A low-light image quality enhancement method and device based on Retinex theory

Abstract

The present invention discloses a method and device for enhancing the quality of low-light images based on Retinex theory, which belongs to the field of image processing technology. The method comprises: constructing a low-light image enhancement model based on Retinex theory; including: reconstructing a loss function and a color correction function; collecting a number of low-light RGB images and their corresponding conventional illumination RGB images in different scenes, inputting them into the low-light image enhancement model for training until the overall loss function converges, and obtaining a trained model; inputting the low-light image to be enhanced into the trained model for processing, and achieving image quality enhancement by suppressing noise and reconstructing the image illumination component. The method can solve the problems of low image visibility and low signal-to-noise ratio. The constructed model and loss function decouple and reconstruct the image illumination component in an interpretable manner to enhance the image contrast, effectively improving the detail reconstruction quality and color accuracy of low-light image enhancement, and achieving better visual effects.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more specifically to a method and device for enhancing the quality of low-illumination images based on Retinex theory, which are used for improving the quality of low-illumination images. Background Art

[0002] At present, low-light image enhancement technology is usually used to solve the problems of low image visibility and low signal-to-noise ratio caused by insufficient lighting, uneven lighting, backlight and other environmental conditions. Traditional methods based on histogram equalization and image decomposition models will amplify noise while improving the overall brightness of the image. With the development of deep learning technology, low-light enhancement methods based on convolutional neural networks such as LLNet, EnlightenGAN, SID, etc. can suppress noise while improving image visibility. However, although the existing low-light image enhancement methods can improve image quality to a certain extent, they often have varying degrees of detail over-smoothing and color cast, which seriously affect the image quality. More importantly, existing methods are often based on end-to-end networks, with poor interpretability, and are difficult to apply to scenarios with high reliability requirements. Summary of the invention

[0003] In view of this, the present invention provides a low-illumination image quality enhancement method and device based on Retinex theory to solve the problems of low visibility and low signal-to-noise ratio of low-illumination images in the prior art.

[0004] In order to achieve the above object, the present invention adopts the following technical solution:

[0005] In a first aspect, an embodiment of the present invention provides a low-illumination image quality enhancement method based on Retinex theory, comprising the following steps:

[0006] Construction steps: constructing a low illumination image enhancement model based on Retinex theory; the low illumination image enhancement model includes: a low illumination degradation feature extractor, a global space feature fusion module, an illumination mapping module, an image restoration module and a multi-level feature fusion module; wherein the loss function of the low illumination image enhancement model includes: a reconstruction loss function and a color correction function;

[0007] Training step: In different scenes, a number of low-light RGB images and their corresponding regular-light RGB images are collected and input into the low-light image enhancement model for training until the overall loss function converges to obtain a trained model;

[0008] Enhancement step: The low-light image to be enhanced is input into the trained low-light image enhancement model for processing, and the image quality is enhanced by suppressing noise and reconstructing the image illumination component.

[0009] Furthermore, in the construction step, the modeling formula of the low-light image enhancement model is:

[0010] (1)

[0011] in, represents the ideal image illumination component, L0 represents the low illumination component, X represents the low illumination image, N represents the image noise; Y represents the enhanced result of the model.

[0012] Furthermore, the low illumination degradation feature extractor is composed of a convolutional coding layer and three NAF blocks in cascade, and the NAF blocks perform spatial 2×2 downsampling and channel number multiplication through convolution with a step size of 2;

[0013] The global spatial feature fusion module is composed of 4 ViT blocks cascaded;

[0014] The illumination mapping module and the image restoration module are both cascaded by three NAF blocks and a convolutional decoding layer, and the number of channels is halved and the resolution is doubled through the pixel shuffle operation between the NAF blocks;

[0015] The multi-level feature fusion module is composed of 3 layers of convolution and residual connection structure; it is arranged on the jump connection structure between the low illumination degradation feature extractor and the illumination mapping module, and on the jump connection structure between the low illumination degradation feature extractor and the image restoration module.

[0016] Furthermore, in the enhancement step, the low-light image to be enhanced is input into a trained low-light image enhancement model for processing, including:

[0017] Inputting the low illumination image to be enhanced into the low illumination degradation feature extractor to extract illumination degradation features;

[0018] Since global consistency is required for reconstructing illumination, the model uses the ViT structure of the global spatial feature fusion module to extract global illumination degradation features;

[0019] The model uses an image restoration module to reconstruct the residual noise component and suppresses the noise through pixel-by-pixel subtraction;

[0020] The illumination mapping module gradually reconstructs the image brightness mapping parameters and multiplies them pixel by pixel with the denoised image to improve the image contrast. The model enhances the image by estimating the noise and illumination mapping parameters as follows:

[0021] (2)

[0022] Among them, Y represents the enhancement result of the model, X is the input low-light image to be enhanced, is the noise estimated by the image restoration module, The mapping relationship estimated by the illumination mapping module.

[0023] Furthermore, the reconstruction loss function is expressed as follows:

[0024] (3)

[0025] The color correction function is expressed as follows:

[0026] (4)

[0027] in, To enhance the model results, is the target enhancement result, || ||1 is the l1 norm, {a,b} represents the image chromaticity channel defined by the CIELab color space, is the chromaticity channel i of the target image Y represented by the CIE Lab color space, Output of the model represented by CIE Lab color space The chromaticity channel i, where channel i∈{a,b}, represents the perceived loss, represents the perceptual loss weight hyperparameter.

[0028] In a second aspect, an embodiment of the present invention further provides a low-illumination image quality enhancement device based on Retinex theory, comprising:

[0029] Construction module: construct a low illumination image enhancement model based on Retinex theory; the low illumination image enhancement model includes: a low illumination degradation feature extractor, a global space feature fusion module, an illumination mapping module, an image restoration module and a multi-level feature fusion module; wherein the loss function of the low illumination image enhancement model includes: a reconstruction loss function and a color correction function;

[0030] Training module: In different scenes, a number of low-light RGB images and their corresponding regular-light RGB images are collected and input into the low-light image enhancement model for training until the overall loss function converges to obtain a trained model;

[0031] Enhancement module: The low-light image to be enhanced is input into the trained low-light image enhancement model for processing, and image quality is enhanced by suppressing noise and reconstructing the image illumination component.

[0032] In a third aspect, an embodiment of the present invention further provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0033] Memory, used to store computer programs;

[0034] The processor, when used to execute the program stored in the memory, can implement a low-illumination image quality enhancement method based on Retinex theory as described in any one of the first aspects.

[0035] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes a low-illumination image quality enhancement method based on Retinex theory as described in any one of the first aspects.

[0036] The descriptions of the second to fourth aspects of the present invention can refer to the detailed description of the first aspect; and the beneficial effects of the descriptions of the second to fourth aspects can refer to the analysis of the beneficial effects of the first aspect, which will not be repeated here.

[0037] It can be seen from the above technical solution that compared with the prior art, it has the following technical advantages:

[0038] The model and loss function constructed by the present invention decouple and reconstruct the image illumination component in an interpretable way to enhance the image contrast, and perform a certain degree of image restoration for the problem of low signal-to-noise ratio in low-illumination images. This method effectively improves the detail reconstruction quality and color accuracy of low-illumination image enhancement, can achieve better visual effects, and can be applied to scenes with high reliability and interpretability requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0040] Figure 1 A schematic diagram of the structure of a low-illumination image illumination reconstruction and denoising network model provided by the present invention;

[0041] Figure 2 Schematic diagram of the structure of each module of the low-light image enhancement model of the method of the present invention;

[0042] Figure 3 A diagram showing a low-light image input of scene 1 of an embodiment of the present invention, an enhancement result of the method of the present invention, and an enhancement result of the ChebyLighter model;

[0043] Figure 4It is a display diagram of the low-light image input of scene 2 of an embodiment of the present invention, the enhancement result of the method of the present invention, and the enhancement result of the ChebyLighter model;

[0044] Figure 5 It is a display diagram of the low-light image input of scene three of an embodiment of the present invention, the enhancement result of the method of the present invention, and the enhancement result of the ChebyLighter model. DETAILED DESCRIPTION

[0045] 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.

[0046] Retinex theory is an image processing method based on the human visual system, which can effectively separate the reflection component and illumination component of an image, thereby enhancing the contrast and details of the image while ensuring color reproduction. Starting from the Retinex theory, the present invention solves the problems of low visibility and low signal-to-noise ratio in low illumination. The constructed model and loss function decouple and reconstruct the image illumination component in an interpretable way to enhance the image contrast, and perform a certain degree of image restoration for the problem of low signal-to-noise ratio of low illumination images. It effectively improves the detail reconstruction quality and color accuracy of low illumination image enhancement, and can achieve better visual effects.

[0047] Embodiment 1:

[0048] The embodiment of the present invention discloses a low-illumination image quality enhancement method based on Retinex theory, comprising the following steps:

[0049] Construction steps: Construct a low-light image enhancement model based on Retinex theory; refer to Figure 1 As shown, the low illumination image enhancement model includes: a low illumination degradation feature extractor, a global space feature fusion module, an illumination mapping module, an image restoration module and a multi-level feature fusion module; wherein the loss function of the low illumination image enhancement model includes: a reconstruction loss function and a color correction function;

[0050] Training step: In different scenes, a number of low-light RGB images and their corresponding regular-light RGB images are collected and input into the low-light image enhancement model for training until the overall loss function converges to obtain a trained model;

[0051] Enhancement step: The low-light image to be enhanced is input into the trained low-light image enhancement model for processing, and the image quality is enhanced by suppressing noise and reconstructing the image illumination component.

[0052] The following is a detailed description of each of the above steps:

[0053] 1. In the construction step, a low-light image enhancement model is designed based on the Retinex theory. According to the Retinex theory, the enhancement process of the low-light image X can be modeled as:

[0054] (1)

[0055] in, represents the ideal image illumination component, L0 represents the low illumination component, N represents the image noise, represents element-by-element multiplication. Figure 1 As shown, the network generates a mapping from low illumination to ideal illumination. And extract the noise component N to enhance the image.

[0056] like Figure 2 As shown in Figure 2, the low-light image enhancement model consists of a low-light degradation feature extractor ( Figure 2 As shown in part a), global spatial feature fusion module ( Figure 2 As shown in part b), illumination mapping module, image restoration module ( Figure 2 c) and a multi-level feature fusion module ( Figure 2 The low-illumination degradation feature extractor is composed of a convolutional coding layer and three NAF blocks in cascade. The NAF blocks perform spatial 2×2 downsampling and channel multiplication through convolution with a step size of 2. That is, if the height×width×channel number of the first layer of convolutional coding features of the input image is H×W×C, the feature extractor outputs a feature size of H / 8×W / 8×8C. Since the ViT structure has good global modeling capabilities, the global spatial feature fusion module is composed of four cascaded ViT blocks; the illumination mapping module is composed of three NAF blocks and a convolutional decoding layer, and the pixel shuffle operation is used between the NAF blocks to halve the number of channels and double the resolution; the image restoration module structure is the same as the illumination mapping module; there is a jump connection structure between the low-illumination degradation feature extractor and the illumination mapping module and the image restoration module, and a multi-level feature fusion module is built on the jump connection structure. The multi-level feature fusion module consists of three layers of convolution and residual connection structure, and the high-resolution shallow-level features are filtered to further retain the clean structural features.

[0057] When enhancing an image, the low-illumination image enhancement model first inputs the image into the low-illumination degradation feature extractor to extract illumination degradation features. Since global consistency is required for illumination reconstruction, the model uses the ViT structure of the global spatial feature fusion module to effectively extract global illumination degradation features. Then, the model uses the image restoration module to reconstruct the residual noise component and suppresses noise by pixel-by-pixel subtraction. Finally, the illumination mapping module gradually reconstructs the image brightness mapping parameters and multiplies them pixel by pixel with the denoised image to effectively improve the image contrast. The process of enhancing the image by estimating the noise and illumination mapping parameters can be described as follows:

[0058] (2)

[0059] Among them, Y represents the enhanced result of the model, X is the input low-light image, is the noise estimated by the image restoration module, The mapping relationship estimated by the illumination mapping module.

[0060] 2. In the training step:

[0061] When training the low-light image enhancement model, the loss function used is composed of the content reconstruction loss L rec and color correction loss L color The expressions are as follows:

[0062] (3)

[0063] (4)

[0064] in, To enhance the model results, is the target enhancement result, || ||1 is the l1 norm, {a,b} represents the image chromaticity channel defined by the CIELab color space, is the chromaticity channel i of the target image Y represented by the CIE Lab color space, Output of the model represented by CIE Lab color space The chromaticity channel i, where channel i∈{a,b}, represents the perceived loss, represents the perceptual loss weight hyperparameter.

[0065] The specific process of training is as follows:

[0066] 2.1 Data preprocessing:

[0067] Data preprocessing is performed on the collected low-illumination RGB images and regular-illumination RGB images, including image normalization, noise filtering, and size adjustment to ensure the consistency of the input images and the stability of the training model.

[0068] Conventional illumination RGB images refer to color images taken under normal or ideal lighting conditions, which have uniform lighting distribution, appropriate brightness and contrast, and true color reproduction. In contrast, low-illuminance RGB images refer to images taken under insufficient lighting conditions, which usually show problems such as insufficient brightness, increased noise, and color distortion.

[0069] Regular illumination RGB images are used as references or targets for low illumination image enhancement models, with the goal of training the model to process low illumination RGB images as close as possible to the image quality under regular illumination. Such images serve as "ideal outputs" or "target images" during training, guiding the model on how to adjust low illumination images to make them close to the effects under regular illumination in terms of brightness, contrast, detail, and color reproduction.

[0070] 2.2 Model initialization:

[0071] According to the characteristics of Retinex theory, the parameters of the low illumination degradation feature extractor, the global space feature fusion module, the illumination mapping module, the image restoration module and the multi-level feature fusion module are initialized so as to be able to effectively extract and process the features of the image.

[0072] 2.3 Model training:

[0073] The preprocessed low-illumination RGB image and normal-illumination RGB image are input into the model. During the forward propagation process, the model is processed by each module in turn, image features are extracted, feature fusion and illumination mapping are performed, and finally the enhanced image is obtained.

[0074] During the back-propagation process, the error between the model output and the target enhanced image is calculated through the loss function, and the model parameters are updated.

[0075] Reconstruction loss function: used to measure the difference in brightness and details between the enhanced image and the conventional illumination image to ensure the overall improvement of image quality.

[0076] Color correction loss function: used to correct the color cast of the enhanced image and ensure the color restoration of the enhanced image.

[0077] 2.4 Training Optimization:

[0078] During the training process, a dynamic learning rate adjustment strategy is adopted to accelerate convergence and prevent overfitting.

[0079] The complexity of the model is suppressed by introducing regularization terms to avoid overfitting of the model on the training data.

[0080] 2.5 Loss function convergence:

[0081] As the training progresses, the changes in the loss function are monitored until the overall loss function converges, that is, the output results of the enhanced model are stable and the model reaches the optimal state.

[0082] 2.6 Model Validation:

[0083] After training, the model is applied to unseen low-light images for verification to evaluate its enhancement effect and generalization ability. Based on the verification results, the model parameters are further adjusted or optimized to ensure the performance of the model in practical applications.

[0084] Through the above training steps, the model can more accurately understand and enhance low-light images, and ultimately achieve high-quality image enhancement effects.

[0085] 3. In the enhancement step, the following processes are included:

[0086] 3.1 Image input:

[0087] The low-light RGB image to be enhanced is input into the trained low-light image quality enhancement model. At this point, the various modules of the model (including the low-light degradation feature extractor, global spatial feature fusion module, illumination mapping module, image restoration module, and multi-level feature fusion module) have been fully trained and can accurately identify and process degradation features in low-light images.

[0088] 3.2 Feature Extraction:

[0089] The model first analyzes the input low-light image through the low-light degradation feature extractor to extract image degradation features, such as uneven illumination, noise, etc. These features are used for image enhancement processing in subsequent steps.

[0090] 3.3 Global inter-domain feature fusion:

[0091] In the global spatial feature fusion module, the model performs a global feature fusion of the extracted degraded features and the input image. By analyzing and fusing the overall brightness, contrast, and local details of the image, a comprehensive enhanced feature representation is generated.

[0092] 3.4 Illuminance Mapping:

[0093] In the illumination mapping module, the model generates an illumination mapping parameter based on the fused feature information. This parameter is used to adjust the brightness and contrast of the image to make it close to the ideal state under normal illumination.

[0094] 3.5 Noise Suppression:

[0095] The model then uses the image restoration module to perform denoising on the image. By estimating and suppressing noise, the noise level in the image is reduced while retaining the effective information and details of the image.

[0096] 3.6 Multi-level feature fusion:

[0097] In the multi-level feature fusion module, the model further fuses features at different levels of the image to ensure the consistency of the enhanced image at all levels, which not only retains the overall brightness and contrast improvement, but also ensures the clarity of the details.

[0098] 3.7 Image Output:

[0099] Finally, the enhanced image is output. At this point, the image's brightness, contrast, detail level, and color reproduction are significantly improved, approaching the image quality captured under normal illumination. The model can effectively enhance low-illumination images, making them close to normal illumination images in terms of brightness, contrast, noise level, and detail retention, significantly improving image quality and providing a better foundation for subsequent image processing tasks (such as recognition, detection, etc.).

[0100] like Figure 3 As shown, the leftmost side is a low-light image of a basketball court, the middle side is the enhancement result after applying the method of the present invention, and the rightmost side is the enhancement result of the ChebyLighter model. The three can form an obvious contrast, and the enhancement result of the present invention is obviously better than the enhancement result of the ChebyLighter model.

[0101] like Figure 4 As shown, the leftmost side is a low-light image of an outdoor scene, the middle side is the enhancement result after applying the method of the present invention, and the rightmost side is the enhancement result of the ChebyLighter model. The three can form an obvious contrast, and the enhancement result of the present invention is obviously better than the enhancement result of the ChebyLighter model.

[0102] like Figure 5 As shown, the leftmost side is a low-light image of merchandise displayed in a window, the middle side is the enhancement result after applying the method of the present invention, and the rightmost side is the enhancement result of the ChebyLighter model. The three can form an obvious contrast, and the enhancement result of the present invention is obviously better than the enhancement result of the ChebyLighter model.

[0103] It can be seen that the low-light image quality enhancement method based on Retina theory provided by the present invention can solve the problems of low visibility and low signal-to-noise ratio in low light conditions. The method can significantly improve the quality of low-light images, including enhancing image brightness, reducing noise, improving details and contrast, and effectively avoiding color distortion problems. The method not only significantly improves subjective visual quality, but also performs well in objective evaluation indicators, and has high practical value.

[0104] Embodiment 2:

[0105] Based on the same inventive concept, an embodiment of the present invention also provides a low-illumination image quality enhancement device based on Retinex theory. Since the principle of the problem solved by the device is similar to that of the aforementioned low-illumination image quality enhancement method based on Retinex theory, the implementation of the device can refer to the implementation of the aforementioned method, and the repeated parts will not be repeated.

[0106] The embodiment of the present invention further provides a low-illumination image quality enhancement device based on Retinex theory, comprising:

[0107] Construction module: construct a low illumination image enhancement model based on Retinex theory; the low illumination image enhancement model includes: a low illumination degradation feature extractor, a global space feature fusion module, an illumination mapping module, an image restoration module and a multi-level feature fusion module; wherein the loss function of the low illumination image enhancement model includes: a reconstruction loss function and a color correction function;

[0108] Training module: In different scenes, a number of low-light RGB images and their corresponding regular-light RGB images are collected and input into the low-light image enhancement model for training until the overall loss function converges to obtain a trained model;

[0109] Enhancement module: The low-light image to be enhanced is input into the trained low-light image enhancement model for processing, and image quality is enhanced by suppressing noise and reconstructing the image illumination component.

[0110] In this embodiment, the device can perform model construction, training, and enhancement processing of the image to be processed, significantly improving the brightness, contrast, and color restoration capabilities of low-light images. Ultimately, the device provides a stable and efficient image enhancement solution that is suitable for a variety of image processing applications in low-light environments.

[0111] Embodiment 3:

[0112] The embodiment of the present invention further provides a computer device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0113] Memory, used to store computer programs;

[0114] The processor, when used to execute the program stored in the memory, can implement the low-illumination image quality enhancement method based on Retinex theory as described in Example 1.

[0115] In this embodiment, the processor is responsible for executing the computer program to realize the image enhancement function. The communication interface is used for communication connection between devices to ensure data transmission and collaborative work. The memory is used to store the computer program, which can realize the low-light image quality enhancement method based on Retinex theory. The communication bus is used to connect the processor, the communication interface and the memory to ensure efficient communication between them.

[0116] In this embodiment, the processor can implement a low-illumination image quality enhancement method based on Retinex theory by executing a program stored in the memory.

[0117] The computer device of this embodiment can run the image enhancement algorithm in a complex computing environment. It can be flexibly integrated into different hardware systems to achieve real-time enhancement processing of low-light images. The implementation of this device enhances the functional extensibility of the device and is suitable for various application scenarios that require improved image quality, such as monitoring systems, camera image processing, etc.

[0118] Embodiment 4:

[0119] An embodiment of the present invention further provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes a low-illumination image quality enhancement method based on Retinex theory as in Embodiment 1.

[0120] The present invention provides a convenient algorithm deployment method by means of a computer-readable storage medium. Users can quickly deploy and run the low-light image enhancement algorithm by loading the program in the storage medium without additional development or configuration work. This greatly simplifies the application and promotion of the algorithm and improves the user experience. This embodiment is particularly suitable for application scenarios that require rapid implementation of image enhancement functions, such as embedded systems, image processing software, etc.

[0121] These embodiments provide a variety of methods and approaches for achieving low-light image quality enhancement for hardware devices, computer devices, and storage media, respectively. Each solution has unique technical advantages and can meet the needs of different application scenarios.

[0122] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0123] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A low-light image quality enhancement method based on Retinex theory, characterized in that: The following steps are involved: Construction steps: constructing a low illumination image enhancement model based on Retinex theory; the low illumination image enhancement model includes: a low illumination degradation feature extractor, a global space feature fusion module, an illumination mapping module, an image restoration module and a multi-level feature fusion module; wherein the loss function of the low illumination image enhancement model includes: a reconstruction loss function and a color correction function; Training step: In different scenes, a number of low-light RGB images and their corresponding regular-light RGB images are collected and input into the low-light image enhancement model for training until the overall loss function converges to obtain a trained model; Enhancement step: input the low-light image to be enhanced into the trained low-light image enhancement model for processing, and achieve image quality enhancement by suppressing noise and reconstructing the image illumination component; The low illumination degradation feature extractor is composed of a convolutional coding layer and three NAF blocks in cascade, and the NAF blocks perform spatial 2×2 downsampling and channel multiplication through convolution with a step size of 2; The global spatial feature fusion module is composed of 4 ViT blocks cascaded; The illumination mapping module and the image restoration module are both cascaded by three NAF blocks and a convolutional decoding layer, and the number of channels is halved and the resolution is doubled through the pixel shuffle operation between the NAF blocks; The multi-level feature fusion module is composed of 3 layers of convolution and residual connection structure; it is arranged on the jump connection structure between the low illumination degradation feature extractor and the illumination mapping module, and on the jump connection structure between the low illumination degradation feature extractor and the image restoration module.

2. The low-light image quality enhancement method based on Retinex theory according to claim 1, characterized in that: In the construction step, the modeling formula of the low-light image enhancement model is: Among them, L ideal represents the ideal image illumination component, L0 represents the low illumination component, X represents the low illumination image, N represents the image noise; Y represents the enhanced result of the model.

3. The low-light image quality enhancement method based on Retinex theory according to claim 1, characterized in that: In the enhancement step, the low-light image to be enhanced is input into the trained low-light image enhancement model for processing, including: Inputting the low illumination image to be enhanced into the low illumination degradation feature extractor to extract illumination degradation features; Since global consistency is required for reconstructing illumination, the model uses the ViT structure of the global spatial feature fusion module to extract global illumination degradation features; The model uses an image restoration module to reconstruct the residual noise component and suppresses the noise through pixel-by-pixel subtraction; The illumination mapping module gradually reconstructs the image brightness mapping parameters and multiplies them pixel by pixel with the denoised image to improve the image contrast. The model enhances the image by estimating the noise and illumination mapping parameters as follows: Among them, Y represents the enhancement result of the model, X is the input low-light image to be enhanced, is the noise estimated by the image restoration module, The mapping relationship estimated by the illumination mapping module.

4. The low-light image quality enhancement method based on Retinex theory according to claim 1, characterized in that: The reconstruction loss function is expressed as follows: The color correction function is expressed as follows: in, is the model enhancement result, Y i is the target enhancement result, ||||1 is the l1 norm, {a,b} represents the image chromaticity channel defined by the CIE Lab color space, is the chromaticity channel i of the target image Y represented by the CIE Lab color space, Output of the model represented by CIE Lab color space The chromaticity channel i, where channel i∈{a,b},L per represents the perceptual loss, λ per represents the perceptual loss weight hyperparameter.

5. A low-light image quality enhancement device based on Retinex theory, characterized in that: include: Building module: Building a low-light image enhancement model based on Retinex theory; The low illumination image enhancement model comprises: a low illumination degradation feature extractor, a global space feature fusion module, an illumination mapping module, an image restoration module and a multi-level feature fusion module; wherein the loss function of the low illumination image enhancement model comprises: a reconstruction loss function and a color correction function; Training module: In different scenes, a number of low-light RGB images and their corresponding regular-light RGB images are collected and input into the low-light image enhancement model for training until the overall loss function converges to obtain a trained model; Enhancement module: The low-light image to be enhanced is input into the trained low-light image enhancement model for processing, and image quality is enhanced by suppressing noise and reconstructing the image illumination component; The low illumination degradation feature extractor is composed of a convolutional coding layer and three NAF blocks in cascade, and the NAF blocks perform spatial 2×2 downsampling and channel multiplication through convolution with a step size of 2; The global spatial feature fusion module is composed of 4 ViT blocks cascaded; The illumination mapping module and the image restoration module are both cascaded by three NAF blocks and a convolutional decoding layer, and the number of channels is halved and the resolution is doubled through the pixel shuffle operation between the NAF blocks; The multi-level feature fusion module is composed of 3 layers of convolution and residual connection structure; it is arranged on the jump connection structure between the low illumination degradation feature extractor and the illumination mapping module, and on the jump connection structure between the low illumination degradation feature extractor and the image restoration module.

6. A computer device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when used to execute the program stored in the memory, can implement a low-illumination image quality enhancement method based on Retinex theory as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, which, when executed by a processor, enable the processor to execute a low-illumination image quality enhancement method based on Retinex theory as described in any one of claims 1 to 4.

Images