Low-illumination image enhancement method based on layered illumination guidance

Abstract

The invention discloses a low-illumination image enhancement method based on layered illumination guidance, and the method employs a trained low-illumination image enhancement model to carry out the image enhancement processing of a low-illumination image, and the low-illumination image enhancement model comprises a layered illumination estimator module, a dual-illumination feature modulation module, and a self-adaptive enhancement module. A low-level visual feature map and a high-level semantic feature map are extracted through a layered illumination estimator module, a dual-illumination feature modulation module modulates the feature maps through an illumination guide wavelet modulation unit and an illumination guide feature modulation unit respectively, and a self-adaptive enhancement module splices the modulated feature maps. And carrying out feature adaptive fusion and image-level enhancement processing on the spliced features, and finally outputting an enhanced illumination enhancement result image. According to the method, multi-domain reconstruction features are fused by combining multi-granularity illumination estimation and wavelet multi-scale analysis, and the robustness of a low-illumination image enhancement effect is improved based on residual learning.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a low-light image enhancement method based on layered illumination guidance. Background Art

[0002] As an important carrier of visual information, the quality of images directly affects the performance of computer vision systems. In actual application scenarios, due to the physical limitations of imaging equipment and insufficient ambient light, low-light images generally have problems such as difficulty in identifying visual information, low local contrast, and severe noise interference. These degradations not only reduce the visual quality of the image, but also significantly affect the accuracy of high-level visual tasks such as target detection, text recognition, and autonomous driving. Studies have shown that under low-light conditions, the performance of existing visual algorithms generally declines to varying degrees. Therefore, low-light image enhancement, as a fundamental issue in the field of computer vision, has important research value.

[0003] In recent years, low-light image enhancement methods have undergone a paradigm shift from traditional image processing to deep learning. Traditional methods mainly include histogram adjustment-based techniques and Retinex theory-based methods. Among them, histogram equalization (HE) improves image contrast by redistributing pixel grayscale values, but it often leads to over-enhancement and detail distortion. Retinex theory achieves enhancement by decomposing images into illumination and reflection components. Since then, a series of works have explored the application of Retinex theory in low-light image enhancement. However, such methods still have obvious deficiencies in the accuracy and robustness of illumination estimation, especially when dealing with complex illumination scenes, they are prone to halo artifacts and color distortion. In addition, although the variational model-based method has advantages in maintaining image structure and details, it is difficult to achieve adaptive local enhancement due to the lack of effective modeling of illumination priors.

[0004] With the development of deep learning, low-light image enhancement methods based on neural networks have made significant progress. Early works such as LLNet and LightenNet achieved joint optimization of image enhancement and denoising through end-to-end training. Subsequently, researchers began to focus on how to introduce physical priors into deep learning frameworks, such as RetinexNet and EnlightenGAN. Recent work has focused more on learning feature representations. For example, SNR-Aware considers the impact of noise level on enhancement quality, while LLFormer introduces the Transformer structure to capture long-range dependencies. However, there are still two key problems with existing deep learning methods: first, the lack of accurate modeling of local illumination distribution makes it difficult for the enhancement process to adapt to the illumination conditions of different regions; second, the illumination prior information is not fully utilized in the process of feature extraction and image reconstruction. Existing methods often do not make full use of illumination information and are difficult to cope with complex illumination degradation. Summary of the invention

[0005] In view of the above-mentioned deficiencies of the prior art, the present invention provides a low-light image enhancement method based on layered illumination guidance, which improves the robustness of the low-light image enhancement effect based on residual learning by combining multi-granularity illumination estimation and wavelet multi-scale analysis, and fusing multi-domain reconstruction features.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A low-light image enhancement method based on layered illumination guidance comprises the following steps:

[0008] S1, obtaining a low-light image to be processed;

[0009] S2, using the low-light image to be processed as an input of a trained low-light image enhancement model, and outputting an image enhancement result of the low-light image to be processed; the low-light image enhancement model includes a hierarchical illumination estimator module, a dual illumination feature modulation module and an adaptive enhancement module;

[0010] The training process of the low-light image enhancement model is as follows:

[0011] S201, the hierarchical illumination estimator module uses the low-light image as a training sample input to the low-light image enhancement model as the input of the shallow illumination estimator unit in the hierarchical illumination estimator module, performs shallow feature extraction through a multi-granular illumination mapping function to obtain a low-level visual feature map, and uses the low-light image as the input of the deep illumination estimator unit to output a high-level semantic feature map;

[0012] S202, the dual illumination feature modulation module modulates the low-level visual feature map and the high-level semantic feature map through an illumination-guided wavelet modulation unit and an illumination-guided feature modulation unit, respectively, to obtain a wavelet modulation feature map and a feature modulation feature map, respectively;

[0013] S203, the adaptive enhancement module splices the wavelet modulation feature map and the feature modulation feature map, performs feature adaptive fusion and image-level enhancement processing on the spliced ​​features, and finally outputs an enhanced illumination enhancement result image;

[0014] S204, using a multi-objective loss function consisting of a pixel-level reconstruction loss function, a structural similarity loss function, and a perceptual loss function to calculate the training loss of the low-light image enhancement model, and updating the parameters of the low-light image enhancement model with the goal of minimizing the multi-objective loss function;

[0015] S205. Repeat steps S201 to S204 until the low-light image enhancement model converges or reaches a preset number of training times, and the training is completed.

[0016] As a preferred solution, in step S201, the hierarchical illumination estimator module includes a shallow illumination estimator unit and a deep illumination estimator unit; wherein the shallow illumination estimator unit includes a cascaded color space transformation module, a convolution layer, a Gaussian filter layer, a multi-granularity illumination mapping function module and a normalization layer, and the deep illumination estimator unit includes a cascaded first convolution layer, an instance normalization layer, a ReLU activation function layer and a second convolution layer;

[0017] In the hierarchical illumination estimator module, the low-light image used as the input of the hierarchical illumination estimator module is used as the input of the shallow illumination estimator unit and the deep illumination estimator unit respectively, and the output of the shallow illumination estimator unit obtains a low-level visual feature map. The output of the shallow illumination estimator unit and the low-light image used as the input are both used as the input of the deep illumination estimator unit, and the output of the deep illumination estimator obtains a high-level semantic feature map.

[0018] As a preferred solution, the processing process of the shallow illumination estimator unit is:

[0019]

[0020]

[0021] I μ =ε shallow (I low )=μ(L norm );

[0022] In the formula, Represents the brightness map extracted by Gaussian filtering introduced by the shallow illumination estimator unit, Indicates the transformation operation from RGB space to HSV space and intercepts the last channel V, G σ represents a Gaussian kernel with standard deviation σ, represents the convolution operation, I low represents the input low-light image, L norm Indicates the normalization of the brightness image after Gaussian filtering, L min and L max They represent the minimum and maximum values ​​of the brightness map, ∈ represents the numerical stability constant, γ represents the normalization coefficient, μ d , μ n , μ b They represent the dark area, normal area, and bright area of ​​the illumination distribution map, respectively. d, n, and b represent dark, normal, and bright, respectively. (a, b) represents the characteristic threshold of the illumination interval. δ and β represent the steepness of the Sigmoid function. represents a three-channel light map for the shallow estimator, Represents each component, I μ Represents ε shallow The output low-level visual feature map, ε shallow (·) denotes a shallow lighting estimator unit.

[0023] The processing of the deep illumination estimator unit is:

[0024]

[0025] In the formula, F μ Represents ε deep The output high-level semantic feature map, ε deep (·) denotes a high-level illumination estimator unit, [·,·] denotes feature concatenation in the channel dimension, ReLU denotes the ReLU activation function, IN denotes instance normalization, and They represent the weights of the two convolutional layers respectively, and k represents the convolution kernel size.

[0026] As a preferred solution, in step S202, the dual illumination feature modulation module includes an illumination-guided wavelet modulation unit and an illumination-guided feature modulation unit; wherein the illumination-guided wavelet modulation unit includes a wavelet coefficient processor network, an illumination modulation network and a regional enhancement unit cascaded in sequence, and the illumination-guided feature modulation unit includes three feature extraction units, a connection layer, a feature fusion layer, a channel attention unit and a spatial attention unit in parallel, and the cascade results of the three parallel feature extraction units, the connection layer and the feature fusion layer are connected in parallel with the cascade results of the attention unit and the spatial attention unit, wherein the three parallel feature extraction units are respectively a feature extraction unit with a convolution kernel of 1, a feature extraction unit with a convolution kernel of 2 and a feature extraction unit with a convolution kernel of 3;

[0027] In the illumination-guided wavelet modulation unit, a low-light image as an input of a low-light image enhancement model is subjected to feature extraction to obtain a basic feature map, the basic feature map is used as an input of a wavelet coefficient processor network in the illumination-guided wavelet modulation unit, the output of the wavelet coefficient processor network is spliced ​​with a low-level visual feature map output by a shallow illumination estimator unit in a hierarchical illumination estimator module, and the result is used as an input of an illumination modulation network, the output of the illumination modulation network is used as an input of a regional enhancement unit, and an inverse wavelet transform is performed on the output of the regional enhancement unit to obtain a wavelet modulation feature map as an output of the illumination-guided wavelet modulation unit;

[0028] In the illumination-guided feature modulation unit, the high-level semantic feature map output by the deep illumination estimator unit in the hierarchical illumination estimator module is used as the input of the three feature extraction units in parallel in the illumination-guided feature modulation unit. The output of the three feature extraction units in parallel is used as the input of the feature fusion layer after passing through the connection layer. Then, the wavelet modulation feature map output by the illumination-guided wavelet modulation unit is used as the input of the fusion attention unit. The output of the fusion attention unit is multiplied with the wavelet modulation feature map and used as the input of the spatial attention unit. The output of the spatial attention unit is multiplied with the output of the fusion attention unit and then multiplied with the output of the feature fusion layer. The output is added to the input wavelet modulation feature map to obtain the feature modulation feature map output by the illumination-guided feature modulation unit.

[0029] As a preferred solution, the wavelet coefficient processor network in the illumination-guided wavelet modulation unit includes a cascaded normalization unit and a convolution activation unit, the normalization unit includes a convolution layer, an instance normalization layer and an RReLU activation function layer connected in sequence, and the convolution activation unit includes a convolution layer and a ReLU activation function layer connected in sequence; wherein, the input feature is used as the input of the normalization unit, the output of the normalization unit is used as the input of the first convolution activation unit, the output of the first convolution activation unit is used as the input of the second convolution activation unit, the output of the second convolution activation unit is added to the output of the normalization unit and then passed through a Sigmoid activation function to obtain the output of the wavelet coefficient processor network;

[0030] The illumination modulation network includes a cascaded normalization unit and a convolution layer; wherein the output of the wave coefficient processor network is input to the illumination modulation network and the low-level visual feature map output by the shallow illumination estimator unit for scaling and concatenation, and the output of the normalization unit is used as the input of the convolution layer, and the output of the convolution layer is subjected to a Sigmoid activation function to obtain the output of the illumination modulation network;

[0031] The region enhancement unit includes a normalization unit, and the normalization unit includes a convolution layer, an instance normalization layer and an RReLU activation function layer connected in sequence.

[0032] As a preferred solution, the processing process of the light-guided wavelet modulation unit is:

[0033]

[0034]

[0035] In the formula, represents the low frequency subband, represents the high frequency subband of the jth layer, L represents the number of decomposition layers, represents discrete wavelet transform, F represents the basic feature map obtained by extracting features from the input low-light image, and Represents low-frequency and high-frequency processors, respectively. and Respectively represent the output after being processed by the low-frequency and high-frequency processors, and represents the illumination modulation network, σ(·) represents the Sigmoid activation function, ⊙ represents element-by-element multiplication, and Y' l and Respectively indicate that and The output after the light modulation network is processed, and They represent the enhancers for different illumination areas, μ i Indicates the degree of membership of the corresponding region, Z l and Respectively indicate that and The output after processing by the region enhancement unit, F w represents the output of the illumination-guided wavelet modulation unit, Represents inverse discrete wavelet transform.

[0036] As a preferred solution, the processing process of the light guidance feature modulation unit is:

[0037]

[0038] F m =F w ⊙W c ⊙W s ⊙F light +F w ;

[0039] In the formula, represents the feature extraction branch with different convolution kernel sizes, k represents the convolution kernel size, and F k Represents the output of three feature extraction units connected in parallel, represents the feature fusion layer, F light represents the output after the feature fusion layer, represents the global average pooling operation, represents a two-layer perceptron, W c represents the output of the channel attention unit, W s represents the output of the spatial attention unit, and Respectively represent the average and maximum pooling operations in the channel dimension, W conv represents the convolutional layer parameters, F m Represents the output of the illumination-guided feature modulation unit.

[0040] As a preferred solution, in step S203, the adaptive enhancement module includes a feature fusion unit and an image level enhancement unit;

[0041] The feature fusion unit includes a first depth convolution unit, two cascaded convolution activation units, a second depth convolution unit, a convolution layer, a Sigmoid activation function layer and an attention unit connected in sequence, wherein the first depth convolution unit includes two cascaded depth convolution layers, the convolution activation unit includes a cascaded convolution layer and a ReLU activation function layer, and the second depth convolution unit is one depth convolution layer;

[0042] Among them, after splicing the wavelet modulation feature map output by the illumination-guided wavelet modulation unit in the dual illumination feature modulation module and the feature modulation feature map of the illumination-guided feature modulation unit, they are respectively used as the input of the first deep convolution unit and the third convolution layer in the feature fusion unit in the adaptive enhancement module, the output of the first deep convolution unit is used as the input of the two cascaded convolution activation units, the output of the two cascaded convolution activation units and the output of the first deep convolution unit are further added, and then used as the input of the third deep convolution layer, the output of the third deep convolution layer and the output of the third convolution layer are multiplied by the output of the attention unit after passing through the Sigmoid activation function layer, and the output is multiplied by the basic feature map obtained by feature extraction of the differential attention and the input low-light image, and then added to obtain the output of the feature fusion unit;

[0043] The image-level enhancement unit includes a first convolution layer, a plurality of residual blocks and a second convolution layer connected in sequence, the residual block includes three cascaded nonlinear convolution activation units, a convolution layer, a maximum pooling layer and a Sigmoid activation function layer, and the nonlinear convolution activation unit includes a cascaded convolution layer and an LReLU activation function layer; wherein, in the residual block, the input feature map is respectively used as the input of the convolution layer and the three cascaded nonlinear convolution activation units, the output of the three cascaded nonlinear convolution activation units and the output of the convolution layer are multiplied after passing through the maximum pooling layer and the Sigmoid activation function layer, and the output result is added to the input of the residual block to obtain the output of the residual block;

[0044] Among them, the output of the feature fusion unit is used as the input of the first convolutional layer in the image-level enhancement unit, the output of the first convolutional layer is used as the input of the first residual block, the output of each residual block is used as the input of the next layer in the image-level enhancement unit, the output of the last residual block is used as the input of the second convolutional layer, and the output of the second convolutional layer is used as the output of the feature fusion unit.

[0045] As a preferred solution, the processing process of the feature fusion unit is:

[0046] A=σ(C([F w ,F m ]));

[0047] F fused =D(F′ m );

[0048] I fusion =A⊙F fused +(1-A)⊙F;

[0049] In the formula, represents a normal convolutional layer, A represents the attention unit matrix, represents a network layer composed of a series of depth-separable convolutions, ordinary convolutions, and residual connections, F fused Indicates passing Processed output, F' m represents the output after concatenating the wavelet modulation feature map and the feature modulation feature map, (1-A) represents the complementary attention unit matrix;

[0050] The processing process of the image level enhancement unit is:

[0051]

[0052] In the formula, F fusion For I fusion The features reprojected into the feature space, F res and Both represent the output of the i-th residual block, The last convolutional layer is 1×1 and maps the features back to the RGB image space. N is the number of residual blocks, and I normal represents the illumination enhancement result image after being enhanced by the image level enhancement unit output. As a preferred solution, in step S204, the pixel level reconstruction loss function is:

[0053]

[0054] In the formula, represents the pixel-level reconstruction loss function, ∥·∥1 represents the L1 norm, and I gt represents the target image;

[0055] The structural similarity loss function is:

[0056]

[0057] In the formula, represents the structural similarity loss function, SSIM(·) represents the indicator for measuring the visual similarity between two images, μ x and μ y Both represent the local mean, σ x and σ y Both represent standard deviation, σ xy represents covariance, C1 and C2 both represent stability constants;

[0058] The perceptual loss function is:

[0059]

[0060] In the formula, is the perceptual loss function, φ l (·) represents the feature extractor of the lth layer of the VGG network, w lrepresents the learned channel weight, H l and W l Represent the height and width of the feature map respectively;

[0061] Then the multi-objective loss function is:

[0062]

[0063] In the formula, is a multi-objective loss function, and λ1, λ2 and λ3 are weight coefficients.

[0064] Compared with the prior art, the present invention has the following technical effects:

[0065] (1) The present invention proposes a hierarchical illumination estimator module for image enhancement, which realizes accurate description of local illumination distribution by defining a pre-designed mapping function. This multi-granular modeling method not only improves the robustness of illumination estimation, but also has more semantic properties than pure numerical division, reduces the semantic difference between the model and the human eye senses, and innovatively decomposes illumination information into shallow and deep layers for processing. The shallow illumination estimator models the local illumination prior through a multi-granular illumination mapping function, realizes accurate modeling of image illumination distribution, effectively captures the illumination change characteristics of different regions that conform to natural language expression, and provides accurate guiding information for feature enhancement. At the same time, the spatial consistency of illumination estimation is enhanced through Gaussian filtering preprocessing; the deep illumination estimator extracts illumination features with semantic perception capabilities through a deep neural network, ensuring the consistency of illumination features with image semantics.

[0066] (2) The present invention designs a synergistic dual illumination feature modulation module including an illumination-guided wavelet modulation unit and an illumination-guided feature modulation unit, which realizes a progressive quality improvement from the wavelet domain to the feature domain. The illumination-guided wavelet modulation unit modulates the image features in the wavelet domain using shallow illumination features in different regions, and achieves precise adjustment of local details by decomposing different frequency components; the illumination-guided feature modulation unit introduces deep illumination priors in the feature domain to ensure that the enhancement process is consistent with the scene illumination conditions. In this way, while maintaining the image structure information, image enhancement is achieved through adaptive modulation of illumination features, making full use of illumination information and suppressing noise amplification.

[0067] (3) The present invention also designs an adaptive enhancement module, which uses an attention mechanism to dynamically weight features at different levels, and realizes adaptive fusion from low-level texture to high-level semantics. Through residual learning and multi-scale feature aggregation, the adaptive enhancement module effectively improves the model's ability to handle complex scenes. At the same time, the present invention uses a multi-objective loss function to achieve balanced optimization of visual quality and numerical indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0069] Figure 1 This is a flow chart of a low-light image enhancement method based on layered illumination guidance according to the present invention;

[0070] Figure 2 Schematic diagram of the framework of the low-light image enhancement model in the method of the present invention;

[0071] Figure 3 is a schematic diagram of a layered illumination estimator module in the method of the present invention;

[0072] Figure 4 Schematic diagram of a light-guided wavelet modulation unit in the method of the present invention;

[0073] Figure 5 This is a schematic diagram of the subnetwork structure of the method of the present invention;

[0074] Figure 6 Schematic diagram of a light-guided feature modulation unit in the method of the present invention;

[0075] Figure 7 This is a qualitative comparison diagram of different methods of the embodiments of the present invention and the method proposed in this embodiment on the LOLv1 dataset

[0076] Figure 8 This is a qualitative comparison chart of different methods of the embodiments of the present invention and the method proposed in this embodiment on the LOLv2-Real dataset;

[0077] Fig. 9 This is a comparison chart of the visual effects of different methods in the embodiments of the present invention on the LOLv2-Syn dataset;

[0078] Fig.10 It is a qualitative comparison diagram of different methods of the embodiments of the present invention on multiple reference-free data sets;

[0079] Fig.11 This is a qualitative result diagram of the ablation experiment of the main components of the embodiment of the present invention on the LOLv2-Real dataset;

[0080] Fig.12 This is an analysis diagram of ablation experiments on the LOLv2-Real dataset of different components of an embodiment of the present invention;

[0081] Fig.13 This is an analysis diagram of the impact of the loss function on image quality according to an embodiment of the present invention. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0083] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0084] Early low-light image enhancement mainly relied on traditional image processing techniques. Histogram adjustment-based methods, such as histogram equalization (HE) and its variant CLAHE, attempt to improve image contrast by redistributing pixel grayscale values. Such methods are computationally simple and easy to implement, but often lead to over-enhancement and unnatural visual effects. Retinex-based methods decompose images into illumination and reflection components, and achieve enhancement by estimating and adjusting the illumination component. However, accurately estimating the illumination component remains challenging in complex scenes. In addition, although the variational model-based methods perform well in maintaining image structure, their iterative optimization process is computationally complex and difficult to meet the needs of real-time processing. Although existing methods have made significant progress in the field of low-light image enhancement, there are still the following key problems: 1) There is a lack of accurate modeling of local illumination. Existing methods often use simple threshold division or global estimation strategies, which are difficult to accurately capture the characteristics of illumination changes in complex scenes; 2) In the process of feature reconstruction, the illumination prior information is not fully utilized, resulting in the enhancement results not matching the actual scene illumination conditions; especially when processing low-light images of real scenes, these problems are more prominent.

[0085] In view of the above problems and shortcomings, the present invention improves the enhancement effect by using multi-level and multi-granular illumination estimation to guide the adjustment of different domain features. Figure 1 As shown, the present invention proposes a low-light image enhancement method based on layered illumination guidance, the method comprising:

[0086] S1, obtaining a low-light image to be processed;

[0087] S2, using the low-light image to be processed as an input of a trained low-light image enhancement model, and outputting an image enhancement result of the low-light image to be processed; the low-light image enhancement model includes a hierarchical illumination estimator module, a dual illumination feature modulation module and an adaptive enhancement module;

[0088] The training process of the low-light image enhancement model is as follows:

[0089] S201, the hierarchical illumination estimator module uses the low-light image as a training sample input to the low-light image enhancement model as the input of the shallow illumination estimator unit in the hierarchical illumination estimator module, performs shallow feature extraction through a multi-granular illumination mapping function to obtain a low-level visual feature map, and uses the low-light image as the input of the deep illumination estimator unit to output a high-level semantic feature map;

[0090] S202, the dual illumination feature modulation module modulates the low-level visual feature map and the high-level semantic feature map through an illumination-guided wavelet modulation unit and an illumination-guided feature modulation unit, respectively, to obtain a wavelet modulation feature map and a feature modulation feature map, respectively;

[0091] S203, the adaptive enhancement module splices the wavelet modulation feature map and the feature modulation feature map, performs feature adaptive fusion and image-level enhancement processing on the spliced ​​features, and finally outputs an enhanced illumination enhancement result image;

[0092] S204, using a multi-objective loss function consisting of a pixel-level reconstruction loss function, a structural similarity loss function, and a perceptual loss function to calculate the training loss of the low-light image enhancement model, and updating the parameters of the low-light image enhancement model with the goal of minimizing the multi-objective loss function;

[0093] S205. Repeat steps S201 to S204 until the low-light image enhancement model converges or reaches a preset number of training times, and the training is completed.

[0094] The core of the present invention is to design a plurality of key modules that work together, including: Hierarchical Illumination Estimator (HIE), dual illumination feature modulation module including illumination-guided wavelet modulation module (IGWM) and illumination-guided feature modulation module (IGFM), and adaptive enhancement module (AEM). Among them, HIE is inspired by the application of fuzzy set theory to construct an illumination prior model, and realizes the accurate description of local illumination distribution by defining a pre-designed mapping function. This multi-granular modeling method not only improves the robustness of illumination estimation, but also has more semantic properties than pure numerical division, reducing the semantic difference between the model and the human eye sense organ. The two modulation modules realize a multi-level feature reconstruction strategy in the image enhancement process. The IGWM module modulates the image features in the wavelet domain using shallow illumination features in different regions, and realizes the accurate adjustment of local details by decomposing different frequency components, which is crucial to maintain the naturalness of the image. The IGFM module introduces deep illumination prior in the feature domain to ensure that the enhancement process is consistent with the scene illumination conditions. This dual modulation mechanism establishes a progressive enhancement framework from the wavelet domain to the feature domain, which not only provides a larger optimization space, but also achieves coordinated enhancement of local details and global structures. AEM uses an attention mechanism to dynamically weight features at different levels, achieving adaptive fusion from low-level textures to high-level semantics. Through residual learning and multi-scale feature aggregation, AEM effectively improves the model's ability to handle complex scenes. At the same time, this embodiment uses a multi-objective loss function to achieve balanced optimization of visual quality and numerical indicators.

[0095] In addition, in the process of training the model of the present invention, the low-light images used as training samples as input to the low-light image model can be low-light images with known defects, wherein the images with known defects include images with low brightness, high noise, low contrast, color distortion, and detail loss.

[0096] The low-light image enhancement method based on layered illumination guidance of the present invention is described in more detail below.

[0097] 1. Low-light image enhancement model

[0098] like Figure 2As shown, the core of the present invention is to design multiple key modules that work together, including: a hierarchical illumination estimator (HIE), a dual illumination feature modulation module including an illumination-guided wavelet modulation unit (IGWM) and an illumination-guided feature modulation unit (IGFM), and an adaptive enhancement module (AEM). Among them, HIE is inspired by the application of fuzzy set theory to construct an illumination prior model, and realizes an accurate description of the local illumination distribution by defining a pre-designed mapping function. This multi-granular modeling method not only improves the robustness of illumination estimation, but also has more semantic properties than pure numerical division, reducing the semantic difference between the model and the human eye senses. The two modulation units implement a multi-level feature reconstruction strategy in the image enhancement process. The IGWM unit modulates the image features in the wavelet domain using shallow illumination features in different regions, and realizes accurate adjustment of local details by decomposing different frequency components, which is crucial to maintaining the naturalness of the image. The IGFM unit introduces deep illumination prior in the feature domain to ensure that the enhancement process is consistent with the scene illumination conditions. This dual modulation mechanism establishes a progressive enhancement framework from the wavelet domain to the feature domain, which not only provides a larger optimization space, but also achieves coordinated enhancement of local details and global structures. AEM uses an attention mechanism to dynamically weight features at different levels, achieving adaptive fusion from low-level textures to high-level semantics. Through residual learning and multi-scale feature aggregation, AEM effectively improves the model's ability to handle complex scenes. At the same time, this embodiment uses a multi-objective loss function to achieve balanced optimization of visual quality and numerical indicators.

[0099] They are introduced in detail below.

[0100] 1.1. Hierarchical Lighting Estimator Module

[0101] The accuracy of illumination estimation directly affects the quality of image enhancement. Traditional global illumination estimation methods estimate global illumination features. This processing method makes it difficult to accurately enhance the local illumination characteristics of an image, and often produces artifacts or over-enhancement. To overcome this limitation, the present invention proposes a hierarchical illumination estimator (HIE), which realizes illumination feature extraction from pixel level to semantic level through the synergy of shallow and deep branches.

[0102] Specifically, if Figure 3As shown, the hierarchical illumination estimator module includes a shallow illumination estimator unit and a deep illumination estimator unit; wherein the shallow illumination estimator unit includes a cascaded color space transformation module, a convolution layer, a Gaussian filter layer, a multi-granularity illumination mapping function module and a normalization layer, and the deep illumination estimator unit includes a cascaded first convolution layer, an instance normalization layer, a ReLU activation function layer and a second convolution layer;

[0103] In the hierarchical illumination estimator module, the low-light image used as the input of the hierarchical illumination estimator module is used as the input of the shallow illumination estimator unit and the deep illumination estimator unit respectively, and the output of the shallow illumination estimator unit obtains a low-level visual feature map. The output of the shallow illumination estimator unit and the low-light image used as the input are both used as the input of the deep illumination estimator unit, and the output of the deep illumination estimator obtains a high-level semantic feature map.

[0104] The shallow illumination estimator unit and the deep illumination estimator unit are introduced below respectively.

[0105] 1.1.1 Shallow Light Estimator Unit

[0106] In this embodiment, for the shallow illumination estimator unit, the processing process is as follows: first, the input image is transformed into a color space and the brightness is extracted. In order to improve the stability of the estimation, a Gaussian filter is introduced for local smoothing:

[0107]

[0108] In the formula, Represents the brightness map extracted by Gaussian filtering introduced by the shallow illumination estimator unit, Indicates the transformation operation from RGB space to HSV space and intercepts the last channel V, G σ represents a Gaussian kernel with standard deviation σ, represents the convolution operation, I low represents the input low-light image;

[0109] Normalize the filtered brightness image:

[0110]

[0111] Where, L norm Indicates the normalization of the brightness image after Gaussian filtering, L min and L max They represent the minimum and maximum values ​​of the brightness map, ∈ represents the numerical stability constant, and γ represents the normalization coefficient (usually 255);

[0112] On this basis, three complementary mapping functions are defined:

[0113]

[0114] In the formula, μ d , μ n , μ b They represent the dark area, normal area, and bright area of ​​the illumination distribution map, respectively. (a, b) represent the characteristic threshold of the illumination interval. δ and β represent the steepness of the Sigmoid function.

[0115] To ensure the complementarity of illumination, a normalization constraint is introduced to obtain the final illumination map:

[0116]

[0117] The calculation of each component is:

[0118]

[0119] In the formula, represents a three-channel light map for the shallow estimator, Represents each component, d, n, b represent dark, normal, bright respectively;

[0120] Finally, the shallow estimator unit is expressed as:

[0121] I μ =ε shallow (I low )=μ(L norm );

[0122] In the formula, I μ Represents ε shallow The output low-level visual feature map, ε shallow (·) denotes a shallow lighting estimator unit.

[0123] 1.1.2. Deep Lighting Estimator Unit

[0124] In this embodiment, the processing process for the deep illumination estimator unit is specifically as follows:

[0125]

[0126] In the formula, F μ Represents ε deep The output high-level semantic feature map, ε deep (·) denotes a high-level illumination estimator unit, [·,·] denotes feature concatenation in the channel dimension, ReLU denotes the ReLU activation function, IN denotes instance normalization, and They represent the weights of the two convolutional layers respectively, and k represents the convolution kernel size.

[0127] This embodiment innovatively decomposes illumination information into two levels, shallow and deep, for processing through the design of a layered architecture. The shallow illumination estimator models local illumination priors through a multi-granularity illumination mapping function, thereby achieving accurate modeling of image illumination distribution, effectively capturing illumination variation characteristics of different regions that are consistent with natural language expression, and providing accurate guidance information for feature enhancement. At the same time, the spatial consistency of illumination estimation is enhanced through Gaussian filtering preprocessing. The deep illumination estimator extracts illumination features with semantic perception capabilities through a deep neural network, thereby ensuring the consistency of illumination features with image semantics.

[0128] 1.2 Dual illumination feature modulation module

[0129] After obtaining the illumination estimation, this embodiment proposes a dual illumination feature modulation module, which achieves progressive enhancement from low-level visual features to high-level semantic features by modulating image features at two levels: frequency domain and feature domain.

[0130] Specifically, the module includes an illumination-guided wavelet modulation unit and an illumination-guided feature modulation unit; wherein, the illumination-guided wavelet modulation unit includes a wavelet coefficient processor network, an illumination modulation network and a regional enhancement unit cascaded in sequence, and the illumination-guided feature modulation unit includes three feature extraction units, a connection layer, a feature fusion layer, a channel attention unit and a spatial attention unit in parallel, and the cascade results of the three parallel feature extraction units, the connection layer and the feature fusion layer are connected in parallel with the cascade results of the attention unit and the spatial attention unit, wherein the three parallel feature extraction units are respectively a feature extraction unit with a convolution kernel of 1, a feature extraction unit with a convolution kernel of 2 and a feature extraction unit with a convolution kernel of 3.

[0131] The illumination-guided wavelet modulation unit and the illumination-guided feature modulation unit are introduced below respectively.

[0132] 1.2.1. Light-guided wavelet modulation unit

[0133] like Figure 4 As shown, in the illumination-guided wavelet modulation unit, a low-light image as an input of a low-light image enhancement model is subjected to feature extraction to obtain a basic feature map, the basic feature map is used as the input of a wavelet coefficient processor network in the illumination-guided wavelet modulation unit, the output of the wavelet coefficient processor network is spliced ​​with a low-level visual feature map output by a shallow illumination estimator unit in a hierarchical illumination estimator module, and the result is used as the input of the illumination modulation network, the output of the illumination modulation network is used as the input of the regional enhancement unit, and the output of the regional enhancement unit is subjected to an inverse wavelet transform to obtain a wavelet modulation feature map as the output of the illumination-guided wavelet modulation unit;

[0134] like Figure 5 As shown, the wavelet coefficient processor network in the illumination-guided wavelet modulation unit includes a cascaded normalization unit and a convolution activation unit, the normalization unit includes a convolution layer, an instance normalization layer and an RReLU activation function layer connected in sequence, and the convolution activation unit includes a convolution layer and a ReLU activation function layer connected in sequence; wherein, the input feature is used as the input of the normalization unit, the output of the normalization unit is used as the input of the first convolution activation unit, the output of the first convolution activation unit is used as the input of the second convolution activation unit, the output of the second convolution activation unit is added to the output of the normalization unit and then passed through the Sigmoid activation function to obtain the output of the wavelet coefficient processor network;

[0135] The illumination modulation network includes a cascaded normalization unit and a convolution layer; wherein the output of the wave coefficient processor network is input to the illumination modulation network and the low-level visual feature map output by the shallow illumination estimator unit for scaling and concatenation, and the output of the normalization unit is used as the input of the convolution layer, and the output of the convolution layer is subjected to a Sigmoid activation function to obtain the output of the illumination modulation network;

[0136] The region enhancement unit includes a normalization unit, and the normalization unit includes a convolution layer, an instance normalization layer and an RReLU activation function layer connected in sequence.

[0137] For the illumination guided wavelet modulation unit, the processing process is: given the basic feature map Through discrete wavelet transform, we can get:

[0138]

[0139] In the formula, represents the low frequency subband, represents the high frequency subband of the jth layer, L represents the number of decomposition layers, represents discrete wavelet transform, and F represents the basic feature map obtained by extracting features from the input low-light image;

[0140] First, each subband coefficient is processed preliminarily, such as Figure 5 (a) as shown:

[0141]

[0142] In the formula, and Represents low-frequency and high-frequency processors, respectively. and Respectively represent the output after being processed by the low-frequency and high-frequency processors;

[0143] Next, introduce the shallow illumination estimation result Iμ The coefficients are modulated and the initial wavelet feature transform is realized using illumination as the guide information, such as Figure 5 (b) as shown:

[0144]

[0145] In the formula, and represents the illumination modulation network, σ(·) represents the Sigmoid activation function, ⊙ represents element-by-element multiplication, and Y' l and Respectively indicate that and The output after the light modulation network processing;

[0146] In order to achieve regional adaptive enhancement and learn different enhancement strategies for different illumination areas, this module also introduces a regional enhancement mechanism of multi-granular illumination mapping. For each sub-band, different areas divided by illumination mapping are enhanced separately, such as Figure 5 (c) as shown:

[0147]

[0148] In the formula, and They represent the enhancers for different illumination areas, μ i Indicates the degree of membership of the corresponding region, Z l and Respectively indicate that and The output after processing by the regional enhancement unit;

[0149] Finally, the features are reconstructed by inverse wavelet transform:

[0150]

[0151] In the formula, F w represents the output of the illumination-guided wavelet modulation unit, Represents inverse discrete wavelet transform.

[0152] 1.2.2 Light-guided feature modulation unit

[0153] Compared with the operation of IGWM in the wavelet domain, the IGFM unit directly modulates in the feature domain, using the deep illumination feature F μ To guide the feature enhancement process, e.g. Figure 6As shown, in the illumination-guided feature modulation unit, the high-level semantic feature map output by the deep illumination estimator unit in the hierarchical illumination estimator module is used as the input of the three feature extraction units in parallel in the illumination-guided feature modulation unit, and the output of the three feature extraction units in parallel is used as the input of the feature fusion layer after passing through the connection layer. Then, the wavelet modulation feature map output by the illumination-guided wavelet modulation unit is used as the input of the fusion attention unit, and the output of the fusion attention unit is multiplied with the wavelet modulation feature map and used as the input of the spatial attention unit. The output of the spatial attention unit is multiplied with the output of the fusion attention unit and then multiplied with the output of the feature fusion layer. The output of the spatial attention unit is added to the input wavelet modulation feature map to obtain the feature modulation feature map output by the illumination-guided feature modulation unit.

[0154] For the illumination-guided feature modulation unit, the processing process is:

[0155] First, multiple parallel branches are used to extract illumination features of different receptive fields:

[0156]

[0157] In the formula, represents the feature extraction branch with different convolution kernel sizes, k represents the convolution kernel size, and F k Represents the output of three feature extraction units connected in parallel;

[0158] These features are then fused to form a comprehensive illumination modulation feature:

[0159]

[0160] In the formula, represents the feature fusion layer, F light Represents the output after the feature fusion layer;

[0161] In order to achieve more accurate feature modulation, IGFM introduces channel attention and spatial attention mechanisms, and its channel attention weight is calculated as follows:

[0162]

[0163] In the formula, represents the global average pooling operation, represents a two-layer perceptron, W c Represents the output of the channel attention unit;

[0164] The calculation of spatial attention weights is based on the statistical characteristics of the feature map:

[0165]

[0166] Where W s represents the output of the spatial attention unit, and Respectively represent the average and maximum pooling operations in the channel dimension, W conv Represents the convolution layer parameters;

[0167] The final feature modulation process can be expressed as:

[0168] F m =F w ⊙W c ⊙W s ⊙F light +F w ;

[0169] In the formula, F m Represents the output of the illumination-guided feature modulation unit.

[0170] 1.3 Adaptive Enhancement Module

[0171] As the last processing unit of the network, the adaptive enhancement module (AEM) mainly completes two tasks: feature adaptive fusion and image level enhancement like Figure 2 As shown, the module first receives the feature F from IGWM w and the characteristic F of IGFM m , and the feature F from the shallow layer, the final output is the enhanced illumination enhancement result image I normal ,in and Structure Figure 5 (d) and Figure 5 (e) as shown.

[0172] Specifically, the adaptive enhancement module includes a feature fusion unit and an image-level enhancement unit;

[0173] The feature fusion unit includes a first depth convolution unit, two cascaded convolution activation units, a second depth convolution unit, a convolution layer, a Sigmoid activation function layer and an attention unit connected in sequence, wherein the first depth convolution unit includes two cascaded depth convolution layers, the convolution activation unit includes a cascaded convolution layer and a ReLU activation function layer, and the second depth convolution unit is one depth convolution layer;

[0174] Among them, after splicing the wavelet modulation feature map output by the illumination-guided wavelet modulation unit in the dual illumination feature modulation module and the feature modulation feature map of the illumination-guided feature modulation unit, they are respectively used as the input of the first deep convolution unit and the third convolution layer in the feature fusion unit in the adaptive enhancement module, the output of the first deep convolution unit is used as the input of the two cascaded convolution activation units, the output of the two cascaded convolution activation units and the output of the first deep convolution unit are further added, and then used as the input of the third deep convolution layer, the output of the third deep convolution layer and the output of the third convolution layer are multiplied by the output of the attention unit after passing through the Sigmoid activation function layer, and the output is multiplied by the basic feature map obtained by feature extraction of the differential attention and the input low-light image, and then added to obtain the output of the feature fusion unit;

[0175] The image-level enhancement unit includes a first convolution layer, a plurality of residual blocks and a second convolution layer connected in sequence, the residual block includes three cascaded nonlinear convolution activation units, a convolution layer, a maximum pooling layer and a Sigmoid activation function layer, and the nonlinear convolution activation unit includes a cascaded convolution layer and an LReLU activation function layer; wherein, in the residual block, the input feature map is respectively used as the input of the convolution layer and the three cascaded nonlinear convolution activation units, the output of the three cascaded nonlinear convolution activation units and the output of the convolution layer are multiplied after passing through the maximum pooling layer and the Sigmoid activation function layer, and the output result is added to the input of the residual block to obtain the output of the residual block;

[0176] Among them, the output of the feature fusion unit is used as the input of the first convolutional layer in the image-level enhancement unit, the output of the first convolutional layer is used as the input of the first residual block, the output of each residual block is used as the input of the next layer in the image-level enhancement unit, the output of the last residual block is used as the input of the second convolutional layer, and the output of the second convolutional layer is used as the output of the feature fusion unit.

[0177] For the feature fusion unit, the processing process is as follows:

[0178]

[0179] I fusion =A⊙F fused +(1-A)⊙F;

[0180] In the formula, represents a normal convolutional layer, A represents the attention unit matrix, represents a network layer composed of a series of depth-separable convolutions, ordinary convolutions, and residual connections, F fused Indicates passing Processed output, F' mrepresents the output after concatenating the wavelet modulation feature map and the feature modulation feature map, (1-A) represents the complementary attention unit matrix;

[0181] For the image-level enhancement unit, a cascade structure of multiple residual blocks is used to further process the fusion features. The processing process is as follows:

[0182]

[0183] In the formula, F fusion For I fusion The features reprojected into the feature space, F res and Both represent the output of the i-th residual block, The last convolutional layer is 1×1 and maps the features back to the RGB image space. N is the number of residual blocks, and I normal Represents the illumination enhancement result image after being enhanced through the output of the image-level enhancement unit.

[0184] The AEM module generates an intermediate fusion image I fusion And the final enhanced result I normal , showing the progressive quality improvement process from feature fusion to image enhancement. The entire module uses depthwise separable convolution instead of ordinary convolution to reduce the amount of computation. At the same time, the use of attention mechanism and residual structure also improves the model's adaptability to different scenarios.

[0185] 1.4 Loss Function

[0186] In order to achieve balanced optimization of visual quality and numerical indicators, this embodiment adopts a combination of multi-objective loss functions. Specifically, the multi-objective loss function consists of three parts: pixel-level reconstruction loss function, structural similarity loss function and perceptual loss function. Among them, the pixel-level reconstruction loss function uses the L1 norm to measure the pixel-level difference between the prediction result and the target image:

[0187]

[0188] In the formula, represents the pixel-level reconstruction loss function, ∥·∥1 represents the L1 norm, and I gt represents the target image;

[0189] The structural similarity loss function is based on the SSIM indicator and is used to maintain the structural information of the image. The SSIM loss can effectively maintain the local structural features of the image, which is particularly important for low-light enhancement tasks:

[0190]

[0191] In the formula, represents the structural similarity loss function, SSIM(·) represents the indicator for measuring the visual similarity between two images, μ x and μ y Both represent the local mean, σ x and σ y Both represent standard deviation, σ xy represents covariance, C1 and C2 both represent stability constants;

[0192] The perceptual loss function uses the LPIPS metric to evaluate the perceptual similarity of images through the pre-trained deep network feature space:

[0193]

[0194] In the formula, is the perceptual loss function, φ l (·) represents the feature extractor of the lth layer of the VGG network, w l represents the learned channel weight, H l and W l Represent the height and width of the feature map respectively;

[0195] Then the multi-objective loss function is:

[0196]

[0197] In the formula, is a multi-objective loss function, and λ1, λ2 and λ3 are weight coefficients.

[0198] The design of this multi-objective loss function fully considers multiple aspects of low-light image enhancement: L1 loss ensures basic pixel-level accuracy, SSIM loss maintains the integrity of the image structure, and LPIPS loss guides the model to generate a natural enhancement effect that conforms to human visual perception. Experiments show that this combination of loss functions can effectively balance detail preservation and visual quality, producing more satisfactory enhancement results. In the specific application embodiment, a reasonable trade-off of each loss term is achieved by setting λ1=1.0, λ2=0.5, and λ3=0.01.

[0199] 2. Example

[0200] 2.1 Experimental Setup

[0201] 2.1.1 Datasets and Indicators

[0202] This embodiment uses multiple widely used datasets for experimental evaluation. In the training and evaluation stages, the LOL and LOLv2 datasets are mainly used. Among them, the LOL dataset contains 485 pairs of low-light-normal-light image pairs, which are obtained through actual indoor environment collection, and the image resolution is 400×600. As an extended version of LOL, the LOLv2 dataset consists of two subsets: the Real World subset (LOLv2-Real) contains 689 pairs of real scene image pairs, including 589 pairs of training sets and 100 pairs of test sets, covering samples collected at different times and scenes; the Synthetic subset (LOLv2-Syn) contains 1000 pairs of synthetic image pairs, including 900 pairs of training sets and 100 pairs of test sets, which are generated by light attenuation and noise addition to normal light images. This combination of real and synthetic data provides the model with richer and more diverse training samples. In order to comprehensively evaluate the generalization performance of the model, this embodiment is also tested on multiple reference-free datasets, including MEF, NPE, LIME, DICM, and VV datasets. These datasets contain 17, 8, 10, 69, and 24 low-light images, respectively, covering different scene types and lighting conditions. In particular, the MEF dataset focuses on scenes with large-scale lighting changes, the LIME dataset highlights the characteristics of uneven lighting, and the VV dataset contains images under extreme lighting conditions. This diverse test set design ensures a comprehensive evaluation of model performance.

[0203] In order to fully verify the effectiveness of the proposed method, two types of image quality metrics were used for evaluation: with reference metrics (PSNR, SSIM and LPIPS) and without reference metrics (BRISQUE, NIQE and PIQE). Such a dataset configuration design achieves effective supervised learning and testing through paired datasets, and uses a diverse test set to ensure reliable evaluation of the model's generalization performance. This reliable evaluation method provides a solid foundation for subsequent model analysis.

[0204] 2.1.2 Implementation Details

[0205] All experiments in this embodiment are performed on a workstation equipped with an NVIDIA RTX 4090 GPU. The low-light image enhancement model is implemented using the PyTorch framework and trained using the Adam optimizer, with the initial learning rate set to 0.0002. In order to achieve a stable training process, the cosine annealing learning strategy is adopted, and the minimum learning rate is set to 1e-6. In terms of data preprocessing, the images of the LOL and LOLv2-Real datasets are randomly cropped to 256×256 size. All datasets are trained for 1000 rounds to ensure full convergence. Common data enhancement strategies are used during training, including random horizontal flipping and random rotation. In order to prevent overfitting, L2 regularization with a weight decay coefficient of 1e-4 is used. In the parameter selection of the illumination mapping function, a combination of a=60, b=160, δ=β=25 is used.

[0206] 2.2. Comparison with existing technologies

[0207] This embodiment is comprehensively compared with a variety of the latest low-light image enhancement methods, including methods based on Retinex theory (RetinexNet, KinD, Kind++, RetinexFormer, LightenDiffusion), GAN-based methods (EnlightenGAN, RUAS), Transformer-based methods (Restormer, LLFormer) and other advanced methods. The performance advantages of the method in this embodiment are fully verified through systematic evaluation on reference datasets and non-reference datasets.

[0208] 2.2.1 Performance of the LOL dataset series

[0209] On the LOLv1 and LOLv2-Real datasets, the method of this embodiment shows significant performance advantages. As shown in Table 1, on the LOLv1 dataset, this method achieved the best score of 0.112 on the perceptual quality index LPIPS, which is significantly improved compared to the suboptimal method MIRNetv2's 0.113. Especially in the SSIM index, this method reached a high value of 0.852, second only to LLFlow's 0.854, showing a strong structural retention ability. On the LOLv2-Real dataset, the PSNR and SSIM indicators of this method reached 23.16 and 0.871 respectively, which are significantly improved compared to the latest methods such as IAT and UHDFour. More noteworthy is that while maintaining high performance, the model parameter volume of this method is only 1.53M, which is significantly lower than MIRNet's 31.76M and Restormer's 26.13M, reflecting a higher computational efficiency.

[0210] Table 1 Comparison of quantitative results of different methods on LOLv1 and LOLv2 datasets

[0211]

[0212] In terms of qualitative analysis, Figure 7 , Figure 8 As shown, the method of this embodiment shows unique advantages in processing low-light images of real scenes. Compared with traditional methods such as LIME and KinD, this method can not only effectively improve the overall brightness, but also better preserve image details and color information. Especially when dealing with areas with uneven lighting, thanks to the multi-granularity lighting estimation mechanism, this method can more accurately capture local lighting changes, thereby achieving a more natural enhancement effect. Compared with the latest RetinexFormer and LightenDiffusion, this method also shows obvious advantages in suppressing noise and maintaining edge structure.

[0213] The evaluation on the LOLv2-Syn dataset further verifies the generalization ability of this method. Fig. 9 As shown in the figure, compared with existing methods, this method shows stronger noise suppression ability and better edge preservation effect when processing synthetic low-light images. In terms of quantitative indicators, the SSIM and LPIPS indicators of this method reach the optimal levels of 0.935 and 0.051 respectively, especially in the illumination transition area, as shown in the sky area in the figure, this method can handle illumination gradient more naturally and effectively suppress common artifact problems. These results fully demonstrate the superiority of the dual feature modulation mechanism in balancing detail enhancement and noise suppression.

[0214] 2.2.2 Performance without reference dataset

[0215] In order to further verify the performance of the model in practical application scenarios, this study conducted comprehensive tests on five widely used no-reference datasets (DICM, LIME, MEF, NPE, and VV). As shown in Table 2, on the three no-reference evaluation indicators (BRISQUE, NIQE, and PIQE) of all datasets, this method achieved the optimal level, showing strong scene adaptability and processing capabilities.

[0216] Table 2 Comparison of quantitative results of different methods on no-reference datasets

[0217]

[0218] Specifically, on the DICM dataset, the BRISQUE, NIQE, and PIQE indicators of this method are 20.07, 3.44, and 30.21, respectively, which are significantly better than the second-best method LLFlow. This result shows that this method has better stability when dealing with complex scenes. On the LIME dataset, which is mainly composed of scenes with uneven lighting, the various indicators of this method are also significantly better than the comparison methods, verifying the unique advantages of layered illumination estimation in dealing with non-uniform illumination. Especially on the MEF dataset, this method achieved the best results in all three indicators (BRISQUE: 14.10, NIQE: 3.15, PIQE: 29.63), indicating that it has good adaptability when dealing with large-scale illumination changes.

[0219] In terms of qualitative analysis, Fig.10 As shown in the figure, this method shows excellent enhancement effects when dealing with various complex scenes. Compared with methods such as EnlightenGAN and KinD, this method can better balance the light levels in different areas and effectively avoid over-enhancement and under-enhancement. Especially when dealing with extreme low-light conditions, thanks to the dual feature modulation mechanism, this method can still maintain a stable enhancement effect while effectively suppressing noise and maintaining image details. The excellent performance on the NPE and VV datasets further confirms the generalization ability of this method in different types of scenes.

[0220] These experimental results fully demonstrate that the method proposed in this embodiment has significant advantages in practical applications. In particular, the method has shown superior performance to existing technologies in terms of handling complex lighting conditions, maintaining image details, and suppressing noise. Combined with the low computational complexity of the model, these results show that the method has good application potential in practical application scenarios.

[0221] 2.3 Ablation Experiment

[0222] In order to fully verify the effectiveness of the proposed method, this example first conducts an overall ablation experiment on the main components of the network, and then deeply analyzes the design choices of key modules. All experiments are conducted on the LOLv2-Real

[25] real-world shooting dataset, using the same training strategy and evaluation metrics.

[0223] 2.3.1 Analysis of main components

[0224] First, we conduct an ablation experiment on the overall components to verify the contribution of each key module to the final performance. As shown in Table 3, the experiment starts with the baseline model and observes the performance changes by gradually adding each component. Unused modules are replaced by a single convolutional layer to maintain the consistency of feature dimensions.

[0225] Table 3 Quantitative results analysis of ablation experiments on main components

[0226]

[0227] Experimental results show that the basic model M1 can only achieve limited enhancement effects, with a PSNR of 18.973dB. After adding the HIE module to model M2, the PSNR increased to 19.428dB, verifying the importance of illumination prior information to the enhancement process. The introduction of the IGWM module to model M3 brought significant performance improvements, with the PSNR increased by 2.942dB, and the use of the IGFM module alone in model M4 increased by 1.217dB. This shows that feature modulation in the wavelet domain has a more significant contribution to detail enhancement. When model M5 uses two modulation mechanisms at the same time, the model PSNR performance is further improved to 22.213dB, indicating that the two modulation methods can achieve a good complementary effect. Finally, the complete model M6 with the addition of the AEM module achieved the best performance, while only increasing the number of parameters by about 0.120M, verifying the high efficiency of the module.

[0228] Fig.11 (a) and Fig.11 (b) shows the ablation results for outdoor and indoor environments respectively. After gradually adding components, the component effects are reflected in the performance of regional illumination, detail enhancement, and noise suppression. It is worth noting that the combined effect of IGWM and IGFM is the most significant, while AEM focuses on overall optimization and detail restoration.

[0229] In terms of computational efficiency, although the computational workload of the complete model (40.960G FLOPs) is higher than that of the basic model, considering the performance improvement, PSNR is improved by 4.194dB and SSIM is improved by 0.042, this computational overhead is reasonable. In particular, compared with some existing methods, such as MIRNet's 785.100G FLOPs and Restormer's 144.250G FLOPs, the method in this embodiment still maintains a high computational efficiency. Fig.11 The visual enhancement effect of each component is further demonstrated in Figure 3. It can be observed that with the gradual addition of components, the model has achieved significant improvements in detail preservation and noise suppression.

[0230] 2.3.2 Effectiveness of Layered Illumination Estimation

[0231] After verifying the overall contribution of HIE, this example further analyzes its internal design choices. As shown in Table 4, a detailed comparative analysis of different component combinations and design strategies of HIE is performed. In the experimental design, M represents the multi-granularity illumination mapping strategy, and removing it means using only a single-granularity μ n For global illumination modeling, S and D represent shallow and deep illumination estimators, respectively, and unused network components are replaced by a single convolutional layer.

[0232] Table 4 Ablation quantitative results analysis table of hierarchical illumination estimator (HIE)

[0233]

[0234] The experimental results reveal the specific contributions of each component. When the illumination mapping function is not used, the performance of model M7 is relatively weak, which is mainly due to the inability to accurately capture the local changes in the dark and bright areas of the illumination. After removing the shallow estimation branch of model M9, the PSNR dropped to 22.542dB, while the model M 10 Removing the deep branches alone results in a relatively small performance drop, with a PSNR of 22.693dB, indicating that shallow features are more critical for accurate illumination estimation. When model M8 only retains illumination mapping processing, the model still maintains good performance, with a PSNR of 22.481dB, highlighting the advantages of this multi-granular gradient mapping strategy in dealing with complex illumination scenes. The complete HIE module achieves optimal performance through the synergy of components.

[0235] like Fig.12 As shown in (a), model M9 removes the shallow estimation branch (w / oε shallow ) retains the deep estimation branch and shows overexposure and enhancement. The brightness histogram shows a shift in the overall peak. 10 Add shallow branches and remove deep branches (w / oε deep ) effectively suppresses the shift of illumination distribution, indicating that the shallow estimation branch is more effective in controlling the overall enhancement amplitude of the image. However, due to the lack of illumination semantic information, the over-enhancement of the region is shown. The complete HIE module (w / ε shallow andε deep ) takes into account both the stability of global illumination of the image and the accuracy of local enhancement.

[0236] As mentioned above Figure 3 As shown in Figure 2, the intermediate results of HIE also verify the rationality of the hierarchical design. The shallow estimator generates three complementary illumination maps (μ d , μ n , μ b ). Dark Area Mapping (μ d ) increases with the decrease of local brightness, providing guidance for key areas for subsequent enhancement. Normal illumination area mapping (μ n ) shows a more concentrated response distribution, providing a stable reference for lighting adjustment. b ) is mainly concentrated on light sources and strong reflection areas, effectively preventing over-enhancement.

[0237] In indoor scenes, the advantages of the layered strategy are more prominent. Light mapping achieves a natural transition in areas with drastic changes in lighting. μ The response map reflects a higher level of scene understanding, with a higher degree of attention paid to dark areas. While maintaining the overall illumination continuity, it shows differentiated response patterns for different object regions, with clear hints at the contrast of object boundaries and a clear focus on dark areas. The resulting illumination guidance map successfully integrates low-level illumination features and high-level semantic information, providing a reliable guidance signal for subsequent feature modulation. Fig.12 The ablation results of (b) and (c) show that after gradually adding shallow branches ε shallow , deep branch ε deep Afterwards, the model's understanding of scene lighting was effectively improved, over-enhancement was suppressed, and relatively accurate enhancement results were provided.

[0238] This multi-granular hierarchical illumination estimation not only overcomes the limitations of global estimation in traditional methods, but also enhances the understanding of scene semantics through deep feature extraction. Especially when dealing with complex illumination scenes, the introduction of this soft segmentation strategy significantly improves the naturalness of illumination transitions, while the hierarchical design ensures comprehensive illumination modeling from the pixel level to the semantic level. Experimental results show that this design can effectively improve the model's adaptability to different illumination conditions while maintaining low computational complexity.

[0239] 2.3.3 Analysis of characteristic modulation mechanism

[0240] After verifying the overall effectiveness of the dual feature modulation mechanism, this embodiment deeply analyzes the working principle and synergistic effect of IGWM and IGFM. As shown in Table 5, the experiment compares the performance of different modulation strategies in detail, where MS stands for multi-scale wavelet decomposition, the decomposition scale used is 3, and RA stands for regional adaptive enhancement.

[0241] Table 5 Analysis of quantitative results of ablation of the feature modulation mechanism

[0242]

[0243] The experimental results show that the model M 12 In the absence of feature interaction and region adaptive enhancement, the PSNR of the model is 22.153dB. 13 After using multi-scale wavelet decomposition, the performance is improved to 22.422dB, while the model M 14 The regional adaptive enhancement alone achieves 22.370dB. The complete modulation mechanism achieves optimal performance through the synergy of components, verifying the complementarity of the two enhancement strategies.

[0244] like Fig.12 As shown in (b), the features select the first three channels with the most significant variance for analysis, and the features of IGWM show a semantic response. Through spectral analysis, it can be observed that different frequency bands show differentiated modulation characteristics. In the low-frequency component area, energy gain is generated after modulation, which mainly affects the overall brightness and contrast of the image. The energy improvement in the mid-frequency area is most significant, which is directly related to the enhancement of edge and main structural information. In the high-frequency component area, the modulation is relatively conservative. This strategy effectively suppresses noise amplification while maintaining the necessary detail information. When the IGWM module is removed, a global modulation response is presented. This global response is difficult to effectively enhance the local illumination changes accurately.

[0245] IGFM achieves more accurate feature enhancement and obtains more lighting details by introducing a lighting-aware spatial attention mechanism. Fig.12 As shown in (c), the module can dynamically adjust the intensity of feature responses according to local lighting conditions. In low-light areas, stronger feature activation is generated; while in normal-light areas, the weight is reduced, effectively avoiding over-enhancement. Especially in the lighting transition area, the weight presents a smooth gradient characteristic, ensuring a natural transition effect. Fig.12 As shown in the brightness histogram and RGB output results in (c), IGFM enhances the lighting details while ensuring the stability of the overall brightness distribution, allowing the image to recover more information from low-light scenes.

[0246] The synergistic effect of the dual modulation mechanism is as mentioned above Figure 2 As shown. IGWM first provides differentiated enhancement for different frequency components in the frequency domain, and then IGFM performs adaptive modulation in the feature domain to ensure that the enhancement effect is consistent with the scene lighting conditions. This cascaded modulation strategy not only expands the feature optimization space, but also achieves progressive enhancement from local details to global structures. It is worth noting that when the multi-scale wavelet decomposition is removed and only single-level wavelet decomposition is used, the PSNR decreases by 1.017dB, highlighting the importance of multi-scale information processing for improving the enhancement effect. Overall, the combination of wavelet domain modulation and feature domain modulation significantly improves the model's adaptability to different scenes, especially showing unique advantages when dealing with complex lighting conditions.

[0247] 2.3.4 Impact of loss function

[0248] The design of the loss function has an important impact on the training and final performance of the model. This embodiment studies the contribution of different loss components and their optimal combination through detailed comparative experiments. As shown in Table 6, the changing trend of the model performance is observed by gradually adding different loss terms.

[0249] Table 6 Analysis of the impact of loss function combination on network performance

[0250]

[0251] The experimental results show that only using Although the loss can ensure the basic reconstruction quality, it often leads to over-smoothing and loss of high-frequency information. Fig.13 As shown, in a scene with rich outdoor textures, only Loss is difficult to preserve detailed structure. Add After the loss, the PSNR is increased to 22.870dB, and the structural information is better preserved, especially the fidelity of the wall edges and texture areas is significantly improved. The color and visual naturalness are improved, such as the background of the text area, and the generated results are more in line with the characteristics of human visual perception.

[0252] The choice of loss function weights has a significant impact on model performance. The optimal weight configuration was determined through experimental comparison: λ1=1.0, λ2=0.5, λ3=0.01. This configuration achieves a good balance between pixel-level accuracy, structure preservation, and perceptual quality, and ultimately achieves a performance level of PSNR 23.167dB. In contrast, other weight configurations such as λ1=0.5 and λ1=2.0 perform well in some indicators, but it is difficult to take into account both reconstruction quality and visual effects at the same time.

[0253] In terms of detail preservation in indoor scenes, the advantage of the complete loss function is more obvious. Through comparative analysis, it is found that the model with the combination of three losses shows stronger robustness when processing complex textures, improving edge clarity while maintaining the naturalness of lighting transitions. This multi-objective optimization strategy not only improves the overall performance of the model, but also enhances its adaptability in different scenarios.

[0254] In general, through detailed ablation experiments and analysis, this embodiment verifies the effectiveness of hierarchical illumination guidance and multi-domain feature modulation. Experimental results show that the HIE module achieves accurate modeling of complex illumination scenes through multi-granular hierarchical illumination estimation, and the dual feature modulation mechanism forms a complementary enhancement effect in the wavelet domain and the feature domain, while the design of the multi-objective loss function ensures a balanced improvement in visual quality and reconstruction accuracy.

[0255] In summary, the present invention proposes a low-light image enhancement method based on hierarchical illumination guidance. By introducing multi-granularity local illumination modeling, an accurate description of complex illumination scenes is achieved. On this basis, a dual feature modulation mechanism is designed. Through the synergy of illumination-guided wavelet modulation and feature modulation, a progressive image quality improvement is achieved in different domains. Finally, an adaptive enhancement module is used to dynamically fuse multi-domain features to further optimize the enhancement effect.

[0256] Experimental results show that the method of this embodiment has achieved advanced performance on multiple public datasets. Especially when dealing with complex lighting scenes, this layered illumination estimation significantly improves the model's adaptability to lighting changes. The necessity of each module is verified through ablation experiments, especially the unique advantages of the dual feature modulation mechanism in maintaining details and suppressing noise. In addition, the method of this paper also has advantages in computational efficiency, with the number of model parameters being only 1.53M, which is lower than the mainstream deep learning method.

[0257] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described with reference to the preferred embodiments of the present invention, it should be understood by those skilled in the art that various changes may be made in form and details without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A low-light image enhancement method based on layered illumination guidance, characterized in that: The steps include: S1, obtaining a low-light image to be processed; S2, using the low-light image to be processed as an input of a trained low-light image enhancement model, and outputting an image enhancement result of the low-light image to be processed; the low-light image enhancement model includes a hierarchical illumination estimator module, a dual illumination feature modulation module and an adaptive enhancement module; The training process of the low-light image enhancement model is as follows: S201, the hierarchical illumination estimator module uses the low-light image as a training sample input to the low-light image enhancement model as the input of the shallow illumination estimator unit in the hierarchical illumination estimator module, performs shallow feature extraction through a multi-granular illumination mapping function to obtain a low-level visual feature map, and uses the low-light image as the input of the deep illumination estimator unit to output a high-level semantic feature map; S202, the dual illumination feature modulation module modulates the low-level visual feature map and the high-level semantic feature map through an illumination-guided wavelet modulation unit and an illumination-guided feature modulation unit, respectively, to obtain a wavelet modulation feature map and a feature modulation feature map, respectively; S203, the adaptive enhancement module splices the wavelet modulation feature map and the feature modulation feature map, performs feature adaptive fusion and image-level enhancement processing on the spliced ​​features, and finally outputs an enhanced illumination enhancement result image; S204, using a multi-objective loss function consisting of a pixel-level reconstruction loss function, a structural similarity loss function, and a perceptual loss function to calculate the training loss of the low-light image enhancement model, and updating the parameters of the low-light image enhancement model with the goal of minimizing the multi-objective loss function; S205. Repeat steps S201 to S204 until the low-light image enhancement model converges or reaches a preset number of training times, and the training is completed.

2. The low-light image enhancement method based on layered illumination guidance according to claim 1, characterized in that: In step S201, the hierarchical illumination estimator module includes a shallow illumination estimator unit and a deep illumination estimator unit; wherein the shallow illumination estimator unit includes a cascaded color space transformation module, a convolution layer, a Gaussian filter layer, a multi-granularity illumination mapping function module and a normalization layer, and the deep illumination estimator unit includes a cascaded first convolution layer, an instance normalization layer, a Re-LU activation function layer and a second convolution layer; In the hierarchical illumination estimator module, the low-light image used as the input of the hierarchical illumination estimator module is used as the input of the shallow illumination estimator unit and the deep illumination estimator unit respectively, and the output of the shallow illumination estimator unit obtains a low-level visual feature map. The output of the shallow illumination estimator unit and the low-light image used as the input are both used as the input of the deep illumination estimator unit, and the output of the deep illumination estimator obtains a high-level semantic feature map.

3. The low-light image enhancement method based on layered illumination guidance according to claim 2, characterized in that: The processing of the shallow illumination estimator unit is: I μ =e shallow (I low )=μ(L norm ); In the formula, Represents the brightness map extracted by Gaussian filtering introduced by the shallow illumination estimator unit, Indicates the transformation operation from RGB space to HSV space and intercepts the last channel V, G σ represents a Gaussian kernel with standard deviation σ, represents the convolution operation, I low represents the input low-light image, L norm Indicates the normalization of the brightness image after Gaussian filtering, L min and L max They represent the minimum and maximum values ​​of the brightness map, ∈ represents the numerical stability constant, γ represents the normalization coefficient, μ d , μ n , μ b They represent the dark area, normal area, and bright area of ​​the illumination distribution map, respectively. d, n, and b represent dark, normal, and bright, respectively. (a, b) represents the characteristic threshold of the illumination interval. δ and β represent the steepness of the Sigmoid function. represents a three-channel light map for the shallow estimator, Represents each component, I μ Represents ε shallow The output low-level visual feature map, ε shallow (·) denotes a shallow lighting estimator unit. The processing of the deep illumination estimator unit is: In the formula, F μ Represents ε deep The output high-level semantic feature map, ε deep (·) denotes a high-level illumination estimator unit, [·,·] denotes feature concatenation in the channel dimension, ReLU denotes the ReLU activation function, IN denotes instance normalization, and They represent the weights of the two convolutional layers respectively, and k represents the convolution kernel size.

4. The low-light image enhancement method based on layered illumination guidance according to claim 1, characterized in that: In step S202, the dual illumination feature modulation module includes an illumination-guided wavelet modulation unit and an illumination-guided feature modulation unit; wherein the illumination-guided wavelet modulation unit includes a wavelet coefficient processor network, an illumination modulation network and a regional enhancement unit cascaded in sequence, and the illumination-guided feature modulation unit includes three feature extraction units, a connection layer, a feature fusion layer, a channel attention unit and a spatial attention unit in parallel, and the cascade results of the three feature extraction units in parallel, the connection layer and the feature fusion layer are connected in parallel with the cascade results of the attention unit and the spatial attention unit, wherein the three feature extraction units in parallel are respectively a feature extraction unit with a convolution kernel of 1, a feature extraction unit with a convolution kernel of 2 and a feature extraction unit with a convolution kernel of 3; In the illumination-guided wavelet modulation unit, a low-light image as an input of a low-light image enhancement model is subjected to feature extraction to obtain a basic feature map, the basic feature map is used as an input of a wavelet coefficient processor network in the illumination-guided wavelet modulation unit, the output of the wavelet coefficient processor network is spliced ​​with a low-level visual feature map output by a shallow illumination estimator unit in a hierarchical illumination estimator module, and the result is used as an input of an illumination modulation network, the output of the illumination modulation network is used as an input of a regional enhancement unit, and an inverse wavelet transform is performed on the output of the regional enhancement unit to obtain a wavelet modulation feature map as an output of the illumination-guided wavelet modulation unit; In the illumination-guided feature modulation unit, the high-level semantic feature map output by the deep illumination estimator unit in the hierarchical illumination estimator module is used as the input of the three feature extraction units in parallel in the illumination-guided feature modulation unit. The output of the three feature extraction units in parallel is used as the input of the feature fusion layer after passing through the connection layer. Then, the wavelet modulation feature map output by the illumination-guided wavelet modulation unit is used as the input of the fusion attention unit. The output of the fusion attention unit is multiplied with the wavelet modulation feature map and used as the input of the spatial attention unit. The output of the spatial attention unit is multiplied with the output of the fusion attention unit and then multiplied with the output of the feature fusion layer. The output is added to the input wavelet modulation feature map to obtain the feature modulation feature map output by the illumination-guided feature modulation unit.

5. The low-light image enhancement method based on layered illumination guidance according to claim 4, characterized in that: The wavelet coefficient processor network in the illumination-guided wavelet modulation unit includes a cascaded normalization unit and a convolution activation unit, the normalization unit includes a convolution layer, an instance normalization layer and an RReLU activation function layer connected in sequence, and the convolution activation unit includes a convolution layer and a ReLU activation function layer connected in sequence; wherein, the input feature is used as the input of the normalization unit, the output of the normalization unit is used as the input of the first convolution activation unit, the output of the first convolution activation unit is used as the input of the second convolution activation unit, the output of the second convolution activation unit is added to the output of the normalization unit and then passed through the Sigmoid activation function to obtain the output of the wavelet coefficient processor network; The illumination modulation network includes a cascaded normalization unit and a convolution layer; wherein the output of the wave coefficient processor network is input to the illumination modulation network and the low-level visual feature map output by the shallow illumination estimator unit for scaling and concatenation, and the output of the normalization unit is used as the input of the convolution layer, and the output of the convolution layer is subjected to a Sigmoid activation function to obtain the output of the illumination modulation network; The region enhancement unit includes a normalization unit, and the normalization unit includes a convolution layer, an instance normalization layer and an RReLU activation function layer connected in sequence.

6. The low-light image enhancement method based on layered illumination guidance according to claim 5, characterized in that: The processing process of the light-guided wavelet modulation unit is as follows: In the formula, represents the low frequency subband, represents the high frequency subband of the jth layer, L represents the number of decomposition layers, represents discrete wavelet transform, F represents the basic feature map obtained by extracting features from the input low-light image, and Represents low-frequency and high-frequency processors, respectively. and Respectively represent the output after being processed by the low-frequency and high-frequency processors, and represents the light modulation network, σ (·) represents the Sigmoid activation function, ⊙ represents element-by-element multiplication, Y' l and Respectively indicate that and The output after the light modulation network is processed, and They represent the enhancers for different illumination areas, μ i Indicates the degree of membership of the corresponding region, Z l and Respectively indicate that and The output after processing by the region enhancement unit, F w represents the output of the illumination-guided wavelet modulation unit, Represents inverse discrete wavelet transform.

7. The low-light image enhancement method based on layered illumination guidance according to claim 4, characterized in that: The processing process of the light guidance feature modulation unit is as follows: F m =F w ⊙W c ⊙W s ⊙F light +F w ; In the formula, represents the feature extraction branch with different convolution kernel sizes, k represents the convolution kernel size, and F k Represents the output of three feature extraction units connected in parallel, represents the feature fusion layer, F light represents the output after the feature fusion layer, represents the global average pooling operation, represents a two-layer perceptron, W c represents the output of the channel attention unit, W s represents the output of the spatial attention unit, and Respectively represent the average and maximum pooling operations in the channel dimension, W conv represents the convolutional layer parameters, F m Represents the output of the illumination-guided feature modulation unit.

8. The low-light image enhancement method based on layered illumination guidance according to claim 2, characterized in that: In step S203, the adaptive enhancement module includes a feature fusion unit and an image level enhancement unit; The feature fusion unit includes a first depth convolution unit, two cascaded convolution activation units, a second depth convolution unit, a convolution layer, a Sigmoid activation function layer and an attention unit connected in sequence, wherein the first depth convolution unit includes two cascaded depth convolution layers, the convolution activation unit includes a cascaded convolution layer and a ReLU activation function layer, and the second depth convolution unit is one depth convolution layer; Among them, after splicing the wavelet modulation feature map output by the illumination-guided wavelet modulation unit in the dual illumination feature modulation module and the feature modulation feature map of the illumination-guided feature modulation unit, they are respectively used as the input of the first deep convolution unit and the third convolution layer in the feature fusion unit in the adaptive enhancement module, the output of the first deep convolution unit is used as the input of the two cascaded convolution activation units, the output of the two cascaded convolution activation units and the output of the first deep convolution unit are further added, and then used as the input of the third deep convolution layer, the output of the third deep convolution layer and the output of the third convolution layer are multiplied by the output of the attention unit after passing through the Sigmoid activation function layer, and the output is multiplied by the basic feature map obtained by feature extraction of the differential attention and the input low-light image, and then added to obtain the output of the feature fusion unit; The image-level enhancement unit includes a first convolution layer, a plurality of residual blocks and a second convolution layer connected in sequence, the residual block includes three cascaded nonlinear convolution activation units, a convolution layer, a maximum pooling layer and a Sigmoid activation function layer, and the nonlinear convolution activation unit includes a cascaded convolution layer and an LReLU activation function layer; wherein, in the residual block, the input feature map is respectively used as the input of the convolution layer and the three cascaded nonlinear convolution activation units, the output of the three cascaded nonlinear convolution activation units and the output of the convolution layer are multiplied after passing through the maximum pooling layer and the Sigmoid activation function layer, and the output result is added to the input of the residual block to obtain the output of the residual block; Among them, the output of the feature fusion unit is used as the input of the first convolutional layer in the image-level enhancement unit, the output of the first convolutional layer is used as the input of the first residual block, the output of each residual block is used as the input of the next layer in the image-level enhancement unit, the output of the last residual block is used as the input of the second convolutional layer, and the output of the second convolutional layer is used as the output of the feature fusion unit.

9. The low-light image enhancement method based on layered illumination guidance according to claim 8, characterized in that: The processing process of the feature fusion unit is as follows: I fusion =A⊙F fused +(1-A)⊙F; In the formula, c represents the ordinary convolution layer, A represents the attention unit matrix, represents a network layer composed of a series of depth-separable convolutions, ordinary convolutions, and residual connections, F fused Indicates passing Processed output, F' m represents the output after concatenating the wavelet modulation feature map and the feature modulation feature map, (1-A) represents the complementary attention unit matrix; The processing process of the image level enhancement unit is: In the formula, F fusion For I fusion The features reprojected into the feature space, F res and Both represent the output of the i-th residual block, The last convolutional layer is 1×1 and maps the features back to the RGB image space. N is the number of residual blocks, and I normal Represents the illumination enhancement result image after being enhanced through the output of the image-level enhancement unit.

10. The low-light image enhancement method based on layered illumination guidance according to claim 2, characterized in that: In step S204, the pixel-level reconstruction loss function is: In the formula, represents the pixel-level reconstruction loss function, ∥·∥1 represents the L1 norm, and I gt represents the target image; The structural similarity loss function is: In the formula, represents the structural similarity loss function, SSIM(·) represents the indicator for measuring the visual similarity between two images, μ x and μ y Both represent the local mean, σ x and σ y Both represent standard deviation, σ xy represents covariance, C1 and C2 both represent stability constants; The perceptual loss function is: In the formula, is the perceptual loss function, φ l (·) represents the feature extractor of the lth layer of the VGG network, w l represents the learned channel weight, H l and W l Represent the height and width of the feature map respectively; Then the multi-objective loss function is: In the formula, is a multi-objective loss function, and λ1, λ2 and λ3 are weight coefficients.

Images