Low-light image enhancement method and system based on global-local illumination perception

Abstract

The present invention discloses a low-light image enhancement method and system based on global-local illumination perception, and relates to the technical field of low-light image enhancement; the method comprises the following steps: S1, obtaining basic features through the first convolutional layer of the original low-light image; S2, inputting the basic features into a codec structure composed of multiple blocks for enhancement to obtain deep features, each block includes a multi-level feature extraction module and a multi-level attention fusion module; the multi-level feature extraction module is used to extract global information and local information, and the multi-level attention fusion module is used to filter and transfer the two parts of information backward; S3, fusing the deep features with the original low-light image to obtain an enhanced image; S4, inputting the enhanced image into an image denoising network based on dense residual connections for denoising. The low-light image enhancement method of the present invention can effectively combine the global and local information of the image, solve the problem of uneven illumination distribution, and reduce the noise of the enhanced image.

Description

Technical Field

[0001] The present invention belongs to the technical field of low-light image enhancement, and in particular to a low-light image enhancement method and system based on global-local illumination perception. Background Art

[0002] Affected by force majeure factors such as ambient light, equipment conditions, and shooting techniques, images taken in low-light environments often suffer from serious detail loss and strong noise, which reduces the recognition of image content. In order to mitigate these effects and generate high-quality images with appropriate brightness, low-light image enhancement methods have emerged one after another.

[0003] Traditional low-light enhancement methods rely on prior knowledge of images or physical models from other tasks, such as dark channel prior knowledge, histogram equalization-based methods, etc. In recent years, more and more deep learning-based methods have been applied to low-light image enhancement. Deep learning methods have better performance and speed due to their powerful learning ability and large-scale synthetic data sets. The success of the Transformer model in the field of natural language processing has made researchers in the field of computer vision begin to think about whether it is possible to introduce the standard Transformer structure, use the long-distance modeling ability of the self-attention mechanism, expand the receptive field, and capture the relationship between long-distance pixel information. With continuous exploration, Transformer-based algorithms have performed well in various visual tasks and have been widely used in the field of computer vision. However, the computational complexity of the Transformer network increases with the increase of the spatial resolution of the image, consuming a lot of computing resources, and cannot be applied to the processing of high-resolution image tasks. At the same time, although the Transformer structure can build long-distance dependencies, it lacks the ability to focus on local information. Therefore, building an efficient Vision Transformer that considers local feature relationships is a key research work in the future.

[0004] The Restoration Transformer (Restormer) provides a general model for applying Transformer in the field of image restoration. The multi-head self-attention module MDTA proposed by it changes the image block restriction in ViT, thereby reducing the computational complexity of the model and solving the limitation that the model cannot be used in high-resolution input images. The convolution kernel of CNN has a powerful local feature extraction capability and retains the original relative position relationship of image pixels. It is still a research hotspot in the field of computer vision. However, considering the computational complexity, the convolution kernel cannot be too large, which limits the ability of the convolutional neural network to capture global information and affects the model performance.

[0005] At the same time, the traditional low-light enhancement method and deep learning method in the existing technology still have the following problems:

[0006] 1) Whether it is traditional methods or deep learning methods, it is difficult to properly improve the brightness of the image, which may cause overexposure or underexposure. For example, histogram equalization adjusts the overall brightness of the image by averaging the brightness of the image, making some dark areas overbrightened or some bright areas overdark, affecting the visual effect; although deep learning methods can use adjacent pixels as a reference, due to the lack of overall brightness perception and normal light distribution characteristics, most images restored by the network still have uneven light distribution.

[0007] 2) Due to insufficient ambient light, some details of the captured images are completely hidden in the dark. When enhancing, traditional methods and unsupervised deep learning algorithms do not have external reference images corresponding to low-light images, and it is difficult to perform fine pixel restoration based on the small amount of information in the image, resulting in the loss of some details in the restored image. Supervised deep learning methods can supplement the detailed information of the image through constraints because they have paired training data. However, due to the limitations of model performance and image brightness distribution, deep learning methods will blur image details during the convolution process.

[0008] 3) Dark areas in low-light images often contain large amounts of noise, which is difficult to detect due to low visibility. When enhancing low-light images, the contrast and brightness of the image are improved, but at the same time, the noise hidden in the darkness will also increase. Some methods propose to eliminate noise first and then increase the image brightness, which will undoubtedly result in the loss of image information. The image noise distribution and brightness distribution do not match, so it is difficult to obtain a general model that can have the best of both worlds. Existing deep learning models introduce random noise in order to train models with strong denoising capabilities. However, this usually results in a more complex noise distribution in the enhanced image, making it difficult to achieve denoising effects.

[0009] Although the Transformer structure can build long-distance dependencies, it lacks the ability to focus on local information. This paper proposes a new network framework LightenTCN that combines Transformer and CNN, which can effectively combine the global and local information of the image, solve the problem of uneven illumination distribution, and reduce the noise of the enhanced image. Summary of the invention

[0010] The purpose of the present invention is to provide a low-light image enhancement method and system based on global-local illumination perception, so as to solve the problems that the restored image in the prior art proposed in the above background technology still has uneven illumination distribution, is prone to losing some details and is difficult to achieve denoising.

[0011] To achieve the above object, the present invention adopts the following technical solutions:

[0012] A first aspect of the present invention provides a low-light image enhancement method based on global-local illumination perception, comprising the following steps:

[0013] S1, the original low-light image is subjected to preliminary feature extraction through the first convolutional layer to obtain basic features;

[0014] S2. Input the basic features into a codec structure composed of multiple blocks for enhancement to obtain deep features; each block includes a multi-level feature extraction module and a multi-level attention fusion module, the multi-level feature extraction module is used to extract global information and local information, and the multi-level attention fusion module is used to filter and transfer the global information and local information backward;

[0015] The extraction of global information in the multi-level feature extraction module is performed by respectively passing the basic features through a multi-head self-attention feature extraction module based on the height axis and a multi-head self-attention feature extraction module based on the width axis, respectively extracting the height axis features and the width axis features to obtain the overall illumination features of the image;

[0016] Extraction of local information in the multi-level feature extraction module, extracting local features through the basic features through the image local feature extraction module built based on the convolutional neural network, obtaining the local distribution characteristics of the image and perceiving the local brightness of the image;

[0017] S3, input the deep features into the second convolutional layer to obtain the brightness difference, and use the residual connection to fuse the deep features with the original low-light image to obtain an enhanced image;

[0018] S4. Input the enhanced image into an image denoising network based on dense residual connections to denoise the enhanced image.

[0019] Preferably, the multi-head self-attention feature extraction module based on the height axis and the multi-head self-attention feature extraction module based on the width axis, the feature extraction process is as follows:

[0020] The basic features of the input are first subjected to three sets of linear mapping to generate Query, Key, and Value matrices. Then, by reshaping the Query and Key, an attention map of the height axis and an attention map of the width axis are generated respectively. The attention map is used to guide the reconstruction of the global illumination features.

[0021] Preferably, the image local feature extraction module constructed based on the convolutional neural network has the following specific structure: it consists of two 3×3 convolutional layers and a PReLU activation layer, which is used to extract local features.

[0022] Preferably, in the multi-level attention fusion module, the feature fusion process is as follows:

[0023] Firstly, three adjustable parameters α, β, and γ are introduced to weight and distinguish the height axis features, width axis features and local features, ensuring that the global and local information used for image enhancement are retained and other information is compressed; then the filtered features are merged through the concat operation to form a fusion feature; then the global average pooling operation is used to purify the key features; finally, the residual learning strategy is used to obtain the final deep features.

[0024] Preferably, the global average pooling operation is performed by a global average pooling layer, two 1×1 convolutional layers, a ReLU activation layer and a Sigmoid layer.

[0025] Preferably, the image denoising network based on dense residual connections enhances the denoising process of the image as follows:

[0026] A 5×5 convolutional layer is used to extract shallow features of the input enhanced image, and the extracted shallow features are fed into two consecutive dense residual blocks for deep processing;

[0027] Each dense residual block consists of two 3×3 convolutional layers and one 1×1 convolutional layer. In the dense residual block, the input features not only pass through the first 3×3 convolutional layer, but are also densely fused with the output features of the first 3×3 convolutional layer, the second 3×3 convolutional layer, and the 1×1 convolutional layer, and the output features of the first 3×3 convolutional layer are also fused with the output features of the second 3×3 convolutional layer.

[0028] The output features after processing by two dense residual blocks are then subjected to a 5×5 convolution to obtain the corresponding image, which is then fused with the initial input enhanced image to obtain the final denoised image.

[0029] A second aspect of the present invention provides a low-light image enhancement system based on global-local illumination perception applied to the above method, wherein the low-light image enhancement system adopts a network structure of LightenTCN, which includes a first convolutional layer, an encoder, a decoder, a second convolutional layer and an enhanced image denoising module;

[0030] The first convolution layer is a 3×3 convolution layer, which is used to perform preliminary feature extraction on the original low-light image to obtain basic features;

[0031] The encoder and decoder are composed of 7 blocks, each of which contains a multi-level feature extraction module and a multi-level attention fusion module. The jump connection technology is introduced between the 7 blocks for feature fusion and transmission. It is used to extract global information and local information from images with basic features, and to filter and transfer the global information and local information accordingly to obtain deep features.

[0032] The multi-level feature extraction module includes a multi-head self-attention feature extraction module based on the height axis, a multi-head self-attention feature extraction module based on the width axis, and an image local feature extraction module based on a convolutional neural network;

[0033] The multi-level attention fusion module includes a feature weighting module, a concat operation module, a global average pooling operation module and a residual learning strategy module;

[0034] The second convolution layer is a 3×3 convolution layer, which is used to process the deep features to obtain brightness difference; after the deep features obtain the brightness difference, they are fused with the original low-light image using residual connection to obtain an enhanced image;

[0035] The enhanced image denoising module is an image denoising network based on dense residual connections, and is used to denoise the enhanced image.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention proposes a new network framework LightenTCN that combines Transformer and CNN, which captures the global and local feature relationships at the same time, so that the illumination of the enhanced image is evenly distributed. The encoder and decoder in the network structure are both composed of a Block, which is internally composed of a multi-level feature extraction module MFEM and a multi-level attention fusion module MAFM. The MFEM module extracts global and local information, and the MAFM module filters and transmits the two parts of information backward. In addition, for the noise problem of low-light enhanced images, the present invention also introduces an enhanced image denoising module EIDM to improve the quality of the enhanced images. The results show that LightenTCN can effectively combine the global and local information of the image, solve the problem of uneven illumination distribution, and reduce the noise of the enhanced image. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a system block diagram of a low-light image enhancement system based on global-local illumination perception in the present invention;

[0039] Figure 2 It is a structural schematic diagram of the multi-level feature extraction module in the present invention;

[0040] Figure 3 Schematic diagram of the structure of the multi-level attention fusion module in the present invention;

[0041] Figure 4 It is a structural schematic diagram of the enhanced image noise reduction module in the present invention;

[0042] Figure 5It is a visual comparison diagram of the low-light image enhancement method in the prior art and the low-light image enhancement method in the present invention on the LOL dataset;

[0043] Figure 6 This is a visual comparison chart of the low-light image enhancement method in the prior art and the low-light image enhancement method in the present invention on the MIT-AdobeFiveK dataset. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 are within the scope of protection of the present invention.

[0045] Embodiment 1:

[0046] The low-light image enhancement system based on global-local illumination perception is specifically designed as follows:

[0047] LightenTCN is a deep learning network specifically designed for image quality enhancement under low light conditions. Its design is based on the classic U-Net architecture and consists of a carefully constructed encoder and decoder. Both the encoder and decoder are composed of blocks. Figure 1 shown.

[0048] The low-light image enhancement method based on global-local illumination perception is as follows:

[0049] When processing the input low-light image, a 3×3 convolutional layer is first used to perform preliminary feature extraction, laying the foundation for subsequent deep information processing. Subsequently, these basic features are fed into the codec structure composed of blocks to further mine and enhance the image information. Each block contains a multi-level feature extraction module (MFEM) and a multi-level attention fusion module (MAFM), which aims to improve the network's processing ability for image features and the accuracy of capturing details. In order to effectively reduce the loss of detail information in the encoder and better restore these features in the decoder, a jump connection technology is introduced between the encoder and the decoder to achieve feature fusion and transmission. After multiple blocks are processed, the deep features are input into a 3×3 convolutional layer to obtain the brightness difference, and the residual connection is used to fuse it with the input image to reconstruct the enhanced image. Finally, the enhanced image is fused with the original input image and processed by an efficient image denoising module (EIDM) to eliminate possible noise and obtain a clear final output image.

[0050] In this embodiment, specifically, regarding the multi-level feature extraction module (MFEM);

[0051] The complex relationship within the image mainly consists of two parts: one is the global relationship between long-distance pixels, and the other is the local relationship between neighborhood pixels. In view of the powerful ability of the Transformer architecture in capturing long-distance positional relationships, it has been widely used in the field of computer vision. In the feature extraction process, the present invention adopts a global-local extraction strategy. Specifically, the extraction of global information is achieved through two multi-head self-attention mechanisms: one is the multi-head self-attention feature extraction (HAE) based on the height dimension, which is responsible for extracting the overall illumination characteristics of the image; the other is the multi-head self-attention feature extraction (WAE) based on the width dimension, which is also committed to capturing global illumination information. For the extraction of local information, the present invention utilizes the powerful local feature extraction capability of the convolutional neural network (CNN) to construct an image local feature (LFE), which not only effectively perceives the local brightness of the image, but also successfully obtains the local distribution characteristics. The design of the entire network structure aims to achieve accurate capture and processing of internal relationships of the image. Its detailed structure is shown in the figure. Figure 2 shown.

[0052] The axis-based multi-head self-attention proposed by LLFormer in the prior art decomposes the traditional feature attention based on the length, width, and channel dimensions of the image into axis-based feature attention. Based on this, the present invention adopts a global-local extraction strategy, uses an axis-based self-attention mechanism to extract the global illumination features of the image, and uses CNN to extract the local illumination features of the image. The LLFormer model adopts a sequential self-attention method, first performing multi-head self-attention processing with the height as the axis, and then continuing to process with the width as the axis, where the output of the former serves as the input of the latter. However, this sequential processing method may be insufficient in retaining the original texture information of the image. In order to better retain the fine structure of the image, the present invention implements multi-head self-attention feature extraction (HAE) based on the height axis and multi-head self-attention feature extraction (WAE) based on the width axis on the original image, and the specific structures are shown respectively. Figure 2 Figures (a) and (b).

[0053] The specific structure of the multi-level feature extraction module (MFEM) is as follows:

[0054] In the HAE and WAE modules, the input features first undergo three sets of linear mappings to generate Query, Key, and Value matrices, and then the attention maps of the height axis and the width axis are generated by reshaping the Query and Key. These attention maps are then used to guide the reconstruction process of the global illumination features. In addition, the convolution kernel of the convolutional neural network (CNN) is good at extracting fine contours in fixed-size image blocks, so it performs well in preserving local spatial details of the image. Based on the local feature extraction capability of CNN, the present invention constructs an image local feature extraction module (LFE), whose structure is as follows: Figure 2 As shown in the figure (c), it consists of two 3×3 convolutional layers and a PReLU activation layer.

[0055] In this embodiment, specifically, regarding the multi-level attention fusion module (MAFM);

[0056] The output of the multi-level feature extraction module (MFEM) consists of features of three different dimensions: height axis features, width axis features, and local features. These three features do not interfere with each other during the generation process. In order to achieve effective fusion between them and optimize the expression of information, the present invention designs a multi-level attention fusion module (MAFM). The detailed structural design of this module is as follows Figure 3 As shown. The purpose of this module is to ensure that the three features are efficiently fused without being coupled to each other. Through this module, the present invention accurately screens the global information, local information and invalid information of the image to ensure that the most critical information for image enhancement is retained. After screening, the three parts of features, namely the height axis features, the width axis features and the local features, are merged through a splicing operation to form a comprehensive fusion feature set. In order to further improve the quality of the fused features, the present invention uses global average pooling technology to process the fused features, and this step helps to purify key global texture details. Through such processing, the enhanced image is closer to the reference image in texture and details, thereby achieving a significant improvement in image quality.

[0057] The specific structure of the multi-level attention fusion module (MAFM) is as follows:

[0058] In MAFM, three adjustable parameters α, β, and γ are first introduced to weight and discriminate the three features of height axis features, width axis features, and local features. During the network training process, these hyperparameters are dynamically adjusted according to the optimization algorithm to ensure that the global and local information that is most critical to image enhancement is retained, while compressing those unimportant information. The screened features are then merged through the concat operation to form a fusion feature. In order to further improve the quality of global texture details in the fusion features, the present invention uses a global average pooling operation to purify key features. This process is completed through a global average pooling layer, two 1×1 convolutional layers, a ReLU activation layer, and a Sigmoid layer, so that the enhanced image is closer to the reference image in texture and details. Finally, through the residual learning strategy, the present invention can obtain the final enhanced image, which helps to reduce the loss of information during the feature fusion process and promote further improvement of network performance.

[0059] In this embodiment, specifically, regarding the enhanced image denoising module (EIDM);

[0060] Images captured under low-light conditions are often accompanied by noise problems, and during the image enhancement process, the noise is often further amplified, which brings challenges to image processing. To address this problem, there are currently two main solutions: the first is to eliminate noise before image enhancement and then increase the image brightness; the second is to first increase the image brightness and then effectively eliminate the noise generated after enhancement. The research of the present invention tends to adopt the second solution. The reason is that if noise is eliminated first, important image detail information may be lost in the subsequent enhancement process, resulting in the inability to restore the characteristics of the original image. On the contrary, enhancement is performed first and then denoising. Although the noise information is amplified, most of the noise can be removed by various means. The present invention uses dense residual connections to construct an image denoising network to perform denoising on the enhanced image. This makes the final output image have better visual effects and index performance. The structure of the denoising module is as follows Figure 4 shown.

[0061] The specific structure of the enhanced image denoising module (EIDM) is as follows:

[0062] The present invention adopts an image denoising network based on dense residual connection, aiming to achieve effective denoising of images processed by image enhancement. The input of the network is the result of fusion of the enhanced image processed by 7 blocks and the original input image. The first step of the network is to use a 5×5 convolution layer to extract the shallow features of the input image, which lays the foundation for the subsequent denoising process. The extracted shallow features are then sent to two consecutive dense residual blocks (DRMs) for deep processing. Each DRM consists of two 3×3 convolution layers and one 1×1 convolution layer. In the DRM, the input features not only pass through the first 3×3 convolution layer, but also densely fuse with the output features of the subsequent two 3×3 convolution layers and the 1×1 convolution layer. In addition, the output of the first 3×3 convolution layer will also be fused with the output of the second 3×3 convolution layer. This design makes it possible to fully utilize and fuse features at different levels. The features processed by the two DRMs are then subjected to a 5×5 convolution to obtain the corresponding image, which is then fused with the enhanced image input at the beginning to obtain the final denoised image. The enhanced image denoising module not only makes the output image have better visual effects, but also further improves the evaluation index results.

[0063] Experimental verification:

[0064] This embodiment selects the LOL dataset and MIT-Adobe FiveK dataset, which are common in low-light image enhancement, to conduct quantitative and qualitative experiments, and selects the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and learning-perceptual image patch similarity (LPIPS) evaluation indicators to evaluate the image quality.

[0065] First, the qualitative and quantitative evaluation is performed on the LOL dataset. The proposed low-light image enhancement method combining Transformer and CNN is compared with a variety of advanced low-light image enhancement methods, including DLN, KinD, RetinexNet, KinD++, ZeroDCE, EnlightenGAN, and Transformer-based methods Uformer, Restormer, LLFormer, etc. For fair comparison, the above methods are retrained with the same experimental parameter settings. The quantitative evaluation results are shown in Table 1.

[0066] Table 1 Quantitative results of LightenTCN and other methods on the LOL dataset

[0067] Method PSNR↑ SSIM↑ LPIPS↓ RetinexNet 16.77 0.56 0.742 ZeroDCE 13.55 0.54 0.222 KinD 20.87 0.80 0.160 EnlightenGAN 17.48 0.65 0.153 KinD++ 21.30 0.82 0.159 DLN 21.94 0.81 0.095 Uformer 18.55 0.72 0.32 Restormer 22.37 0.82 0.14 LLFormer 23.43 0.81 0.094 LightenTCN(Ours) 24.29 0.83 0.087

[0068] As shown in Table 1, LightenTCN performs well in reducing the noise of enhanced images and has the highest PSNR value. Figure 5As shown in the figure, ZeroDCE needs to be improved in improving image brightness. The image restored by EnlightenGAN has slight color distortion. KinD and KinD++ add brightness smoothing constraints during re-decomposition, resulting in excessive smoothing of the enhanced image. The Transformer-based method can improve the global illumination of the image, but the image restored by Uformer has vertical artifacts. The image enhanced by LightenTCN in the present invention effectively reduces noise and restores global illumination.

[0069] Secondly, the qualitative and quantitative evaluation is conducted on the MIT-Adobe FiveK dataset. The effectiveness of the proposed LightenTCN is verified by comparing it with a variety of advanced methods such as DeepLPF, DeepUPE, DSLR, EnlightenGAN, Uformer, Restormer, MIRNet and LLFormer. The quantitative evaluation results are shown in Table 2.

[0070] Table 2 Quantitative results of LightenTCN and other methods on the MIT-Adobe FiveK dataset

[0071]

[0072]

[0073] As shown in Table 2, the quantitative indicators PSNR, SSIM and LPIPS of LightenTCN are better than most of the comparison methods. The method proposed in the present invention can effectively compress noise while improving brightness, and the restored image structure is close to the reference image.

[0074] In the LOL dataset above, we verified the effectiveness of LightenTCN in enhancing global brightness. Here, we selected reference images with low local brightness in the MIT-Adobe FiveK dataset to verify the local brightness enhancement performance of LightenTCN. The results are shown in Figure 6 In the case of RetinexNet, serious color distortion occurs when enhancing images, and ElightenGAN also distorts the color of the sky area when restoring local dark areas. DSLR and MIRNet are prone to artifacts when restoring continuous local illumination areas, such as uneven illumination on the wall. Relatively speaking, the Transformer-based method can better restore local and global illumination and preserve the texture details of the image.

[0075] The above description is only used to help understand the method of the present invention and its core essence, but the protection scope of the present invention is not limited thereto. For those skilled in the art in the art, equivalent replacement or change according to the technical solution and inventive concept of the present invention within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A low-light image enhancement method based on global-local illumination perception, characterized in that: The steps include: S1, the original low-light image is subjected to preliminary feature extraction through the first convolutional layer to obtain basic features; S2. Input the basic features into a codec structure composed of multiple blocks for enhancement to obtain deep features; each block includes a multi-level feature extraction module and a multi-level attention fusion module, the multi-level feature extraction module is used to extract global information and local information, and the multi-level attention fusion module is used to filter and transfer the global information and local information backward; The extraction of global information in the multi-level feature extraction module is performed by respectively passing the basic features through a multi-head self-attention feature extraction module based on the height axis and a multi-head self-attention feature extraction module based on the width axis, respectively extracting the height axis features and the width axis features to obtain the overall illumination features of the image; The multi-head self-attention feature extraction module based on the height axis and the multi-head self-attention feature extraction module based on the width axis, the feature extraction process is as follows: The basic features of the input are firstly subjected to three sets of linear mapping to generate Query, Key, and Value matrices. Then, by reshaping Query and Key, the attention map of the height axis and the attention map of the width axis are generated respectively. The attention map is used to guide the reconstruction of the global illumination features. Extraction of local information in the multi-level feature extraction module, extracting local features through the basic features through the image local feature extraction module built based on the convolutional neural network, obtaining the local distribution characteristics of the image and perceiving the local brightness of the image; The multi-level attention fusion module, the feature fusion process is as follows: First, three adjustable parameters α, β, and γ are introduced to weight and distinguish the height axis features, width axis features, and local features, ensuring that the global information and local information used for image enhancement are retained and other information is compressed; then the filtered features are merged through the concat operation to form a fusion feature; then the global average pooling operation is used to purify the key features; finally, the final deep features are obtained through the residual learning strategy; S3, input the deep features into the second convolutional layer to obtain the brightness difference, and use the residual connection to fuse the deep features with the original low-light image to obtain an enhanced image; S4, inputting the enhanced image into an image denoising network based on dense residual connections to denoise the enhanced image; The image denoising network based on dense residual connections enhances the denoising process of the image as follows: A 5×5 convolutional layer is used to extract shallow features of the input enhanced image, and the extracted shallow features are fed into two consecutive dense residual blocks for deep processing; Each dense residual block consists of two 3×3 convolutional layers and one 1×1 convolutional layer. In the dense residual block, the input features not only pass through the first 3×3 convolutional layer, but are also densely fused with the output features of the first 3×3 convolutional layer, the second 3×3 convolutional layer, and the 1×1 convolutional layer, and the output features of the first 3×3 convolutional layer are also fused with the output features of the second 3×3 convolutional layer. The output features after processing by two dense residual blocks are then subjected to a 5×5 convolution to obtain the corresponding image, which is then fused with the initial input enhanced image to obtain the final denoised image.

2. The low-light image enhancement method based on global-local illumination perception according to claim 1, characterized in that: The image local feature extraction module constructed based on the convolutional neural network has the following specific structure: it consists of two 3×3 convolutional layers and a PReLU activation layer, which is used to extract local features.

3. The low-light image enhancement method based on global-local illumination perception according to claim 1, characterized in that: The global average pooling operation is completed through a global average pooling layer, two 1×1 convolutional layers, a ReLU activation layer and a Sigmoid layer.

4. A low-light image enhancement system based on global-local illumination perception applied to the method of any one of claims 1 to 3, characterized in that: The low-light image enhancement system adopts the network structure of LightenTCN, which includes a first convolutional layer, an encoder, a decoder, a second convolutional layer and an enhanced image denoising module; The first convolution layer is a 3×3 convolution layer, which is used to perform preliminary feature extraction on the original low-light image to obtain basic features; The encoder and decoder are composed of 7 blocks, each of which contains a multi-level feature extraction module and a multi-level attention fusion module. The jump connection technology is introduced between the 7 blocks for feature fusion and transmission. It is used to extract global information and local information from images with basic features, and to filter and transfer the global information and local information accordingly to obtain deep features. The multi-level feature extraction module includes a multi-head self-attention feature extraction module based on the height axis, a multi-head self-attention feature extraction module based on the width axis, and an image local feature extraction module based on a convolutional neural network; The multi-level attention fusion module includes a feature weighting module, a concat operation module, a global average pooling operation module and a residual learning strategy module; The second convolution layer is a 3×3 convolution layer, which is used to process deep features to obtain brightness differences; After the deep features obtain the brightness difference, they are fused with the original low-light image using residual connection to obtain an enhanced image; The enhanced image denoising module is an image denoising network based on dense residual connections, and is used to denoise the enhanced image.

Images