Image enhancement method and system for low-lightness scene

Abstract

This invention belongs to the field of low-light image enhancement technology, specifically relating to an image enhancement method and system for low-light scenes, including the following steps: 1. Acquire a low-light image, perform dimensionality reduction processing, and divide it into illuminance and reflectance components; 2. Perform local brightness adaptive enhancement and structural detail preservation on the illuminance component, generating illuminance component enhancement features and detail features; 3. Perform chromaticity calibration on the reflectance component, generating chromaticity features; 4. Perform interaction and aggregation of complementary features to generate fused features; 5. Perform feature normalization processing and cross-spatial information transfer between features on the input fused features to generate refined detail features, and supplement high-frequency information, fusing the original features with the detail features to generate a normal light image. This method achieves synergistic optimization of brightness restoration and color fidelity, reduces visual distortion, model computational complexity and parameter overhead, and improves the feasibility and operational efficiency of edge deployment.

Description

Technical Field

[0001] This application belongs to the field of low-light image enhancement technology, specifically relating to an image enhancement method and system for low-light scenes. Background Technology

[0002] In recent years, low-light image enhancement (LLIE) has become a key task in the field of computer vision. The core task of LLIE is to recover normal-light images with appropriate brightness, clear details, and natural colors from low-light input in low-light scenes such as nighttime and tunnels. These images must simultaneously meet the requirements of "discernible details in dark areas," "suppressed noise," and "colors consistent with the real scene" to support the reliable operation of downstream computer vision tasks such as nighttime target detection and autonomous driving environmental perception.

[0003] Compared to traditional frame-based image enhancement methods, the existing LLIE technique, based on Retinex, has a roughly defined network diagram as follows: Figure 1 As shown, by decomposing the image, enhancing details, and then reconstructing, the dark areas become clearly visible. End-to-end mapping optimization is achieved using deep learning, significantly improving adaptability in complex lighting scenarios. However, the following technical issues still exist:

[0004] (1) Brightness degradation is severe and noise is amplified under low light conditions.

[0005] (2) It is difficult to ensure color consistency while enhancing brightness and restoring structure, and problems such as color distortion, over-enhancement or loss of detail are likely to occur.

[0006] (3) Some complex architectures have a large number of parameters and computational overhead, which is not conducive to lightweight hardware deployment.

[0007] Therefore, a new method is urgently needed to achieve synergistic optimization of brightness restoration and color fidelity. Summary of the Invention

[0008] To address the aforementioned technical problems in the existing technology, the purpose of this invention is to achieve synergistic optimization of brightness restoration and color fidelity, and the technical solution is as follows:

[0009] An image enhancement method for low-light scenes includes the following steps:

[0010] Step 1: Acquire a low-light image, and perform dimensionality reduction processing on the low-light image to separate it into illuminance and reflectance components;

[0011] Step 2: Through multi-scale coding and gated modulation, the illuminance component is locally enhanced with adaptive brightness and structural detail preserved to generate illuminance component enhancement features and detail features;

[0012] Step 3: Perform chromaticity calibration on the reflection component by analyzing the statistical deviation of the chromaticity channel to generate chromaticity characteristics;

[0013] Step 4: Through multi-hypothesis modeling and cross-attention mechanism, complementary features are interacted and aggregated to generate fused features;

[0014] Step 5: Perform feature normalization on the input fusion features to eliminate distribution differences. Through normalization and multi-head attention matrix operations, perform cross-spatial information transfer between features to generate refined detail features and supplement high-frequency information. Then, fuse the original features with the refined detail features to generate a normal light image.

[0015] Furthermore, the specific process of step 2 is as follows:

[0016] Step 2.1: Extract brightness features using deep convolution with multiple dilation rates, and generate multi-scale aggregated features through channel fusion;

[0017] Step 2.2: The aggregated features are split into gain branches and activation branches after depthwise convolution and point projection through a gated multiplication interaction mechanism;

[0018] Step 2.3: Generate dynamic gain coefficients through Sigmoid activation, the expression of which is:

[0019]

[0020] in, Indicates the first One scale, Indicates a branch, This represents the global weight hyperparameters of the gain branch. This represents the kernel weights of the gain branch. The feature input representing the activation branch, This indicates element-wise multiplication. This represents the activation function. The feature input represents the Sigmoid activation. This is a multi-scale aggregation feature;

[0021] Step 2.4: A lightweight high-frequency detail recovery block is used. Through GELU activation and residual connections, suppressed high-frequency structural information is recovered. The original luminance features are then fused using learnable weighted residual interpolation to constrain the enhancement amplitude. The expression for this is:

[0022]

[0023] in, This represents suppressed high-frequency feature information. Indicates the first High-frequency feature fusion weights at various scales Indicates the first Weights for gain features at each scale Indicates high frequency.

[0024] Furthermore, in step 2.1, the brightness features include: local fine details, mesoscale brightness distribution, and global illumination trend;

[0025] The expansion rates are set to 1, 3 and 5 respectively.

[0026] Furthermore, the specific process of step 3 is as follows:

[0027] Step 3.1: Channel adaptive calibration performs global average pooling on low-light chromaticity features and learns channel importance weights by combining them with 1D convolution;

[0028] Step 3.2: Based on the chromaticity statistical benchmark of the normal light image, align the chromaticity distribution through scale-offset calibration, the expression of which is:

[0029]

[0030] in, , These are the learnable scale parameter and the offset parameter, respectively. Represents the original input chromaticity features. Indicates global average pooling. This represents a one-dimensional convolutional layer;

[0031] Step 3.3: Low-rank correction and residual interpolation are used to suppress redundant chromatic noise through rank-reduced projection, and learnable weights are applied. Integrate original color characteristics.

[0032] Furthermore, in step 3.2, the chromaticity statistical benchmark based on the normal light image includes: mean. and variance .

[0033] Furthermore, the specific process of step 4 is as follows:

[0034] Step 4.1: Luminosity branching generates multiple luminosity hypotheses in parallel. ;

[0035] Step 4.2: Perform feature-level cross-attention calculation on each brightness hypothesis and chromaticity feature respectively. Use the brightness feature as the query vector and the chromaticity feature as the key-value pair to perform complementary feature transfer. The chromaticity feature absorbs brightness information through the operation.

[0036] Step 4.3: Perform mean aggregation on the brightness and chromaticity features after interaction to generate fused features.

[0037] Furthermore, in step 4.1, the various brightness assumptions include: conservative brightening, balanced brightening, and aggressive brightening;

[0038] In step 4.2, the calculation expression for the feature-level cross-attention calculation of the brightness assumption and chromaticity features is as follows:

[0039]

[0040] in, Indicates the first A brightness assumption, Indicates the first The query vector corresponding to each brightness assumption The key vector representing the chromaticity features. Represents the key vector transpose, This represents the value vector generated by the chromaticity features. Indicates the dimensions of the query vector and key vector.

[0041] Furthermore, the specific process of step 5 is as follows:

[0042] Step 5.1: Perform feature normalization processing on the features to be fused using LayerNorm to eliminate the distribution differences of the features to be fused;

[0043] Step 5.2: Generate query vectors and key-value pairs through 1×1 convolution, and perform channel dimension matching and feature projection;

[0044] Step 5.3: Extract local spatial context features through depthwise convolution;

[0045] Step 5.4: Optimize the attention weight allocation strategy by adaptively adjusting the steepness of the attention distribution through learnable multi-head temperature parameters;

[0046] Step 5.5: Through normalization and multi-head attention matrix operations, cross-spatial information transfer between the two sets of features is performed to generate refined detailed features;

[0047] Step 5.6: By using channel stitching, 1×1 convolution fusion, and PReLU activation technology, the original features are fused with the refined detail features to generate a normal light image.

[0048] Furthermore, in step 5.1, the features to be fused include deep global semantic features and shallow high-frequency detail features.

[0049] An image enhancement system for low-light scenes includes the following modules:

[0050] Brightness-oriented structural feature enhancement module ISFE: It uses deep convolution with multiple dilation rates to capture local fine details, mesoscale brightness distribution and global illumination trends. It generates multi-scale aggregated features through channel fusion. According to the gated multiplication cross-control mechanism, it performs deep convolution and point projection on the aggregated features, splits them into gain branches and activation branches and performs adaptive adjustment.

[0051] The Chromaticity Statistical Consistency Enhancement Module (CSCE) performs global average pooling on low-light chromaticity features through adaptive calibration. Based on the channel importance weights learned by convolution and the chromaticity statistical benchmark based on normal light images, it aligns the chromaticity distribution through scale-offset calibration. It employs low-rank correction and residual difference techniques, suppresses redundant chromaticity noise through rank reduction projection, and fuses the original chromaticity features through learnable weights.

[0052] Multi-hypothesis Cross-Attention Module (MHCA): Through multi-hypothesis modeling and cross-attention mechanism, complementary features are interacted and aggregated to generate fused features;

[0053] The Attention-Driven Fusion Module (AGSC) performs feature normalization using LayerNorm to eliminate feature distribution differences. It generates query vectors and key-value pairs through 1×1 convolutions, performs channel dimension matching and feature projection, and extracts local spatial context features through depthwise convolutions. Based on learnable multi-head temperature parameters, it adaptively adjusts the steepness of the attention distribution and optimizes the attention weight allocation strategy. Through normalization and multi-head attention matrix operations, it performs cross-spatial information transfer between features, generating semantically guided refined detail features and supplementing high-frequency information. Finally, it fuses the original features with the refined detail features to output a normal light image.

[0054] Beneficial effects: (1) It achieves synergistic optimization of brightness restoration and color fidelity, significantly reducing visual distortion. (2) The structure is simple and conducive to lightweight hardware deployment. (3) It reduces the performance loss caused by feature entanglement. (4) It reduces the computational complexity and parameter overhead of the model. (5) It improves the feasibility and operating efficiency of edge deployment. Attached Figure Description

[0055] Figure 1 This is a diagram of the image enhancement network framework for low-light scenes according to the present invention;

[0056] Figure 2 This is a flowchart of the image enhancement method for low-light scenes according to the present invention;

[0057] Figure 3 This is a block diagram of the image enhancement system for low-light scenes according to the present invention;

[0058] Figure 4 This is a schematic diagram illustrating the working principle of the brightness-guiding structure feature enhancement module of the present invention.

[0059] Figure 5 This is a schematic diagram illustrating the working principle of the colorimetric statistical consistency enhancement module of the present invention.

[0060] Figure 6 This is a schematic diagram illustrating the working principle of the multi-hypothesis cross-attention module of the present invention.

[0061] Figure 7 This is a schematic diagram illustrating the working principle of the attention-driven fusion module of the present invention. Detailed Implementation

[0062] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0063] like Figure 2 As shown, an image enhancement method for low-light scenes according to the present invention includes the following steps:

[0064] S1: Acquire a low-light image, perform dimensionality reduction processing on the low-light image, and divide it into illuminance component and reflectance component;

[0065] S2: In low-light scenes, the exposure difference between dark and bright areas of an image is significant. Uniform brightening easily leads to loss of detail in dark areas and overexposure in bright areas, and simple brightness stretching amplifies noise. This invention achieves dual optimization of brightness and detail through a collaborative design of multi-scale coding and gated modulation. The specific process is as follows:

[0066] S2.1: Use multi-dilation depth convolution to extract brightness features, and set depth convolution with dilation rates of 1, 3 and 5 in parallel to capture local fine details, mesoscale brightness distribution and global illumination trend respectively. Generate multi-scale aggregated features through channel fusion.

[0067] S2.2: To avoid the inadequacy of a single scale in adapting to complex lighting, a gated multiplication interaction mechanism is used to split the aggregated features into gain branches and activation branches after depth convolution and point projection.

[0068] S2.3: Dynamic gain coefficients are generated through Sigmoid activation to achieve adaptive adjustment of high gain for brightening dark areas and low gain for bright areas. The expression is as follows:

[0069]

[0070] in, Indicates the first One scale, Indicates a branch, This represents the global weight hyperparameters of the gain branch. This represents the kernel weights of the gain branch. The feature input representing the activation branch, This indicates element-wise multiplication. This represents the activation function. The feature input represents the Sigmoid activation. This is a multi-scale aggregation feature;

[0071] S2.4: To prevent over-enhancement from causing detail smoothing, a lightweight high-frequency detail recovery block is used. This block recovers suppressed high-frequency structural information through GELU activation and residual connections. The original luminance features are then fused using learnable weighted residual interpolation to constrain the enhancement amplitude. The expression for this block is:

[0072]

[0073] in, This represents suppressed high-frequency feature information. Indicates the first High-frequency feature fusion weights at various scales Indicates the first Weights for gain features at each scale Indicates high frequency.

[0074] S3: To address chromatic degradation in low-light images, statistical bias, noise-induced color fluctuations, and hue shift techniques are employed. Through statistical calibration, hue constraint, and noise suppression, chromaticity features are enhanced with fidelity, avoiding distortions such as color cast and oversaturation during brightness enhancement. A three-layer progressive processing method ensures chromaticity consistency, as detailed below:

[0075] S3.1: Channel adaptive calibration performs global average pooling on low-light chromaticity features and combines 1D convolution to learn channel importance weights;

[0076] S3.2: Chromaticity statistical baseline mean based on normal light image and variance The chromaticity distribution is aligned using scale-offset calibration, and its expression is:

[0077]

[0078] in, , These are the learnable scale parameter and the offset parameter, respectively. Represents the original input chromaticity features. Indicates global average pooling. This represents a one-dimensional convolutional layer;

[0079] S3.3: Low-rank correction and residual interpolation are employed, and redundant chromatic noise is suppressed through rank-reduced projection, and learnable weights are used. It integrates the original color characteristics to prevent color distortion caused by overcorrection.

[0080] S4: Low-light images lack brightness information, and a single brightness feature is insufficient to support accurate matching of chroma and brightness, easily leading to "brightness-color mismatch," such as oversaturation of colors in dark areas. This invention achieves controllable and precise interaction of decoupled components through multi-hypothesis modeling and cross-attention mechanisms, specifically in two stages:

[0081] (1) Encoder stage

[0082] The luminance branch generates multiple luminance hypotheses in parallel. These correspond to three lighting scenarios: conservative brightening, balanced brightening, and aggressive brightening, respectively, covering the potential possibilities of low brightness; the chroma branch maintains a single feature representation to avoid computational redundancy introduced by multiple assumptions;

[0083] For each brightness assumption, feature-level cross-attention is calculated with the chromaticity feature. The brightness feature is used as the query vector, and the chromaticity feature is used as the key-value pair for complementary feature transfer. The expression is as follows:

[0084]

[0085] in, Indicates the first A brightness assumption, Indicates the first The query vector corresponding to each brightness assumption The key vector representing the chromaticity features. Represents the key vector transpose, This represents the value vector generated by the chromaticity features. Indicates the dimensions of the query vector and key vector;

[0086] At the same time, chromaticity features absorb luminance information through computation.

[0087] (2) Decoder stage

[0088] The brightness and chromaticity features after interaction are aggregated by mean value to eliminate the uncertainty of a single assumption and generate stable fused features.

[0089] S5: In low-light image enhancement and semantic segmentation, there are differences in feature distribution between the two sets of features to be fused, namely deep global semantic features and shallow high-frequency detail features. This invention achieves adaptive and accurate interaction between features through feature normalization, attention-driven refinement, and cross-feature fusion, taking into account both semantic integrity and detail richness. The specific process is as follows:

[0090] S5.1: The input features to be fused are normalized using an independent LayerNorm to eliminate the distribution differences of the features to be fused;

[0091] S5.2: Generate query vectors using 1×1 convolutions. key-value pairs Perform channel dimension matching and feature projection;

[0092] S5.3: Extracting local spatial context features through depthwise convolution;

[0093] S5.4: By learning multi-head temperature parameters, the steepness of the attention distribution is adaptively adjusted to optimize the attention weight allocation strategy;

[0094] S5.5: Through normalization and multi-head attention matrix operations, cross-spatial information transfer between two sets of features is performed to generate semantically guided refined detailed features and accurately complete high-frequency information.

[0095] S5.6: Employing channel stitching, 1×1 convolutional fusion, and PReLU activation techniques, the original features are fused with attention-refined detail features, compressing channel dimensions while enhancing non-linear feature expression, ultimately generating a normal light image.

[0096] like Figure 3 As shown, an image enhancement system for low-light scenes includes the following modules:

[0097] Luminance-Guided Structure Feature Enhancement Module (ISFE): such as Figure 4 As shown, deep convolution with multiple dilation rates is used to capture local fine details, mesoscale brightness distribution and global illumination trends. Multi-scale aggregated features are generated through channel fusion. The aggregated features are then subjected to deep convolution and point projection according to the gated multiplication cross-control mechanism, and are split into gain branches and activation branches and adaptively adjusted.

[0098] Colorimetric Statistical Consistency Enhancement Module (CSCE): such as Figure 5As shown, global average pooling is performed on low-light chromaticity features through adaptive calibration. Based on the channel importance weights learned by convolution and the chromaticity statistical benchmark based on normal light images, the chromaticity distribution is aligned through scale-offset calibration. Low-rank correction and residual difference techniques are used. Redundant chromaticity noise is suppressed by rank reduction projection. The original chromaticity features are fused through learnable weights.

[0099] Multiple Hypothesis Cross-Attention Module (MHCA): such as Figure 6 As shown, through multi-hypothesis modeling and cross-attention mechanism, complementary features are interacted and aggregated to generate fused features;

[0100] Attention-Driven Fusion Module (AGSC): such as Figure 7 As shown, LayerNorm is used to normalize the features to be fused, eliminating the differences in feature distribution. Query vectors and key-value pairs are generated through 1×1 convolution, and channel dimension matching and feature projection are performed. Local spatial context features are extracted through depthwise convolution. Based on the learnable multi-head temperature parameters, the steepness of the attention distribution is adaptively adjusted to optimize the attention weight allocation strategy. Cross-spatial information transfer between features is performed through normalization and multi-head attention matrix operations, generating semantically guided refined detail features and supplementing high-frequency information. The original features and the refined detail features are then fused to output a normal light image.

Claims

1. A method for image enhancement for low-lightness scenes, characterized in that, Includes the following steps: Step 1: Acquire a low-light image, and perform dimensionality reduction processing on the low-light image to separate it into illuminance and reflectance components; Step 2: Through multi-scale coding and gated modulation, the illuminance component is locally enhanced with adaptive brightness and structural detail preserved to generate illuminance component enhancement features and detail features; Step 3: Perform chromaticity calibration on the reflection component by analyzing the statistical deviation of the chromaticity channel to generate chromaticity characteristics; Step 4: Through multi-hypothesis modeling and cross-attention mechanism, complementary features are interacted and aggregated to generate fused features; Step 5: Perform feature normalization on the input fusion features to eliminate distribution differences. Through normalization and multi-head attention matrix operations, perform cross-spatial information transfer between features to generate refined detail features and supplement high-frequency information. Then, fuse the original features with the refined detail features to generate a normal light image.

2. The image enhancement method for low-lightness scene according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Extract brightness features using deep convolution with multiple dilation rates, and generate multi-scale aggregated features through channel fusion; Step 2.2: The aggregated features are split into gain branches and activation branches after depthwise convolution and point projection through a gated multiplication interaction mechanism; Step 2.3: Generate dynamic gain coefficients through Sigmoid activation, the expression of which is: wherein, represents the i-th scale, represents a branch, represents a global weight hyperparameter for the gain branch, represents a convolution kernel weight for the gain branch, represents a feature input for the activation branch, represents an element-wise multiplication, represents an activation function, represents a feature input for the Sigmoid activation, is a multi-scale aggregated feature;​ Step 2.4: A lightweight high-frequency detail recovery block is used. Through GELU activation and residual connections, suppressed high-frequency structural information is recovered. The original luminance features are then fused using learnable weighted residual interpolation to constrain the enhancement amplitude. The expression for this is: wherein, represents the suppressed high-frequency characteristic information, represents the high-frequency characteristic fusion weight of the first scale, represents the gain characteristic fusion weight of the first scale, represents the high-frequency.

3. The image enhancement method for low-lightness scene according to claim 2, characterized in that, In step 2.1, the brightness features include: local fine details, mesoscale brightness distribution, and global illumination trend; The expansion rates are set to 1, 3 and 5 respectively.

4. The image enhancement method for low-lightness scene according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3.1: Channel adaptive calibration performs global average pooling on low-light chromaticity features and learns channel importance weights by combining them with 1D convolution; Step 3.2: Based on the chromaticity statistical benchmark of the normal light image, align the chromaticity distribution through scale-offset calibration, the expression of which is: wherein, , are learnable scale and shift parameters, respectively, denotes the original input chroma feature, denotes global average pooling, denotes a one-dimensional convolutional layer; Step 3.3: Low-rank correction and residual interpolation are used to suppress redundant chromatic noise through rank-reduced projection, and learnable weights are applied. Integrate original color characteristics.

5. The image enhancement method for low-light scenes according to claim 5, characterized in that, In step 3.2, the chromaticity statistical benchmark based on the normal light image includes: mean. and variance .

6. The image enhancement method for low-light scenes according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1: Luminosity branching generates multiple luminosity hypotheses in parallel. ; Step 4.2: Perform feature-level cross-attention calculation on each brightness hypothesis and chromaticity feature respectively. Use the brightness feature as the query vector and the chromaticity feature as the key-value pair to perform complementary feature transfer. The chromaticity feature absorbs brightness information through the operation. Step 4.3: Perform mean aggregation on the brightness and chromaticity features after interaction to generate fused features.

7. The image enhancement method for low-light scenes according to claim 6, characterized in that, In step 4.1, the various brightness assumptions include: conservative brightening, balanced brightening, and aggressive brightening; In step 4.2, the calculation expression for the feature-level cross-attention calculation of the brightness assumption and chromaticity features is as follows: in, Indicates the first A brightness assumption, Indicates the first The query vector corresponding to each brightness assumption The key vector representing the chromaticity features. Represents the key vector transpose, This represents the value vector generated by the chromaticity features. Indicates the dimensions of the query vector and key vector.

8. The image enhancement method for low-light scenes according to claim 1, characterized in that, The specific process of step 5 is as follows: Step 5.1: Perform feature normalization processing on the features to be fused using LayerNorm to eliminate the distribution differences of the features to be fused; Step 5.2: Generate query vectors and key-value pairs through 1×1 convolution, and perform channel dimension matching and feature projection; Step 5.3: Extract local spatial context features through depthwise convolution; Step 5.4: Optimize the attention weight allocation strategy by adaptively adjusting the steepness of the attention distribution through learnable multi-head temperature parameters; Step 5.5: Through normalization and multi-head attention matrix operations, cross-spatial information transfer between the two sets of features is performed to generate refined detailed features; Step 5.6: By using channel stitching, 1×1 convolution fusion, and PReLU activation technology, the original features are fused with the refined detail features to generate a normal light image.

9. The image enhancement method for low-light scenes according to claim 8, characterized in that, In step 5.1, the features to be fused include deep global semantic features and shallow high-frequency detail features.

10. An image enhancement system for low-light scenes, characterized in that, Includes the following modules: Brightness-oriented structural feature enhancement module ISFE: It uses deep convolution with multiple dilation rates to capture local fine details, mesoscale brightness distribution and global illumination trends. It generates multi-scale aggregated features through channel fusion. According to the gated multiplication cross-control mechanism, it performs deep convolution and point projection on the aggregated features, splits them into gain branches and activation branches and performs adaptive adjustment. The Chromaticity Statistical Consistency Enhancement Module (CSCE) performs global average pooling on low-light chromaticity features through adaptive calibration. Based on the channel importance weights learned by convolution and the chromaticity statistical benchmark based on normal light images, it aligns the chromaticity distribution through scale-offset calibration. It employs low-rank correction and residual difference techniques, suppresses redundant chromaticity noise through rank reduction projection, and fuses the original chromaticity features through learnable weights. Multi-hypothesis Cross-Attention Module (MHCA): Through multi-hypothesis modeling and cross-attention mechanism, complementary features are interacted and aggregated to generate fused features; The Attention-Driven Fusion Module (AGSC) performs feature normalization using LayerNorm to eliminate feature distribution differences. It generates query vectors and key-value pairs through 1×1 convolutions, performs channel dimension matching and feature projection, and extracts local spatial context features through depthwise convolutions. Based on learnable multi-head temperature parameters, it adaptively adjusts the steepness of the attention distribution and optimizes the attention weight allocation strategy. Through normalization and multi-head attention matrix operations, it performs cross-spatial information transfer between features, generating semantically guided refined detail features and supplementing high-frequency information. Finally, it fuses the original features with the refined detail features to output a normal light image.

Images