System and method for enhancing image quality of low-illumination image

Abstract

The invention relates to the technical field of image processing, in particular to a low-illumination image quality enhancement system and method, and the system comprises a frequency domain processing network and a space domain processing network. In the frequency domain processing network, guiding a multi-scale gating convolution unit to perform image enhancement through a space-frequency mapping weight with frequency domain information, and capturing local and global information through parallel multi-scale convolution; in the space domain processing network, space information and frequency domain information are combined to generate space-frequency joint attention, and texture information of detail areas such as image edges and the like is recovered in a space-frequency feature fusion process.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and particularly relates to a low-light image quality enhancement system and method. Background Technology

[0002] In low-light environments such as nighttime, indoors, or backlit conditions, images captured by imaging devices often suffer from severe quality problems, such as low overall brightness, poor contrast, color distortion, and significant noise. These degradations not only greatly affect the visual appeal of the images but also pose a serious challenge to downstream computer vision tasks, such as object detection, image segmentation, and recognition, leading to a sharp decline in performance. Therefore, researching high-performance, robust low-light image enhancement algorithms aimed at recovering implicit information and improving the quality of low-light images has profound theoretical value and broad practical application significance.

[0003] Traditional methods, based on physical models and hand-designed features, achieve illumination correction through mathematical transformations or filtering strategies. While these methods provide interpretable illumination correction in specific scenarios through physical modeling and mathematical transformations, focusing on illumination estimation, they often introduce significant noise or local color distortion. This is because these methods assume noise-free and color-distortion-free images, which is inconsistent with real-world underexposed scenes. Furthermore, their performance is limited by the need for hand-designed parameters. In contrast, deep learning-based methods have become more popular in recent years. However, most current deep learning-based low-light enhancement methods rely on skip-connection feature extraction, which introduces shallow features but also significant noise in dark areas, making image quality restoration difficult. To suppress noise, many models learn "smooth" priors, leading to the loss of realistic texture and detail in the enhanced image, resulting in pixel-level loss, blurred edges, and damage to image details. Finding a balance between enhancing brightness, suppressing noise, preserving detail and color fidelity, and achieving a natural visual effect remains challenging. Summary of the Invention

[0004] In view of this, the present invention aims to provide a low-light image quality enhancement system and method. This system recursively iterates the amplitude component in the image's frequency domain using a frequency-domain-guided multi-scale convolutional gated recurrent unit to improve image brightness. Furthermore, it jointly generates a spatial-frequency joint attention mechanism by combining the image's spatial and frequency domain information, further restoring the texture of detailed areas such as image edges and contours during the spatial-frequency feature fusion process. This invention effectively improves the restoration of image details, fully utilizes the image's dual-domain information, and provides more comprehensive processing for noise suppression and color consistency.

[0005] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A low-light image quality enhancement system includes: a frequency domain processing network, comprising a domain transformation unit, a space-frequency mapping weight extraction unit, and a multi-scale gated convolution unit, wherein: the domain transformation unit converts the input low-light image into a first frequency domain image; the space-frequency mapping weight extraction unit performs high-frequency noise compression on the first frequency domain image and extracts global illumination features to obtain space-frequency mapping weights; the multi-scale gated convolution unit, guided by the space-frequency mapping weights, performs gated convolution operations at different scales on the first frequency domain image, and the obtained features are then input into the domain transformation unit to obtain a preliminary enhanced image; and a spatial domain processing network, including a space-frequency joint attention module, converts the preliminary enhanced image into a second frequency domain image, extracts frequency domain weights from the second frequency domain image, obtains spatial domain weights from the preliminary enhanced image, combines the frequency domain weights and spatial domain weights, and applies them to the preliminary enhanced image to obtain the final enhanced image.

[0006] Furthermore, in the space-frequency mapping weight extraction unit, a convolution operation is performed on the first frequency domain image, and the features after the convolution operation are then processed by a fully connected layer to obtain the space-frequency mapping weights.

[0007] Furthermore, in the multi-scale gated convolutional unit, multi-scale gated convolution is performed using the following formula: G k =Conv k ([X,h t-1 ]); Among them, G k Conv represents the output features of the input first frequency domain image X after being gated convolution at the k-th scale. k H represents gated convolution. t-1 [,] represents the previous hidden state in the gated convolution at the current scale, and [,] represents the first frequency domain image X and the previous hidden state h. t-1 Channel concatenation is performed; multi-scale extraction of space-frequency mapping weights is performed, and the multi-scale gated convolution output features are fused using the space-frequency mapping weights guided by the following formula: ; Among them, G freq This indicates the characteristics after guided fusion. The k-th scale spatial-frequency mapping weights are represented; the fused features are combined with the first frequency domain image to obtain the output features of the multi-scale gated convolutional unit.

[0008] Furthermore, in the multi-scale gated convolutional unit, a penalty is applied to the hidden state using the following formula: ; Among them, h tLet represent the current hidden state after penalty, z represent the update gate in the gated convolution, c represent the candidate state, λ represent the penalty coefficient, and MLP represent the multilayer perceptron. This indicates element-wise multiplication.

[0009] Furthermore, the fused features are added to the amplitude component of the first frequency domain image and then combined with the phase component of the first frequency domain image, and input into the domain conversion unit to obtain a preliminary enhanced image.

[0010] Furthermore, in the spatial-frequency joint attention module: the process of extracting frequency domain weights from the second frequency domain image includes: performing channel attention operation on the amplitude component in the second frequency domain image, and the resulting weights are used together with the phase component in the second frequency domain image as frequency domain weights; the process of obtaining spatial domain weights from the preliminary enhancement image includes: performing multiple consecutive convolution operations on the preliminary enhancement image to obtain spatial domain weights; and applying the weighted sum of the frequency domain weights and spatial domain weights to the preliminary enhancement image to obtain the final enhancement image.

[0011] A method for enhancing image quality in low-light images includes: S1: Obtain the low-light image dataset and preprocess it to obtain the training set; S2: Construct a low-light image quality enhancement system as provided in this invention, and train the low-light image quality enhancement system using the training set constructed in step S1 to obtain an image quality enhancement model. S3: Input the low-light image to be enhanced into the image quality enhancement model obtained in step S2, and output the corresponding final enhanced image.

[0012] Furthermore, in step S2, the low-light image quality enhancement system is trained using the total loss function expressed in the following formula: L=ξL s1 +μL s2 +ηL vgg ; Where L represents the total loss function, L s1 L represents the loss function of the frequency domain processing network. s2 L represents the loss function of the spatial processing network. vgg Let ξ represent the perceptual loss function implemented based on the VGG network, and let ξ, μ, and η represent the loss weights.

[0013] Furthermore, the loss function L of the frequency domain processing network s1 for: ; Where F represents the Fourier transform, A represents the amplitude component extraction operation, and Output s1This represents the initial enhanced image, where GT represents the corresponding true specular image. This represents the 2-norm operation; Loss function L of spatial processing network s2 for: ; Among them, Output s2 This represents the final enhanced image; Perceptual loss function L vgg for: ; in, This represents the features obtained from VGG network processing.

[0014] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) In the low-light image quality enhancement system and method described in this invention, a frequency-domain guided multi-scale convolutional gated recurrent unit is proposed to replace skip connections, dynamically adjust the weights of encoder features, suppress the feature response of noise regions, and strengthen effective information such as edges and textures. This makes up for the defect that skip connections, which are widely used in current enhancement networks, may transmit noise information along with the low-level features of the encoder to the decoder, resulting in artifacts in the enhancement results. In the spatial domain stage, a spatial-frequency joint attention module is designed to make full use of dual-domain information. By combining spatial domain attention and frequency domain attention, global illumination adjustment and local texture repair are optimized synchronously. The feature complementarity of dual-domain information is fully utilized to specifically strengthen high-frequency detail textures and edge contour regions, suppress low-frequency redundant information, and further suppress local noise. This achieves the synergistic effect of global and local features, and can more effectively restore high-frequency edge and other detail texture information. (2) In the low-light image quality enhancement system and method described in this invention, the gradient penalty mechanism is used to suppress the abrupt update of the hidden state, realize the smooth constraint of the brightness variable, force the brightness change of adjacent pixels to be gradual, make the enhancement result more in line with the light distribution law of the natural scene, improve the performance of the model, and effectively reduce the occurrence of overexposure and other situations. Attached Figure Description

[0015] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of the low-light image quality enhancement system described in an embodiment of the present invention; Figure 2A schematic diagram of the multi-scale gated convolutional unit described in an embodiment of the present invention; Figure 3 A schematic diagram of the space-frequency joint attention module described in an embodiment of the present invention; Figure 4 This is a schematic flowchart of the low-light image quality enhancement method described in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.

[0017] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this invention can be combined with each other. In the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0018] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] like Figure 1 As shown, the low-light image quality enhancement system described in this embodiment of the invention includes a frequency domain processing network and a spatial domain processing network.

[0020] The frequency domain processing network includes a domain transformation unit, a space-frequency mapping weight extraction unit, and a multi-scale gated convolution unit. The domain transformation unit converts the input low-light image into a first frequency domain image. The space-frequency mapping weight extraction unit performs high-frequency noise compression on the first frequency domain image and extracts global illumination features to obtain space-frequency mapping weights. The multi-scale gated convolution unit, guided by the space-frequency mapping weights, performs gated convolution operations at different scales on the first frequency domain image. The resulting features are then input into the domain transformation unit to obtain a preliminary enhanced image.

[0021] Images with non-uniform lighting often contain severe noise (such as dark noise and color distortion). Currently used skip connections directly pass low-level features from the encoder to the decoder, potentially transmitting noise along with the signal, leading to artifacts in the enhancement result. Furthermore, the degradation levels vary significantly across different regions of non-uniformly lit images (e.g., more noise in dark areas and less noise in bright areas), making it difficult for traditional skip connections to flexibly adjust feature importance through static stitching or addition. Therefore, this invention addresses the limitations of skip connections by proposing the replacement of skip connections with frequency-domain guided multi-scale gated convolutional units. A space-frequency mapping weight extraction unit dynamically adjusts the weights of the output features of the multi-scale gated convolutional units, suppressing feature responses in noisy regions while enhancing effective information such as edges and textures.

[0022] In this embodiment of the invention, the domain transformation unit converts the input low-light image into a first frequency domain image through Fourier transform. The first frequency domain image includes a low-frequency component (characterizing global illumination) and a high-frequency component (characterizing details / noise). The low-frequency component is concentrated in the center of the first frequency domain image, reflecting the overall brightness distribution of the image, while the high-frequency component is located at the edge of the first frequency domain image, corresponding to texture details.

[0023] In some embodiments, in the space-frequency mapping weight extraction unit, a convolution operation is performed on the first frequency domain image to compress high-frequency noise and extract global illumination features; the features after the convolution operation are then processed by a fully connected layer to obtain space-frequency mapping weights, which are used as weights in subsequent multi-scale gated convolution units.

[0024] In some embodiments, the processing procedure of a multi-scale gated convolutional unit is as follows: Figure 2 As shown, this includes: multi-scale gated convolution performed using the following formula: G k =Conv k ([X,h t-1 ]); Among them, G k Conv represents the output features of the input first frequency domain image X after being gated convolution at the k-th scale. k H represents gated convolution. t-1 [,] represents the previous hidden state in the gated convolution at the current scale, and [,] represents the first frequency domain image X and the previous hidden state h. t-1Channel splicing is performed. In this embodiment of the invention, to reduce parameters, three convolutional kernel scales are set in the convolutional layer of gated convolution (ConvGRU): small scale (3×3) captures fine-grained textures, focuses on local details, and captures high-frequency features (such as textures and noise), suitable for enhancing or suppressing subtle lighting changes; medium scale (5×5) has a medium receptive field, balances local and regional features, and handles medium-range brightness contrast (such as shadow transitions); global scale (7×7) learns the overall lighting distribution through a large receptive field, corresponding to k={3,5,7}.

[0025] The space-frequency mapping weights are extracted at multiple scales, and the multi-scale gated convolution output features are fused using the space-frequency mapping weights guided by the following formula: ; Among them, G freq This indicates the characteristics after guided fusion. This represents the space-frequency mapping weight at the k-th scale. In this embodiment of the invention, if low-frequency components are significant in the input low-light image (such as uniform illumination areas), then the space-frequency mapping weight corresponding to the global scale (7×7) is... Increase the size of the receptive field; if the low-light image contains many high-frequency components (such as a high-texture scene), then the space-frequency mapping weights corresponding to the small scale (3×3) should be increased. Enlarging and strengthening the role of the small kernel makes it easier to process detailed textures.

[0026] The guided fusion features are combined with the first frequency domain image to obtain the output features of the multi-scale gated convolutional unit. In this embodiment of the invention, this process specifically involves adding the guided fusion features to the amplitude component of the first frequency domain image, and combining them with the phase component of the first frequency domain image to form the output features of the multi-scale gated convolutional unit.

[0027] In low-light image enhancement tasks, enhancement models may over-boost the brightness of dark areas, leading to unnatural halos or overexposure in adjacent regions. To suppress this phenomenon, this invention incorporates a penalty constraint on the hidden conditions into the multi-scale gated convolutional unit. This gradient penalty mechanism suppresses abrupt updates of the hidden state, resulting in smoother brightness changes. Specifically, in some embodiments, the multi-scale gated convolutional unit applies a penalty to the hidden state using the following formula: ; Among them, h t Let represent the current hidden state after penalty, z represent the update gate in the gated convolution, c represent the candidate state, λ represent the penalty coefficient used to control the penalty strength, and MLP represent a multilayer perceptron. This represents element-wise multiplication. In this embodiment of the invention, the Multilayer Perceptron (MLP) includes a 3×3 convolutional layer, a ReLU activation function, and a 1×1 convolutional layer. The MLP can encode local brightness changes in the hidden state. If the brightness of a certain area changes drastically (such as overexposure or object edges), the output value of the MLP increases, resulting in a relatively high penalty. Furthermore, in this embodiment of the invention, the penalty coefficient λ is preferably set to 0.1. Further explanation: When the output value of the MLP is MLP(h... t-1 When ) is large (generally in regions of abrupt changes in brightness), -λMLP(h) t-1 The item will significantly reduce the current hidden state h. t The update amplitude suppresses unnatural changes in the brightness of the output result; when the output value of the multilayer perceptron (MLP) is MLP(h t-1 When the value is small (generally in a smooth region), the constraint effect is weak, allowing the current hidden state h to be allowed to remain hidden. t Normal update. This is equivalent to applying L2 regularization to the gradient of the hidden state, forcing the brightness changes of adjacent pixels to be gradual, making the enhancement result conform to the lighting distribution law of the natural scene, improving the performance of the model, and effectively reducing the occurrence of overexposure and other issues.

[0028] After the guided fusion features are added to the amplitude component of the first frequency domain image, they are then input together with the phase component of the first frequency domain image into the domain transformation unit. The domain transformation unit converts the output of the multi-scale gated convolution unit into a preliminary enhanced image through inverse Fourier transform.

[0029] When performing image brightness enhancement using a frequency domain processing network, some detail degradation is inevitable because suppressing noise may also suppress some detail information. Therefore, it is necessary to integrate frequency and spatial information in the spatial domain processing network to restore details. To address this, this invention proposes setting up a joint spatial-frequency attention module in the spatial domain processing network. This module leverages the complementarity of spatial and frequency domain information to enhance the network model's ability to capture detailed features while suppressing redundant or noise components. Specifically, the spatial domain processing network includes a joint spatial-frequency attention module. This module converts the initial enhanced image into a second frequency domain image, extracts frequency domain weights from the second frequency domain image, obtains spatial domain weights from the initial enhanced image, and combines the frequency and spatial weights before applying them to the initial enhanced image.

[0030] In some embodiments, the processing procedure of the space-frequency joint attention module is as follows: Figure 3As shown, the process of extracting frequency domain weights from the second frequency domain image includes: performing channel attention operations on the amplitude components in the second frequency domain image, and using the resulting weights together with the phase components in the second frequency domain image as the frequency domain weights. This process extracts the global energy distribution (such as high-frequency details and low-frequency background) through Fourier transform, decomposes the signal into different frequency components, and assigns dynamic weights to different frequency bands through an attention mechanism, achieving efficient utilization of frequency domain information. The process of obtaining spatial domain weights from the preliminary enhancement image includes: performing multiple consecutive convolution operations on the preliminary enhancement image to obtain spatial domain weights; combining the frequency domain weights and spatial domain weights through a weighted sum and applying them to the preliminary enhancement image. The weighted fusion of the dual-domain features ensures that the final enhancement image retains both local details and enhances global structural information.

[0031] In this embodiment of the invention, the process of extracting frequency domain weights from the second frequency domain image specifically includes: performing a 1×1 convolution operation on the amplitude component in the second frequency domain image, followed by a channel attention operation, and then performing a 1×1 convolution operation on the output features of the channel attention operation. The weights obtained at this point are then combined with the phase component in the second frequency domain image to form the frequency domain weights. The process of obtaining spatial domain weights from the preliminary enhancement image specifically includes: performing a 3×3 convolution operation, a ReLU activation operation, and another 3×3 convolution operation on the preliminary enhancement image consecutively to obtain the spatial domain weights. The frequency domain weights and spatial domain weights are then combined by a weighted sum using the following formula: M sf =αM freq +βM spa ; Among them, M sf M represents the combined weights. freq M represents the frequency domain weights. spa This represents the spatial weight.

[0032] The final enhanced image is obtained by applying the combined weights to the initial enhanced image using the following formula: ; Among them, X output X represents the final enhanced image. input This indicates the initial image enhancement.

[0033] This invention also provides a method for enhancing the image quality of low-light images, such as... Figure 4 As shown, it includes: S1: Obtain a low-light image dataset and preprocess it to obtain a training set. In this embodiment of the invention, the preprocessing process includes cropping the input low-light images to 384×384 pixels and performing data augmentation using random rotation to obtain the training set.

[0034] S2: Construct a low-light image quality enhancement system as provided in this invention, and train the low-light image quality enhancement system using the training set constructed in step S1 to obtain an image quality enhancement model.

[0035] In some embodiments, the low-light image enhancement system is trained using the total loss function expressed in the following formula: L=ξL s1 +μL s2 +ηL vgg ; Where L represents the total loss function, L s1 L represents the loss function of the frequency domain processing network. s2 L represents the loss function of the spatial processing network. vgg Let ξ represent the perceptual loss function implemented based on the VGG network, and let ξ, μ, and η represent the loss weights.

[0036] Loss function L of frequency domain processing network s1 for: ; Where F represents the Fourier transform, A represents the amplitude component extraction operation, and Output s1 This represents the initial enhanced image, where GT represents the corresponding true specular image. This indicates the operation of taking the 2-norm.

[0037] Loss function L of spatial processing network s2 for: ; Among them, Output s2 This indicates the final enhanced image.

[0038] Perceptual loss function L vgg for: ; in, This represents the features obtained from VGG network processing.

[0039] This embodiment of the invention was performed on Ubuntu 20.04, using an NVIDIA RTX 3090 GPU processor for system training. The software environment consisted of CUDA 11.2, Python 3.9, and a 64-bit operating system. During training, all modules were initialized from scratch using random weights and the Adam optimizer was used with a momentum set to 0.9. The learning rate was initialized to 4.0 × 10⁻⁶. 4 A multi-step scheduler was used with a batch size of 4 and a total number of training iterations of 2.0 × 10⁻⁶. 5 .

[0040] S3: Input the low-light image to be enhanced into the image quality enhancement model obtained in step S2, and output the corresponding final enhanced image.

[0041] To verify the network's image enhancement capabilities for complex lighting images, this embodiment of the invention selected three widely used paired LLIE datasets for qualitative and quantitative evaluation: LOL, LSRW-Huawei, and LSRW-Nikon. Simultaneously, images were captured in real-world environments, and the images were processed using the interface provided by Adobe Lightroom to match the image histograms with the illumination distribution of low-light images in conventional datasets. A self-made test dataset was then synthesized to verify the network's generalization ability in real-world scenarios. Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) were used as objective evaluation metrics.

[0042] The low-light images from the self-made test dataset were input into the image quality enhancement model obtained in step S2. It was found that the image quality enhancement model obtained by the present invention improved the PSNR index by 4.47% after enhancing the low-light images in the LOL dataset, and improved the SSIM index by 1.45% after enhancing the low-light images in the LSRW dataset. It can be seen that the present invention can significantly improve the image quality under non-uniform illumination and has good generalization ability.

[0043] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0044] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A low-light image quality enhancement system, characterized in that, include: The frequency domain processing network includes a domain transformation unit, a space-frequency mapping weight extraction unit, and a multi-scale gated convolution unit. Specifically: the domain transformation unit converts the input low-light image into a first frequency domain image; the space-frequency mapping weight extraction unit performs high-frequency noise compression on the first frequency domain image and extracts global illumination features to obtain space-frequency mapping weights; the multi-scale gated convolution unit, guided by the space-frequency mapping weights, performs gated convolution operations at different scales on the first frequency domain image, and the resulting features are then input into the domain transformation unit to obtain a preliminary enhanced image. The spatial processing network includes a spatial-frequency joint attention module. The spatial-frequency joint attention module converts the initial enhanced image into a second frequency domain image, extracts frequency domain weights from the second frequency domain image, obtains spatial domain weights from the initial enhanced image, and applies the combination of frequency domain weights and spatial domain weights to the initial enhanced image to obtain the final enhanced image.

2. The low-light image quality enhancement system according to claim 1, characterized in that, In the space-frequency mapping weight extraction unit, a convolution operation is performed on the first frequency domain image, and the features after the convolution operation are then processed by a fully connected layer to obtain the space-frequency mapping weights.

3. The low-light image quality enhancement system according to claim 1, characterized in that, In multi-scale gated convolutional units, multi-scale gated convolution is performed using the following formula: G k =Conv k ([X,h t-1 ]); Among them, G k Conv represents the output features of the input first frequency domain image X after being gated convolution at the k-th scale. k H represents gated convolution. t-1 [,] represents the previous hidden state in the gated convolution at the current scale, and [,] represents the first frequency domain image X and the previous hidden state h. t-1 Perform channel splicing; The space-frequency mapping weights are extracted at multiple scales, and the multi-scale gated convolution output features are fused using the space-frequency mapping weights guided by the following formula: ； Among them, G freq This indicates the characteristics after guided fusion. Represents the space-frequency mapping weights at the k-th scale; The fused features are combined with the first frequency domain image to obtain the output features of the multi-scale gated convolutional unit.

4. The low-light image quality enhancement system according to claim 3, characterized in that, In multi-scale gated convolutional units, the hidden state is penalized using the following formula: ； Among them, h t Let represent the current hidden state after penalty, z represent the update gate in the gated convolution, c represent the candidate state, λ represent the penalty coefficient, and MLP represent the multilayer perceptron. This indicates element-wise multiplication.

5. The low-light image quality enhancement system according to claim 3, characterized in that, After the guided fusion features are added to the amplitude component of the first frequency domain image, they are then input together with the phase component of the first frequency domain image into the domain conversion unit to obtain a preliminary enhanced image.

6. The low-light image quality enhancement system according to claim 1, characterized in that, In the space-frequency joint attention module: The process of extracting frequency domain weights from the second frequency domain image includes: performing channel attention operation on the amplitude component in the second frequency domain image, and using the obtained weights together with the phase component in the second frequency domain image as frequency domain weights; The process of obtaining spatial weights from the initial enhanced image includes: performing multiple consecutive convolution operations on the initial enhanced image to obtain the spatial weights; The combined weighted sum of the frequency domain weights and the spatial domain weights is applied to the initial enhanced image to obtain the final enhanced image.

7. A method for enhancing image quality in low-light conditions, characterized in that, include: S1: Obtain the low-light image dataset and preprocess it to obtain the training set; S2: Construct a low-light image quality enhancement system as described in any one of claims 1 to 6, and train the low-light image quality enhancement system using the training set constructed in step S1 to obtain an image quality enhancement model; S3: Input the low-light image to be enhanced into the image quality enhancement model obtained in step S2, and output the corresponding final enhanced image.

8. The low-light image quality enhancement method according to claim 7, characterized in that, In step S2, the low-light image enhancement system is trained using the total loss function expressed in the following formula: L=ξL s1 +μL s2 +ηL vgg ; Where L represents the total loss function, L s1 L represents the loss function of the frequency domain processing network. s2 L represents the loss function of the spatial processing network. vgg Let ξ represent the perceptual loss function implemented based on the VGG network, and let ξ, μ, and η represent the loss weights.

9. The low-light image quality enhancement method according to claim 8, characterized in that, Loss function L of frequency domain processing network s1 for: ； Where F represents the Fourier transform, A represents the amplitude component extraction operation, and Output s1 This represents the initial enhanced image, where GT represents the corresponding true specular image. This represents the 2-norm operation; Loss function L of spatial processing network s2 for: ； Among them, Output s2 This represents the final enhanced image; Perceptual loss function L vgg for: ； in, This represents the features obtained from VGG network processing.

Images