principle is as follows:

[0042]

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

[0044] Y∈R H×W×N is reshaped using one convolution operation, and the feature dimension of Y is changed from N to C; and the original image X is summed to obtain the final feature mapping I c .

[0045] Further, 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 image domain decoder D(.) and restored back to the image domain, so as to obtain the final enhancement result I s and I c respectively represent the processed feature information;

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

[0047]

[0048] Wherein, F(·) represents a comprehensive operation, F(·) includes fusion operation of feature maps I s and I c , and spatial and channel attention operation; D(·) represents a decoder operation, which converts the obtained feature information to the image domain, is the obtained enhanced image.

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

[0050] 1. The original low-light image is gamma corrected to generate a series of enhanced images highlighting different regional features. In order to avoid 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 fused with the original image.

[0051] 2. The application proposes a structure texture enhancement network guided by graph convolution, which can enhance the global structure features and local texture features of low-light images in spatial and channel dimensions. In the spatial dimension, a plurality of anchor points are defined, and the similarity of pixel features is used for graph node aggregation to reduce the propagation of noise. In the channel dimension, the feature map with good performance is used to guide the enhancement of blurred feature information between the same scene. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0053] Figure 1 The schematic diagram of the architecture of the WFGNN model of the application.

[0054] Figure 2 The principle diagram of generating images with different exposure levels using Gama-Net.

[0055] Figure 3 The schematic diagram of the structure of the interaction of the image space information and the channel information of the application.

[0056] Figure 4 The qualitative comparison contrast diagram of the LOL synthesis data set of the application, wherein, Fig. (a) 1, Fig. (a) 2, Fig. (a) 3 are original low-light images; Fig. (b) 1, Fig. (b) 2, Fig. (b) 3 are schematic diagrams after enhancement by the method LIME; Fig. (c) 1, Fig. (c) 2, Fig. (c) 3 are schematic diagrams after enhancement by the method Retinex-Net; Fig. (d) 1, Fig. (d) 2, Fig. (d) 3 are schematic diagrams after enhancement by the method MBLLEN; Fig. (e) 1, Fig. (e) 2, Fig. (e) 3 are schematic diagrams after enhancement by the method SCI; Fig. (f) 1, Fig. (f) 2, Fig. (f) 3 are schematic diagrams after enhancement by the method URetinex; Fig. (g) 1, Fig. (g) 2, Fig. (g) 3 are schematic diagrams after enhancement by the method Semantic; Fig. (h) 1, Fig. (h) 2, Fig. (h) 3 are schematic diagrams after enhancement by the application. DETAILED DESCRIPTION

[0057] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application, so that those skilled in the art can better understand the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of the present application.

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

[0059] As shown in Figures 1-4 , the present application provides a double-layer graph reasoning method for low-light image enhancement, comprising the following steps:

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

[0061] S02: inputting the clear image I clear at this stage into the spatial information interaction module, obtaining the spatial node feature similar area relationship of the image content I clear , and performing information interaction, updating the graph nodes using a graph convolutional neural network to obtain the final result image I s at this stage;

[0062] S03: inputting the clear image I clear at this stage into the channel information interaction module, obtaining the clear image I clearThe relationship of the node features with similar channel dimension content is captured, the relationship between the similar texture information of the relevant channel node features is captured and information interaction is performed, the graph node is updated using a graph convolutional neural network, and a final result image I of this stage is obtained c ;

[0063] S04: The result I of the spatial information interaction module s and the channel information interaction module I c is input into a feature fusion module, a squeezing operation and an excitation operation are used to obtain the feature information between channels, the features are spliced, and the final enhanced result is generated by inputting the features into a decoder

[0064] In the present application, the denoising module comprises a generation stage and a weighted fusion stage.

[0065] In the generation stage, a low-light image I low is input, a series of parameters [λ1γ 1 ,λ1γ 1 ,…,λ n γ n ] are generated by a gamma-net network, and the parameters are added to the low-light image I low to generate images with different exposure levels [I1,I2,…,I n ];

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

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

[0068] Wherein, DNM(·) represents the denoising module operation, I i represents the i-th generated exposure intensity image, which is obtained by combining the low-light image I low and the exposure parameters [λ1γ 1 ,λ1γ 1 ,…,λ n γ n ], 1≤i≤n; I clear represents the generated final result image; the Gamma-Net network outputs a corresponding set of pixel-level parameters.

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

[0070]

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

[0072] As a preferred embodiment, the weighted fusion stage adopts a weighted fusion method to fuse the feature information between different regions.

[0073] Two 3x3 convolutions are used for feature extraction, and a sigmod function is used for normalization, the input is the fusion image obtained by the i-th exposure and the i-1th exposure, and the output is two different weight matrices; for the weight matrix, it represents the health status of each pixel, and is represented as:

[0074] ω i-1 , ω i = WF(I i-1 , I i );

[0075]

[0076] wherein WF() represents the network WF-Net for generating the weight matrix; ω i represents the weight map corresponding to I i , ω i-1 represents the weight map corresponding to I i-1 ; c∈{R,G,B} represents the red (R) green (G) blue (B) three channels in the low-light image.

[0077] As a preferred embodiment, in the step S02, the spatial information interaction module is composed of three convolutions and a graph convolution structure, wherein two convolutions are used for restoring and expanding the input feature X and the output Y, and one convolution is used for generating double projections B and The graph convolution network is used for global information inference based on the adjacency matrix of the interaction space and the channel node.

[0078] In the present application, as a preferred embodiment, the purpose of the spatial information interaction module using spatial information interaction is to obtain the relationship of the relevant spatial node feature similar region and perform information interaction; given an image X s ∈R H ×W×C , where HxW represents the size of the feature map, and C represents the feature dimension; the input image is first subjected to twice 1x1 convolution φ(x), θ(x), then X is reduced and reshaped by φ(x) to obtain φ(x) in X ∈R L×C , φ(x) transposes X to generate a graph projection matrix φ(x) with the same dimension as B ∈R N×L , and N nodes are used to represent the constructed graph information:

[0079] V = g(X) = BxX;

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

[0081] In order to realize the information interaction of the spatial graph node, an adjacency matrix A s ∈NxN is constructed, and 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 interaction update operation is:

[0083]

[0084] , where W is the graph convolution weight, A s ∈NxN is the constructed adjacency matrix, and V ∈R N×C represents the new reconstructed graph node feature; at this time, the graph node dimension is NxC, and the graph convolution is:

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

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

[0087] The feature information is restored from the graph node to the feature domain, X is transposed using θ(x) to generate a new projection matrix , and the graph space is projected 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 r H×W×C At this time, Y is the residual image of the input image X; adding Y and the original image X obtains the final feature map I s .

[0090] As preferred, the channel information interaction module uses the channel information interaction to capture the relationship between the similar texture information of the relevant channel node features and perform information interaction; the channel information interaction module is given an image X c ∈R H ×W×C First, 1x1 convolution φ(x) is performed on the image, and φ(x) uses convolution operation to change the feature dimension of the image X c ∈R H ×W×N N represents the number of nodes; the average pooling operation is introduced to reduce the size of X c feature;

[0091] X’ c =Avgpool(X c );

[0092] Wherein, X’ c ∈R Q×Q×N QxQ represents the size of the feature obtained again;

[0093] Then, X′ c is repositioned and transformed to obtain X c ∈R N×S At this time, the obtained X c represents the node features projected in the channel dimension, and the feature information of different nodes is interacted through the GCN to obtain the context information; then an adjacency matrix A c ∈R N×S corresponding to X c is created to store the updated node feature state, which is used for information updating between different nodes; the node updating principle is as follows:

[0094]

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

[0096] Y∈R H×W×NReshape the feature dimension of Y from N to C, and sum it with the original image X to obtain the final feature map I c .

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

[0098] respectively, I s and I c use spatial self-attention operation and channel self-attention operation to fuse the obtained feature map; then the obtained feature information is converted to the image domain to obtain the final enhancement result:

[0099]

[0100] wherein F(·) represents a comprehensive operation, F(·) includes the fusion operation of the feature maps I s and I c , and spatial and channel attention operations; D(·) represents a decoder operation, which converts the obtained feature information to the image domain, is the obtained enhanced image.

[0101] Embodiment:

[0102] In the present application, the training data is cropped into 224x224 pixel blocks to obtain pairs of 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, wherein β1 and β2 are taken as 0.9 and 0.999 respectively. When training to 400 epochs, the learning rate is reduced to 0.5 times of the original. All experiments are carried out 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 above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. In the above-mentioned embodiments of the present application, the description of each embodiment is focused on, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments. In the several embodiments provided in the present application, it should be understood that the disclosed technical content can be realized by other means.

[0106] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A two-layer graph inference method for low-light image enhancement, characterized in that, Includes the following steps: S01: Acquire the original RGB low-light image I low Given a low-light image I low Simultaneously inputting both Gamma-net and WF-Net networks, the Gamma-net network generates a series of parameters [λ1γ]. 1 ,λ1γ 1 ,…,λ n γ n The images are then input into the WF-Net network, where they are weighted and fused to generate the clear image I for this stage. clear ; S02: Clear image I at this stage clear Input into the spatial information interaction module to obtain I clear The relationship between spatial node features and similar regions with similar image content is determined, and information exchange is performed. A graph convolutional neural network is then used to update the graph nodes, yielding the final result image I for this stage. s ; S03: Clear the image at this stage I clear The image I is input into the channel information interaction module to obtain the clear image at this stage. clear This involves identifying relationships between similar node features across different channel dimensions, capturing the connections between similar texture information of related channel node features, and enabling information exchange. A graph convolutional neural network is then used to update the graph nodes, resulting in the final image I for this stage. c ; S04: Transfer the results from the spatial information interaction module I s The final result image I of the channel information interaction module c The input is fed into the feature fusion module, where squeezing and excitation operations are used to obtain inter-channel feature information. The features are then concatenated and fed into the decoder to generate the final enhanced result. The denoising module includes: a generation stage and a weighted fusion stage; During the generation stage, low-light image I low As input, a series of parameters [λ1γ] are generated through the Gamma-net network. 1 ,λ1γ 1 ,…,λ n γ n ], and add the parameters to the low-light image I. low Generate images with different exposure levels [I1,I2,…,I] n ]; In the weighted fusion stage, the WF-Net network generates a corresponding weight matrix, combining the feature information of images with different exposures with the original image to achieve denoising. The denoising module operates as follows: I clear =DNM(I)=[I low ,I1,…,I i ,…,I n ]; Where DNM(·) represents the denoising module operation, I i This represents the generated exposure intensity image I, which has a low-light image. low and exposure parameters [λ1γ] 1 ,λ1γ 1 ,…,λ n γ n Combining these, we get 1 ≤ i ≤ n; I clear The final image generated is represented; the output of the gamma-net network is a set of corresponding pixel-level parameters; the weighted fusion stage uses a weighted fusion method to fuse feature information between different regions; Feature extraction is performed using two 3x3 convolutions, followed by normalization using the sigmoid function. The input is the fused image obtained from the i-th and (i-1)-th exposures, and the output is two distinct weight matrices. The weight matrix represents the health state of each pixel, expressed as: oh i-1 ,oh i =WF(I i-1 ,I i ); Where WF(·) represents the WF-Net network that generates the weight matrix; ω i Indicates with I i The corresponding weighted graph, ω i-1 Indicates with I i-1 The corresponding weighted graph; c∈{R,G,B} represents the three channels of red (R), green (G), and blue (B) in a low-light image; The purpose of the channel information interaction module is to capture the relationship between similar texture information of related channel node features and to perform information interaction; provide the channel information interaction module with an image X. c ∈R H×W×C Where H×W represents the size of the feature map and C represents the feature dimension; first, a 1x1 convolution φ(x) is performed on the image, and φ(x) changes the feature dimension X of the image using the convolution operation. c ∈R H×W×N N represents the number of nodes; average pooling is introduced to reduce X. c The size of the feature; X’ c =Avgpool(X c ); Where X' c ∈R Q×Q×N Q×Q represents the size of the reacquired feature; Then for X c After repositioning and transformation, X is obtained. c ∈R N×S Where S represents the graph node information obtained at this time, X c This represents the node features projected along the channel dimension. Feature information from different nodes interacts through the GCN to obtain context information; then, a connection is created with X. c ∈R N×S The corresponding adjacency matrix A c ∈N×N is used to store the updated node feature states for information updates 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; a new projection matrix B∈R is generated by transposing X with θ(x). L×N Then, the graph space is projected back into the original feature space to obtain Y∈R. N×L ; Use a single convolution operation on Y∈R H×W×N The feature dimension of Y is reshaped from N to C, and then summed with the original image X to obtain the final feature map I. c .

2. The two-layer graph inference method for low-light image enhancement according to claim 1, characterized in that, The gamma-net network includes two convolutional layers and ReLU activation functions, average pooling functions, and sigmoid activation functions; to highlight the feature information between different regions, an iterative method is used to generate images with different exposures. Among them, I n-1 In-1 represents the generated exposure image; 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 RGB channels of the image, compared with directly generating a uniform γ value to recover the overall color difference.

3. The two-layer graph inference 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 one graph convolution structure. Two convolutions are used to restore and expand the input features X and output Y, and one convolution is used to generate a dual projection B between the coordinates and the potential interaction space. Graph convolutional networks are used for global information inference based on the interaction space and the adjacency matrix of channel nodes.

4. The two-layer graph inference method for low-light image enhancement according to claim 3, characterized in that, The purpose of the spatial information interaction module is to obtain the relationship between similar regions of related spatial nodes 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 is first subjected to two 1x1 convolutions φ(x) and θ(x), and then X is reshaped by φ(x) to obtain the image in X∈R. L×C Transpose φ(x) with respect to X to generate a form with B∈R N×L A graph projection matrix φ(x) of the same dimension is used, and N nodes are used to represent the information of the constructed graph: V = g(X) = B × X; Where V∈R N×C Let B represent the new reconstructed graph node features, B be the graph node projection matrix, and g(X) be the linear projection operation. To enable information exchange between nodes in a spatial graph, an adjacency matrix A needs to be constructed. 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 can be inferred.

5. A two-layer graph inference method for low-light image enhancement according to claim 4, characterized in that, The graph node interaction update operation is as follows: Where W is the graph convolution weight, A s ∈N×N is the constructed adjacency matrix, V∈R N×C This represents the new reconstructed graph node features; at this point, the graph node dimension is N×C, and the graph convolution is: SCI(S)=Conv1D(Conv1D(S) T ) T ; Where Conv1D(·) represents a 1x1 convolution, and SCI(·) represents a spatial information interaction module; The feature information is restored from the graph nodes to the feature domain, and X is transposed using θ(x) to generate a new projection matrix. Then, the graph space is projected 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 in R L×C In the middle, Y is reshaped into R. H×W×C At this point, Y is the residual map of the input image X; adding Y to the original image X yields the final feature map I. s .

6. A two-layer graph inference method for low-light image enhancement according to claim 1, characterized in that, In step S03, the channel attention mechanism uses squeezing and excitation operations to acquire inter-channel feature information; the fused feature map is placed into the image domain decoder D(.) and recovered back to the image domain, thus obtaining the final enhancement result. For I respectively s and I c Spatial self-attention and channel self-attention operations are used to fuse the acquired feature maps; then the obtained feature information is transformed into the image domain to obtain the final enhancement result. Where F(·) represents a synthesis operation, and F(·) includes feature map I s and I c The fusion operation, as well as spatial and channel attention operations; D(·) represents the decoder operation, which transforms the obtained feature information into the image domain. This is the enhanced image obtained.

Images