Low-light image enhancement method based on retina theory and implicit neural representation

Abstract

This invention discloses a low-light image enhancement method based on retinal theory and implicit neural expression, comprising: acquiring a low-light image; inputting the low-light image into a latent decomposition module based on retinal theory to decompose the image into a reflection component and an illumination component; inputting the obtained illumination component into an illumination enhancer based on implicit neural expression to generate a normal illumination image; recombining the normal illumination image and the reflection component to obtain an illumination-corrected image; inputting the original low-light image into a multi-scale structure enhancement auxiliary branch to generate a structure enhancement image, and using the illumination information of the main branch to align the brightness of the auxiliary branch output; and adaptively fusing the illumination-corrected image and the brightness-aligned structure enhancement image through a lightweight fusion network to output the final enhanced image. This invention solves the problems of existing methods, such as loss of reflection details due to high coupling between illumination and structure under extremely low illumination, spatial artifacts generated by discrete illumination modeling, and over-smoothing of high-frequency details by implicit neural expression.

Description

Technical Field

[0001] This invention belongs to the field of computer digital image processing technology, specifically relating to a low-light image enhancement method based on retinal theory and implicit neural expression. Background Technology

[0002] Images captured by cameras in low-light environments often suffer from insufficient brightness, reduced contrast, and significant noise, severely impacting subjective visual quality and limiting the performance of high-level visual tasks such as object detection and semantic segmentation. Directly stretching brightness or adjusting contrast in low-light images, while increasing brightness, amplifies existing noise. Furthermore, lighting in real-world scenes is inherently continuous, smooth, and spatially dependent, while existing methods mostly model lighting in discrete pixel spaces or feature maps. This discretization approach struggles to capture the continuous changes in lighting, easily leading to spatial discontinuities, blocky artifacts, and leakage of structural information. To balance brightness restoration and structural preservation during enhancement, this invention proposes a low-light image enhancement method based on retinal theory and implicit neural expression.

[0003] Extensive research has been conducted in the field of low-light image enhancement. Traditional methods, such as histogram equalization, gamma correction, and decomposition methods based on retinal theory, enhance brightness through pixel remapping or illumination estimation. However, these methods rely on simplified physical priors, are sensitive to parameters, and tend to amplify noise and destroy local structures in extremely dark areas. In recent years, deep learning-based methods have made significant progress, such as decomposition enhancement methods that combine retinal theory with neural networks, and end-to-end models that directly learn low-light to normal-light mappings. However, these methods still estimate illumination in the image domain or discrete feature domain. Under extremely low illumination conditions, illumination changes are highly coupled with texture structure, leading to the erroneous absorption of structural information into the illumination component, resulting in the loss of reflection details and structural artifacts in the illumination map. Furthermore, recently emerging implicit neural representation methods attempt to continuously model illumination, but most existing methods directly regress illumination values ​​from pixel coordinates, ignoring the intrinsic coupling between illumination and reflection. Moreover, the low-pass characteristics of implicit neural networks can easily over-smooth high-frequency details. Summary of the Invention

[0004] The purpose of this invention is to provide a low-light image enhancement method based on retinal theory and implicit neural expression, which solves the problems of existing methods such as loss of reflection details due to high coupling between illumination and structure under extremely low illumination, spatial artifacts generated by discrete illumination modeling, and over-smoothing of high-frequency details by implicit neural expression.

[0005] The technical solution adopted in this invention is a low-light image enhancement method based on retinal theory and implicit neural expression, comprising the following steps: acquiring an input low-light image; inputting the low-light image into a latent decomposition module based on retinal theory to decompose the image into a reflection component and an illumination component; inputting the decomposed illumination component into an illumination enhancer based on implicit neural expression to generate a normal illumination image; recombining the normal illumination image and the reflection component to obtain an illumination-corrected image; simultaneously inputting the original low-light image into a multi-scale structural enhancement auxiliary branch to generate a structural enhancement image, and using the illumination information of the main branch to perform brightness alignment on the output of the auxiliary branch; adaptively fusing the illumination-corrected image and the brightness-aligned structural enhancement image through a lightweight fusion network to output the final enhanced image.

[0006] The invention is further characterized by: The specific steps are as follows: Step 1: Input a low-light image I For digital images created in low-light environments, input a low-light image. I The spatial resolution is H×W; where H represents the height in pixels, W represents the width in pixels, and there are 3 color channels, corresponding to the red, green, and blue channels, i.e., I∈R. (H×W×3) Input low-light image I The pixel value range of each pixel is normalized to the interval [0,1], where 0 represents all black and 1 represents all white; Step 2: Input low-light image I Input a latent decomposition module based on retinal theory, which decomposes the image into reflectance components in the latent feature space. R ^ and light component L ^ ; Step 3: Decompose the obtained light components L ^ Input a light intensifier based on implicit neural expression, the intensifier for L ^ Downsampling is performed to obtain a low-resolution illumination map. L lr Using spatial coordinates and local context neighborhood as conditions, a sinusoidal representation network is employed to predict illumination residuals, generating spatially continuous and smoothly varying normal illumination maps. L Fixed lr Guided filtering is used to restore the illumination map to its original resolution, resulting in a corrected normal illumination map. L Fixed ; Step 4: Extract the original low-light image IThe input multi-scale structure enhancement auxiliary branch employs an encoder-decoder structure, introducing channel attention and spatial attention mechanisms at different spatial resolutions to explicitly model high-frequency structures such as edges and textures, generating structure-enhanced images. I MSA The expression is shown in equation (1); (1); Step 5: Design a brightness alignment module, using the main branch output as a brightness reference, and perform global brightness normalization on the auxiliary branch output; Specifically, the luminance component of the RGB image is calculated. For any RGB image, the luminance component is obtained by weighted summation of each color channel, as shown in equation (2). (2); in, I (c) This represents the c-th color channel of the image. oh c This refers to the brightness coefficient. Calculate the illumination correction images of the main branches separately. I Fixed luminance component Y Fixed Image enhancement with auxiliary branch structures I MSA luminance component Y MSA The global scaling factor s is estimated by matching the spatial average brightness of the two, as shown in equation (3). (3); in, m(·) This indicates that a spatial averaging operation is performed on the image plane, which calculates the average brightness of all pixels; e It is a small constant used to ensure numerical stability; The auxiliary branch output is rescaled as a whole using the global scaling factor to obtain the brightness-aligned structural image, as shown in equation (4). (4); Step 6: Obtain the structure enhancement result after brightness alignment. I BA Then, a lightweight fusion network is designed to process the illumination-corrected image. I Fixed Structure image aligned with brightness I BA Adaptive fusion is performed, the fusion network predicts pixel-level weight mapping, and the contributions of illumination correction and structural enhancement are dynamically balanced according to spatial location to output the final enhanced image. The mathematical expression is shown in Equation (5). (5); in, F Fuse ( By predicting pixel-level weight mappings W∈[0,1] H×W The weighted fusion of the two results is achieved, and the mathematical expression is shown in equation (6). (6).

[0007] Input a low-light image in step 1 I Any low-light image captured in nighttime surveillance, indoor low-light, or backlit scenarios.

[0008] Input a low-light image in step 1 I This includes degradation features such as insufficient brightness, low contrast, noticeable noise, and blurred details.

[0009] Step 2 includes: estimating coarse illumination by channel-wise maximum operation, introducing a feature interaction mechanism between the reflection branch and the illumination branch, and redistributing structural information from the illumination component to the reflection component, thereby obtaining a structurally complete reflection component and a spatially smooth illumination component.

[0010] Step 2 is implemented in the following steps: Step 2.1: Transfer the low-light image I Modeled as reflection component R With light component L The element-wise product is expressed as shown in equation (7); (7); Where ⊙ represents element-wise multiplication; R It represents the inherent content of the scene, including the inherent properties of texture, edges, and color; L This represents a smooth spatial distribution of light, under ideal conditions. R Preserve complete high-frequency details. L Maintain low-frequency, smooth, and continuous changes; Step 2.2: In order to achieve decomposition in the latent feature space, the input low-light image is... I By mapping to the feature space through convolutional layers, we obtain the feature representation F(I) ∈ R. H×W×C , where C is the total number of channels; the maximum value operation of each channel is used to extract a rough estimate of the illumination intensity from the feature response. This operation assumes that at each spatial location, the maximum response of the feature channel can reflect the dominant illumination intensity at that location, as shown in Equation (8); (8); in, F c(I)Feature representation F ( I ) c Each channel response, max c(·) This indicates taking the maximum value along the channel dimension. t It is a small constant that ensures numerical stability; Step 2.3: Introduce a feature interaction mechanism between the reflection branch and the illumination branch to realize the redistribution and constraint of structural information; Step 2.4: Introduce a cross-branch feature modulation strategy: On the one hand, use illumination features to guide and enhance reflection features; on the other hand, apply structural suppression constraints to the illumination features themselves to achieve separation and redistribution of structural information, as shown in Equation (9). (9); in, CA ( ) represents the cross-attention mechanism, used to model reflection features. R With light characteristics L The dependencies between them are determined by dynamically weighting the reflection features guided by illumination characteristics, thereby enhancing the structure-related response and obtaining a clearer reflection representation of the structure. R′ ; SA ( () represents a self-attention mechanism used to model long-range dependencies within illumination features, extracting embedded structural responses from the illumination features to obtain structural residual components. L′ ; Step 2.5: After completing the feature interaction, the final reflection component and illumination component are reconstructed, and the expression is shown in Equation (10); (10); in, Convs ( ) indicates the convolution reconstruction operation.

[0011] Step 3 is implemented in the following steps: Step 3.1: Analyze the illumination components output by the latent decomposition module based on retina theory. L ^ Perform downsampling to obtain a low-resolution illumination map. L Fixed lr This serves as the initial estimate for subsequent implicit refinement; Step 3.2: Obtain a low-resolution illumination map L Fixed lr Based on this, the illumination enhancement process is modeled as a continuous residual field conditioned on spatial coordinates and local context information; for any pixel position = ( , Extract the corresponding local context neighborhood. N ( P i The neighborhood is defined as... The illumination of the k×k image patch centered on the image is defined as the sum of the initial estimate and the prediction residual, as shown in equation (11). (11); Where Θ(·) represents the position An implicit mapping function for predicting illumination residuals is used to compensate for local spatial inconsistencies while maintaining smooth changes in overall illumination. Step 3.3: The implicit mapping Θ(·) is parameterized using a sinusoidal representation network. The sinusoidal representation network is a multilayer perceptron with a sinusoidal function as the activation function, and the definition of each layer is shown in equation (12); (12); in, W This is the weight matrix. b For bias terms, sin (·) is the sinusoidal activation function; The Sigmoid activation function is used in the output layer to constrain the magnitude of the predicted residual to fall within the [0,1] interval, thereby ensuring the numerical stability of the illumination refinement process and avoiding over-enhancement or oscillation.

[0012] Step 3 is implemented in the following steps: Step 3.4: The implicit mapping network Θ(·) is generated by the local image patch encoder. F p Spatial coordinate encoder F s and output head F o It consists of three parts; among them, the local image block encoder F p With neighborhood N ( P i The input is a convolutional neural network, which extracts local contextual features; the spatial coordinate encoder... F s In pixel position P i Fourier position coding c ( Pi The input is 256 channels, which enhances the network's ability to express high-frequency location information. The hidden layer channels of both encoders are set to 256 and compressed to 128 dimensions before feature concatenation, thus controlling computational complexity while maintaining expressive power. The two feature streams are concatenated along the channel dimension and then input into the output head. F o The final residual prediction result is obtained, and the expression is shown in equation (13); (13); Where || represents the feature concatenation operation; Step 3.5: After performing illumination residual prediction at low resolution, a refined low-resolution illumination map is obtained. L Fixed lr Using guided filtering L Fixed lr Upsampling is performed, and the guide image is the brightness channel of the original low-light image. The guide filter can ensure the smoothness of the illumination transition while maintaining the structural boundary. The calculation formula is shown in Equation (14). (14); Where r is the filtering radius and ε is the regularization parameter; Step 3.6: Restore the illumination components to their original resolution. L Fixed The reflection component output by the latent decomposition module R ^ Perform element-wise multiplication to obtain the final illumination correction result, as shown in equation (15); (15).

[0013] The beneficial effects of this invention are: This invention provides a low-light image enhancement method based on retinal theory and implicit neural representation. It performs retinal theory decomposition in the latent feature space and designs a feature interaction mechanism between the reflection and illumination branches. By redistributing structural residuals from the illumination component to the reflection component, it effectively decouples the coupled illumination and structural information in low-light images. This avoids the structural content leakage problem common in traditional retinal theory methods, allowing the reflection component to retain more complete texture and edge details, while making the illumination component smoother and free of structural artifacts, thus significantly improving the realism and visual quality of the enhanced image. Implicit neural representation is used to continuously model illumination, transforming the illumination enhancement problem into a continuous residual field prediction task conditioned by spatial coordinates and local context. Compared to existing discrete pixel-level illumination modeling methods, this continuous modeling approach is highly consistent with the smooth, continuous, and spatially correlated physical characteristics of illumination in real scenes. It can effectively eliminate block artifacts and spatial discontinuities caused by discretization and supports sub-pixel-level illumination queries at arbitrary resolutions, providing a flexible solution for high-resolution low-light image enhancement. A multi-scale structural enhancement auxiliary branch, including a brightness alignment module and an adaptive fusion module, is designed. The auxiliary branch employs an encoder-decoder structure and introduces an attention mechanism to explicitly extract high-frequency structural information such as edges and textures from the original low-light image. The brightness alignment module uses the illumination correction result of the main branch as a brightness reference and performs global brightness normalization on the output of the auxiliary branch. This allows the auxiliary branch to purely provide structural compensation without interfering with the illumination modeling of the main branch, effectively compensating for the shortcomings of implicit neural expression in smoothing high-frequency details. The adaptive fusion module predicts pixel-level weight mapping based on spatial location, dynamically balancing the illumination correction of the main branch and the structural enhancement of the auxiliary branch: in well-lit areas, the main branch is prioritized to maintain natural brightness, while in extremely dark areas, the auxiliary branch is prioritized to restore structural information. Through the synergistic effect of these two modules, the final enhanced image achieves both global brightness balance and clear local details in complex low-light scenes. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the network structure of the network model in Embodiment 6 of the present invention; Figure 2 This is a schematic diagram of the potential decomposition module structure based on retinal theory in Embodiment 6 of the present invention; Figure 3 This is a schematic diagram of the enhancement effect of the light intensifier based on implicit neural expression in Embodiment 6 of the present invention; Figure 4 This is a visual comparison diagram of the low-light enhancement method in Comparative Example 1 of this invention on the LOL dataset. Detailed Implementation

[0015] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0016] The low-light image enhancement method based on retinal theory and implicit neural expression provided by this invention includes the following steps: acquiring an input low-light image; inputting the low-light image into a latent decomposition module based on retinal theory, decomposing the image into a reflection component and an illumination component in the latent feature space; inputting the decomposed illumination component into an illumination enhancer based on implicit neural expression to generate a normal illumination map with continuous and smooth changes in the generation space; recombining the normal illumination map and the reflection component to obtain an illumination-corrected image; simultaneously inputting the original low-light image into a multi-scale structure enhancement auxiliary branch to generate a structure enhancement image, and using the illumination information of the main branch to perform brightness alignment on the output of the auxiliary branch; adaptively fusing the illumination-corrected image and the brightness-aligned structure enhancement image through a lightweight fusion network to output the final enhanced image; The specific steps are as follows: Step 1: Input a low-light image I For digital images created in low-light environments, input a low-light image. I The spatial resolution is H×W; where H represents the height in pixels, W represents the width in pixels, and there are 3 color channels, corresponding to the red, green, and blue channels, i.e., I∈R. (H×W×3) Input low-light image I The pixel value range of each pixel is normalized to the interval [0,1], where 0 represents all black and 1 represents all white; Input low-light image I Any low-light image captured in nighttime surveillance, indoor low-light or backlight scenarios; Input low-light image I This includes degradation features such as insufficient brightness, low contrast, noticeable noise, and blurred details; Step 2: Input low-light image I Input a latent decomposition module based on retinal theory, which decomposes the image into reflectance components in the latent feature space. R ^ and light component L ^ ; This includes: estimating coarse illumination through channel-wise maximum operation, introducing a feature interaction mechanism between the reflection branch and the illumination branch, and redistributing structural information from the illumination component to the reflection component, thereby obtaining a structurally complete reflection component and a spatially smooth illumination component. The specific steps are as follows: Step 2.1: Inspired by the retinal theory, low-light images... I Modeled as reflection component R With light component LThe element-wise product is expressed as shown in equation (7); (7); Where ⊙ represents element-wise multiplication; R It represents the inherent content of the scene, including the inherent properties of texture, edges, and color; L This represents a smooth spatial distribution of light, under ideal conditions. R Preserve complete high-frequency details. L Maintain low-frequency, smooth, and continuous changes; Step 2.2: In order to achieve decomposition in the latent feature space, the input low-light image is... I By mapping to the feature space through convolutional layers, we obtain the feature representation F(I) ∈ R. H×W×C , where C is the total number of channels; the maximum value operation of each channel is used to extract a rough estimate of the illumination intensity from the feature response. This operation assumes that at each spatial location, the maximum response of the feature channel can reflect the dominant illumination intensity at that location, as shown in Equation (8); (8); in, F c(I) Feature representation F ( I ) c Each channel response, max c(·) This indicates taking the maximum value along the channel dimension. t It is a small constant that ensures numerical stability; by performing a channel-by-channel maximum operation on the feature channels, the illumination intensity can be estimated from the dominant, color-invariant feature response, thereby effectively capturing the most significant brightness information in the image, alleviating the problem of underestimation of illumination under low illumination conditions, and ensuring that the estimated illumination distribution is more robust and continuous. Step 2.3: Under extremely low illumination conditions, due to the extremely low signal-to-noise ratio, illumination changes and texture structures are highly coupled, and the retinal theory decomposition exhibits pathological behavior. At this time, structural responses (such as edges and fine textures) are often incorrectly absorbed into the illumination branch, resulting in incomplete reflection estimation. At the same time, unnatural structural artifacts appear in the illumination component. To solve the above-mentioned structural leakage problem, a feature interaction mechanism is introduced between the reflection branch and the illumination branch to realize the redistribution and constraint of structural information. Step 2.4: Introduce a cross-branch feature modulation strategy: On the one hand, use illumination features to guide and enhance reflection features; on the other hand, apply structural suppression constraints to the illumination features themselves to achieve separation and redistribution of structural information, as shown in Equation (9). (9); in, CA ( ) represents the cross-attention mechanism, used to model reflection features. R With light characteristics L The dependencies between them are determined by dynamically weighting the reflection features guided by illumination characteristics, thereby enhancing the structure-related response and obtaining a clearer reflection representation of the structure. R′ ; SA ( () represents a self-attention mechanism used to model long-range dependencies within illumination features, extracting embedded structural responses from the illumination features to obtain structural residual components. L′ The residual component characterizes the structural information contained in the illumination branch; Through the above mechanism, on the one hand, structural cues in the illumination branch are used to guide the optimization of the reflection branch, and on the other hand, the structural components in the illumination branch are explicitly separated to provide a basis for subsequent structural suppression. Step 2.5: After completing the feature interaction, the final reflection component and illumination component are reconstructed, and the expression is shown in Equation (10); (10); in, Convs ( () indicates the convolution reconstruction operation; By L′ join in R′ This allows for the recovery of structural information that was erroneously absorbed by the illumination branches, resulting in a more complete reflection representation; simultaneously, it allows for the extraction of information from the original illumination features. L Subtract L′ This effectively removes structural contamination, making the illumination component smoother and more consistent with physical priors; finally, the reconstructed reflection and illumination components are denoted as... R ^ and L ^ As the output of this module, it is used in subsequent image enhancement and reconstruction processes; through the above-mentioned inter-branch structural response redistribution mechanism, the retinal decomposition in the latent space is more realistic and decoupled, effectively alleviating the content leakage problem; Step 3: Decompose the obtained light components L ^ Input a light intensifier based on implicit neural expression, the intensifier for L ^ Downsampling is performed to obtain a low-resolution illumination map. L lr Using spatial coordinates and local context neighborhood as conditions, a sinusoidal representation network is employed to predict illumination residuals, generating spatially continuous and smoothly varying normal illumination maps. L Fixed lrGuided filtering is used to restore the illumination map to its original resolution, resulting in a corrected normal illumination map. L Fixed ; The specific steps are as follows: Step 3.1: Analyze the illumination components output by the latent decomposition module based on retina theory. L ^ Perform downsampling to obtain a low-resolution illumination map. L Fixed lr This serves as the initial estimate for subsequent implicit refinement. This step significantly reduces the computational complexity of implicit modeling and, by utilizing the characteristic that illumination components are predominantly low-frequency, effectively suppresses the interference of high-frequency noise on the continuous modeling process, thereby improving the stability and robustness of subsequent illumination estimation. Step 3.2: Obtain a low-resolution illumination map L Fixed lr Building upon this foundation, the illumination enhancement process is modeled as a continuous residual field conditioned by spatial coordinates and local context information. Unlike directly predicting illumination values ​​in the discrete pixel domain, this method achieves spatial consistency refinement of illumination by learning a continuous function and supports sub-pixel level queries, thereby effectively alleviating the block artifacts and discontinuities caused by discrete sampling in traditional methods. For any pixel position... = ( , Extract the corresponding local context neighborhood. N ( P i The neighborhood is defined as... The illumination of the k×k image patch centered on the image is defined as the sum of the initial estimate and the prediction residual, as shown in equation (11). (11); Where Θ(·) represents the position An implicit mapping function for predicting illumination residuals is used to compensate for local spatial inconsistencies while maintaining smooth changes in overall illumination. Step 3.3: The implicit mapping Θ(·) is parameterized using a sinusoidal representation network. The sinusoidal representation network is a multilayer perceptron with a sinusoidal function as the activation function, and the definition of each layer is shown in equation (12); (12); in, WThis is the weight matrix. b For bias terms, sin (·) is the sinusoidal activation function; By introducing a sinusoidal activation function, SiRen can effectively model spatially continuous signals with smooth first derivatives, which is particularly important for illumination modeling because drastic changes in real illumination are physically implausible. Furthermore, SiRen exhibits a significant inductive bias for low-frequency signals while maintaining good expressiveness under coordinate-based input conditions, making it well-suited for modeling smoothly varying illumination fields. The Sigmoid activation function is used in the output layer to constrain the magnitude of the predicted residual to fall within the [0,1] interval, thereby ensuring the numerical stability of the illumination refinement process and avoiding over-enhancement or oscillation. Step 3.4: The implicit mapping network Θ(·) is generated by the local image patch encoder. F p Spatial coordinate encoder F s and output head F o It consists of three parts; among them, the local image block encoder F p With neighborhood N ( P i The input is a convolutional neural network, which extracts local contextual features; the spatial coordinate encoder... F s In pixel position P i Fourier position coding c ( P i The input is 256 channels, which enhances the network's ability to express high-frequency location information. The hidden layer channels of both encoders are set to 256 and compressed to 128 dimensions before feature concatenation, thus controlling computational complexity while maintaining expressive power. The two feature streams are concatenated along the channel dimension and then input into the output head. F o The final residual prediction result is obtained, and the expression is shown in equation (13); (13); Where || represents the feature concatenation operation; Step 3.5: After performing illumination residual prediction at low resolution, a refined low-resolution illumination map is obtained. L Fixed lr To restore the original image resolution, guided filtering is used. L Fixed lrUpsampling is performed, and the guide image is the brightness channel of the original low-light image. The guide filter can ensure the smoothness of the illumination transition while maintaining the structural boundary. The calculation formula is shown in Equation (14). (14); Where r is the filtering radius and ε is the regularization parameter; Step 3.6: Restore the illumination components to their original resolution. L Fixed The reflection component output by the latent decomposition module R ^ Perform element-wise multiplication to obtain the final illumination correction result, as shown in equation (15); (15); Through the above process, we have achieved fine lighting modeling and structure preservation enhancement based on continuous implicit representation, which effectively improves the brightness consistency and visual naturalness of low-light images, while avoiding the artifact problem common in traditional discrete enhancement methods. Step 4: Extract the original low-light image I The input multi-scale structure enhancement auxiliary branch employs an encoder-decoder structure, introducing channel attention and spatial attention mechanisms at different spatial resolutions to explicitly model high-frequency structures such as edges and textures, generating structure-enhanced images. I MSA The expression is shown in equation (1); (1); Step 5: Since the auxiliary branch's primary objective is structural restoration, the output... I MSA The brightness level is often close to the original low-light input, and the overall image is too dark. This is when directly compared to the illumination-corrected image of the main branch. I Fixed Merging will reduce the brightness of the final result; therefore, a brightness alignment module is designed to use the main branch output as a brightness reference and perform global brightness normalization on the auxiliary branch output. Specifically, the luminance component of the RGB image is calculated. For any RGB image, the luminance component is obtained by weighted summation of each color channel, as shown in equation (2). (2); in, I (c) This represents the c-th color channel of the image. oh c The luminance coefficient is used; the standard luminance conversion coefficient ω is employed. r =0.299, ω g =0.587, ω b =0.114; Calculate the illumination correction images of the main branches separately. I Fixed luminance component Y Fixed Image enhancement with auxiliary branch structures I MSA luminance component Y MSA The global scaling factor s is estimated by matching the spatial average brightness of the two, as shown in equation (3). (3); in, m(·) This indicates that a spatial averaging operation is performed on the image plane, which calculates the average brightness of all pixels; e Let be a small constant, take e =10 -6 This is used to ensure numerical stability; The auxiliary branch output is rescaled as a whole using the global scaling factor to obtain the brightness-aligned structural image, as shown in equation (4). (4); Step 6: Obtain the structure enhancement result after brightness alignment. I BA Then, a lightweight fusion network is designed to process the illumination-corrected image. I Fixed Structure image aligned with brightness I BA Adaptive fusion is performed, the fusion network predicts pixel-level weight mapping, and the contributions of illumination correction and structural enhancement are dynamically balanced according to spatial location to output the final enhanced image. The mathematical expression is shown in Equation (5). (5); in, F Fuse ( By predicting pixel-level weight mappings W∈[0,1] H×W The weighted fusion of the two results is achieved, and the mathematical expression is shown in equation (6). (6).

[0017] Example 1 The low-light image enhancement method proposed in this embodiment, based on retinal theory and implicit neural expression, includes the following steps: acquiring an input low-light image; inputting the low-light image into a latent decomposition module based on retinal theory to decompose the image into a reflection component and an illumination component; inputting the decomposed illumination component into an illumination enhancer based on implicit neural expression to generate a normal illumination image; recombining the normal illumination image and the reflection component to obtain an illumination-corrected image; simultaneously inputting the original low-light image into a multi-scale structural enhancement auxiliary branch to generate a structural enhancement image, and using the illumination information of the main branch to perform brightness alignment on the output of the auxiliary branch; and adaptively fusing the illumination-corrected image and the brightness-aligned structural enhancement image through a lightweight fusion network to output the final enhanced image.

[0018] Example 2 The low-light image enhancement method proposed in this embodiment, based on retinal theory and implicit neural expression, includes the following steps: acquiring an input low-light image; inputting the low-light image into a latent decomposition module based on retinal theory to decompose the image into a reflection component and an illumination component; inputting the decomposed illumination component into an illumination enhancer based on implicit neural expression to generate a normal illumination image; recombining the normal illumination image and the reflection component to obtain an illumination-corrected image; simultaneously inputting the original low-light image into a multi-scale structure enhancement auxiliary branch to generate a structure enhancement image, and using the illumination information of the main branch to perform brightness alignment on the output of the auxiliary branch; adaptively fusing the illumination-corrected image and the brightness-aligned structure enhancement image through a lightweight fusion network to output the final enhanced image. The specific steps are as follows: Step 1: Input a low-light image I For digital images created in low-light environments, input a low-light image. I The spatial resolution is H×W; where H represents the height in pixels, W represents the width in pixels, and there are 3 color channels, corresponding to the red, green, and blue channels, i.e., I∈R. (H×W×3) Input low-light image I The pixel value range of each pixel is normalized to the interval [0,1], where 0 represents all black and 1 represents all white; Step 2: Input low-light image I Input a latent decomposition module based on retinal theory, which decomposes the image into reflectance components in the latent feature space. R ^ and light component L ^ ; Step 3: Decompose the obtained light components L ^ Input a light intensifier based on implicit neural expression, the intensifier for L ^Downsampling is performed to obtain a low-resolution illumination map. L lr Using spatial coordinates and local context neighborhood as conditions, a sinusoidal representation network is employed to predict illumination residuals, generating spatially continuous and smoothly varying normal illumination maps. L Fixed lr Guided filtering is used to restore the illumination map to its original resolution, resulting in a corrected normal illumination map. L Fixed ; Step 4: Extract the original low-light image I The input multi-scale structure enhancement auxiliary branch employs an encoder-decoder structure, introducing channel attention and spatial attention mechanisms at different spatial resolutions to explicitly model high-frequency structures such as edges and textures, generating structure-enhanced images. I MSA The expression is shown in equation (1); (1); Step 5: Design a brightness alignment module, using the main branch output as a brightness reference, and perform global brightness normalization on the auxiliary branch output; Specifically, the luminance component of the RGB image is calculated. For any RGB image, the luminance component is obtained by weighted summation of each color channel, as shown in equation (2). (2); in, I (c) This represents the c-th color channel of the image. oh c This refers to the brightness coefficient. Calculate the illumination correction images of the main branches separately. I Fixed luminance component Y Fixed Image enhancement with auxiliary branch structures I MSA luminance component Y MSA The global scaling factor s is estimated by matching the spatial average brightness of the two, as shown in equation (3). (3); in, m(·) This indicates that a spatial averaging operation is performed on the image plane, which calculates the average brightness of all pixels; e It is a small constant used to ensure numerical stability; The auxiliary branch output is rescaled as a whole using the global scaling factor to obtain the brightness-aligned structural image, as shown in equation (4). (4); Step 6: Obtain the structure enhancement result after brightness alignment. I BA Then, a lightweight fusion network is designed to process the illumination-corrected image. I Fixed Structure image aligned with brightness I BA Adaptive fusion is performed, the fusion network predicts pixel-level weight mapping, and the contributions of illumination correction and structural enhancement are dynamically balanced according to spatial location to output the final enhanced image. The mathematical expression is shown in Equation (5). (5); in, F Fuse ( By predicting pixel-level weight mappings W∈[0,1] H×W The weighted fusion of the two results is achieved, and the mathematical expression is shown in equation (6). (6).

[0019] Example 3 The low-light image enhancement method proposed in this embodiment, based on retinal theory and implicit neural expression, includes the following steps: acquiring an input low-light image; inputting the low-light image into a latent decomposition module based on retinal theory to decompose the image into a reflection component and an illumination component; inputting the decomposed illumination component into an illumination enhancer based on implicit neural expression to generate a normal illumination image; recombining the normal illumination image and the reflection component to obtain an illumination-corrected image; simultaneously inputting the original low-light image into a multi-scale structure enhancement auxiliary branch to generate a structure enhancement image, and using the illumination information of the main branch to perform brightness alignment on the output of the auxiliary branch; adaptively fusing the illumination-corrected image and the brightness-aligned structure enhancement image through a lightweight fusion network to output the final enhanced image. The specific steps are as follows: Step 1: Input a low-light image I For digital images created in low-light environments, input a low-light image. I The spatial resolution is H×W; where H represents the height in pixels, W represents the width in pixels, and there are 3 color channels, corresponding to the red, green, and blue channels, i.e., I∈R. (H×W×3) Input low-light image I The pixel value range of each pixel is normalized to the interval [0,1], where 0 represents all black and 1 represents all white; Input low-light image I Any low-light image captured in nighttime surveillance, indoor low-light or backlight scenarios; Input low-light imageI This includes degradation features such as insufficient brightness, low contrast, noticeable noise, and blurred details; Step 2: Input low-light image I Input a latent decomposition module based on retinal theory, which decomposes the image into reflectance components in the latent feature space. R ^ and light component L ^ ; Step 3: Decompose the obtained light components L ^ Input a light intensifier based on implicit neural expression, the intensifier for L ^ Downsampling is performed to obtain a low-resolution illumination map. L lr Using spatial coordinates and local context neighborhood as conditions, a sinusoidal representation network is employed to predict illumination residuals, generating spatially continuous and smoothly varying normal illumination maps. L Fixed lr Guided filtering is used to restore the illumination map to its original resolution, resulting in a corrected normal illumination map. L Fixed ; Step 4: Extract the original low-light image I The input multi-scale structure enhancement auxiliary branch employs an encoder-decoder structure, introducing channel attention and spatial attention mechanisms at different spatial resolutions to explicitly model high-frequency structures such as edges and textures, generating structure-enhanced images. I MSA The expression is shown in equation (1); (1); Step 5: Design a brightness alignment module, using the main branch output as a brightness reference, and perform global brightness normalization on the auxiliary branch output; Specifically, the luminance component of the RGB image is calculated. For any RGB image, the luminance component is obtained by weighted summation of each color channel, as shown in equation (2). (2); in, I (c) This represents the c-th color channel of the image. oh c This refers to the brightness coefficient. Calculate the illumination correction images of the main branches separately. I Fixed luminance component Y Fixed Image enhancement with auxiliary branch structures I MSAluminance component Y MSA The global scaling factor s is estimated by matching the spatial average brightness of the two, as shown in equation (3). (3); in, m(·) This indicates that a spatial averaging operation is performed on the image plane, which calculates the average brightness of all pixels; e It is a small constant used to ensure numerical stability; The auxiliary branch output is rescaled as a whole using the global scaling factor to obtain the brightness-aligned structural image, as shown in equation (4). (4); Step 6: Obtain the structure enhancement result after brightness alignment. I BA Then, a lightweight fusion network is designed to process the illumination-corrected image. I Fixed Structure image aligned with brightness I BA Adaptive fusion is performed, the fusion network predicts pixel-level weight mapping, and the contributions of illumination correction and structural enhancement are dynamically balanced according to spatial location to output the final enhanced image. The mathematical expression is shown in Equation (5). (5); in, F Fuse ( By predicting pixel-level weight mappings W∈[0,1] H×W The weighted fusion of the two results is achieved, and the mathematical expression is shown in equation (6). (6).

[0020] Example 4 The low-light image enhancement method proposed in this embodiment, based on retinal theory and implicit neural expression, includes the following steps: acquiring an input low-light image; inputting the low-light image into a latent decomposition module based on retinal theory to decompose the image into a reflection component and an illumination component; inputting the decomposed illumination component into an illumination enhancer based on implicit neural expression to generate a normal illumination image; recombining the normal illumination image and the reflection component to obtain an illumination-corrected image; simultaneously inputting the original low-light image into a multi-scale structure enhancement auxiliary branch to generate a structure enhancement image, and using the illumination information of the main branch to perform brightness alignment on the output of the auxiliary branch; adaptively fusing the illumination-corrected image and the brightness-aligned structure enhancement image through a lightweight fusion network to output the final enhanced image. The specific steps are as follows: Step 1: Input a low-light imageI For digital images created in low-light environments, input a low-light image. I The spatial resolution is H×W; where H represents the height in pixels, W represents the width in pixels, and there are 3 color channels, corresponding to the red, green, and blue channels, i.e., I∈R. (H×W×3) Input low-light image I The pixel value range of each pixel is normalized to the interval [0,1], where 0 represents all black and 1 represents all white; Input low-light image I Any low-light image captured in nighttime surveillance, indoor low-light or backlight scenarios; Input low-light image I This includes degradation features such as insufficient brightness, low contrast, noticeable noise, and blurred details; Step 2: Input low-light image I Input a latent decomposition module based on retinal theory, which decomposes the image into reflectance components in the latent feature space. R ^ and light component L ^ ; This includes: estimating coarse illumination through channel-wise maximum operation, introducing a feature interaction mechanism between the reflection branch and the illumination branch, and redistributing structural information from the illumination component to the reflection component, thereby obtaining a structurally complete reflection component and a spatially smooth illumination component. The specific steps are as follows: Step 2.1: Transfer the low-light image I Modeled as reflection component R With light component L The element-wise product is expressed as shown in equation (7); (7); Where ⊙ represents element-wise multiplication; R It represents the inherent content of the scene, including the inherent properties of texture, edges, and color; L This represents a smooth spatial distribution of light, under ideal conditions. R Preserve complete high-frequency details. L Maintain low-frequency, smooth, and continuous changes; Step 2.2: In order to achieve decomposition in the latent feature space, the input low-light image is... I By mapping to the feature space through convolutional layers, we obtain the feature representation F(I) ∈ R. H×W×C, where C is the total number of channels; the maximum value operation of each channel is used to extract a rough estimate of the illumination intensity from the feature response. This operation assumes that at each spatial location, the maximum response of the feature channel can reflect the dominant illumination intensity at that location, as shown in Equation (8); (8); in, F c(I) Feature representation F ( I ) c Each channel response, max c(·) This indicates taking the maximum value along the channel dimension. t It is a small constant that ensures numerical stability; Step 2.3: Introduce a feature interaction mechanism between the reflection branch and the illumination branch to realize the redistribution and constraint of structural information; Step 2.4: Introduce a cross-branch feature modulation strategy: On the one hand, use illumination features to guide and enhance reflection features; on the other hand, apply structural suppression constraints to the illumination features themselves to achieve separation and redistribution of structural information, as shown in Equation (9). (9); in, CA ( ) represents the cross-attention mechanism, used to model reflection features. R With light characteristics L The dependencies between them are determined by dynamically weighting the reflection features guided by illumination characteristics, thereby enhancing the structure-related response and obtaining a clearer reflection representation of the structure. R′ ; SA ( () represents a self-attention mechanism used to model long-range dependencies within illumination features, extracting embedded structural responses from the illumination features to obtain structural residual components. L′ ; Step 2.5: After completing the feature interaction, the final reflection component and illumination component are reconstructed, and the expression is shown in Equation (10); (10); in, Convs ( () indicates the convolution reconstruction operation; Step 3: Decompose the obtained light components L ^ Input a light intensifier based on implicit neural expression, the intensifier for L ^ Downsampling is performed to obtain a low-resolution illumination map. L lrUsing spatial coordinates and local context neighborhood as conditions, a sinusoidal representation network is employed to predict illumination residuals, generating spatially continuous and smoothly varying normal illumination maps. L Fixed lr Guided filtering is used to restore the illumination map to its original resolution, resulting in a corrected normal illumination map. L Fixed ; Step 4: Extract the original low-light image I The input multi-scale structure enhancement auxiliary branch employs an encoder-decoder structure, introducing channel attention and spatial attention mechanisms at different spatial resolutions to explicitly model high-frequency structures such as edges and textures, generating structure-enhanced images. I MSA The expression is shown in equation (1); (1); Step 5: Design a brightness alignment module, using the main branch output as a brightness reference, and perform global brightness normalization on the auxiliary branch output; Specifically, the luminance component of the RGB image is calculated. For any RGB image, the luminance component is obtained by weighted summation of each color channel, as shown in equation (2). (2); in, I (c) This represents the c-th color channel of the image. oh c This refers to the brightness coefficient. Calculate the illumination correction images of the main branches separately. I Fixed luminance component Y Fixed Image enhancement with auxiliary branch structures I MSA luminance component Y MSA The global scaling factor s is estimated by matching the spatial average brightness of the two, as shown in equation (3). (3); in, m(·) This indicates that a spatial averaging operation is performed on the image plane, which calculates the average brightness of all pixels; e It is a small constant used to ensure numerical stability; The auxiliary branch output is rescaled as a whole using the global scaling factor to obtain the brightness-aligned structural image, as shown in equation (4). (4); Step 6: Obtain the structure enhancement result after brightness alignment.I BA Then, a lightweight fusion network is designed to process the illumination-corrected image. I Fixed Structure image aligned with brightness I BA Adaptive fusion is performed, the fusion network predicts pixel-level weight mapping, and the contributions of illumination correction and structural enhancement are dynamically balanced according to spatial location to output the final enhanced image. The mathematical expression is shown in Equation (5). (5); in, F Fuse ( By predicting pixel-level weight mappings W∈[0,1] H×W The weighted fusion of the two results is achieved, and the mathematical expression is shown in equation (6). (6).

[0021] Example 5 The low-light image enhancement method proposed in this embodiment, based on retinal theory and implicit neural expression, includes the following steps: acquiring an input low-light image; inputting the low-light image into a latent decomposition module based on retinal theory to decompose the image into a reflection component and an illumination component; inputting the decomposed illumination component into an illumination enhancer based on implicit neural expression to generate a normal illumination image; recombining the normal illumination image and the reflection component to obtain an illumination-corrected image; simultaneously inputting the original low-light image into a multi-scale structure enhancement auxiliary branch to generate a structure enhancement image, and using the illumination information of the main branch to perform brightness alignment on the output of the auxiliary branch; adaptively fusing the illumination-corrected image and the brightness-aligned structure enhancement image through a lightweight fusion network to output the final enhanced image. The specific steps are as follows: Step 1: Input a low-light image I For digital images created in low-light environments, input a low-light image. I The spatial resolution is H×W; where H represents the height in pixels, W represents the width in pixels, and there are 3 color channels, corresponding to the red, green, and blue channels, i.e., I∈R. (H×W×3) Input low-light image I The pixel value range of each pixel is normalized to the interval [0,1], where 0 represents all black and 1 represents all white; Input low-light image I Any low-light image captured in nighttime surveillance, indoor low-light or backlight scenarios; Input low-light image I This includes degradation features such as insufficient brightness, low contrast, noticeable noise, and blurred details; Step 2: Input low-light image I Input a latent decomposition module based on retinal theory, which decomposes the image into reflectance components in the latent feature space. R ^ and light component L ^ ; This includes: estimating coarse illumination through channel-wise maximum operation, introducing a feature interaction mechanism between the reflection branch and the illumination branch, and redistributing structural information from the illumination component to the reflection component, thereby obtaining a structurally complete reflection component and a spatially smooth illumination component. The specific steps are as follows: Step 2.1: Transfer the low-light image I Modeled as reflection component R With light component L The element-wise product is expressed as shown in equation (7); (7); Where ⊙ represents element-wise multiplication; R It represents the inherent content of the scene, including the inherent properties of texture, edges, and color; L This represents a smooth spatial distribution of light, under ideal conditions. R Preserve complete high-frequency details. L Maintain low-frequency, smooth, and continuous changes; Step 2.2: In order to achieve decomposition in the latent feature space, the input low-light image is... I By mapping to the feature space through convolutional layers, we obtain the feature representation F(I) ∈ R. H×W×C , where C is the total number of channels; the maximum value operation of each channel is used to extract a rough estimate of the illumination intensity from the feature response. This operation assumes that at each spatial location, the maximum response of the feature channel can reflect the dominant illumination intensity at that location, as shown in Equation (8); (8); in, F c(I) Feature representation F ( I ) c Each channel response, max c(·) This indicates taking the maximum value along the channel dimension. t It is a small constant that ensures numerical stability; Step 2.3: Introduce a feature interaction mechanism between the reflection branch and the illumination branch to realize the redistribution and constraint of structural information; Step 2.4: Introduce a cross-branch feature modulation strategy: On the one hand, use illumination features to guide and enhance reflection features; on the other hand, apply structural suppression constraints to the illumination features themselves to achieve separation and redistribution of structural information, as shown in Equation (9). (9); in, CA ( ) represents the cross-attention mechanism, used to model reflection features. R With light characteristics L The dependencies between them are determined by dynamically weighting the reflection features guided by illumination characteristics, thereby enhancing the structure-related response and obtaining a clearer reflection representation of the structure. R′ ; SA ( () represents a self-attention mechanism used to model long-range dependencies within illumination features, extracting embedded structural responses from the illumination features to obtain structural residual components. L′ ; Step 2.5: After completing the feature interaction, the final reflection component and illumination component are reconstructed, and the expression is shown in Equation (10); (10); in, Convs ( () indicates the convolution reconstruction operation; Step 3: Decompose the obtained light components L ^ Input a light intensifier based on implicit neural expression, the intensifier for L ^ Downsampling is performed to obtain a low-resolution illumination map. L lr Using spatial coordinates and local context neighborhood as conditions, a sinusoidal representation network is employed to predict illumination residuals, generating spatially continuous and smoothly varying normal illumination maps. L Fixed lr Guided filtering is used to restore the illumination map to its original resolution, resulting in a corrected normal illumination map. L Fixed ; The specific steps are as follows: Step 3.1: Analyze the illumination components output by the latent decomposition module based on retina theory. L ^ Perform downsampling to obtain a low-resolution illumination map. L Fixed lr This serves as the initial estimate for subsequent implicit refinement; Step 3.2: Obtain a low-resolution illumination mapL Fixed lr Based on this, the illumination enhancement process is modeled as a continuous residual field conditioned on spatial coordinates and local context information; for any pixel position = ( , Extract the corresponding local context neighborhood. N ( P i The neighborhood is defined as... The illumination of the k×k image patch centered on the image is defined as the sum of the initial estimate and the prediction residual, as shown in equation (11). (11); Where Θ(·) represents the position An implicit mapping function for predicting illumination residuals is used to compensate for local spatial inconsistencies while maintaining smooth changes in overall illumination. Step 3.3: The implicit mapping Θ(·) is parameterized using a sinusoidal representation network. The sinusoidal representation network is a multilayer perceptron with a sinusoidal function as the activation function, and the definition of each layer is shown in equation (12); (12); in, W This is the weight matrix. b For bias terms, sin (·) is the sinusoidal activation function; The Sigmoid activation function is used in the output layer to constrain the magnitude of the predicted residual to fall within the [0,1] interval, thereby ensuring the numerical stability of the illumination refinement process and avoiding over-enhancement or oscillation. Step 3.4: The implicit mapping network Θ(·) is generated by the local image patch encoder. F p Spatial coordinate encoder F s and output head F o It consists of three parts; among them, the local image block encoder F p With neighborhood N ( P i The input is a convolutional neural network, which extracts local contextual features; the spatial coordinate encoder... F s In pixel positionP i Fourier position coding c ( P i The input is 256 channels, which enhances the network's ability to express high-frequency location information. The hidden layer channels of both encoders are set to 256 and compressed to 128 dimensions before feature concatenation, thus controlling computational complexity while maintaining expressive power. The two feature streams are concatenated along the channel dimension and then input into the output head. F o The final residual prediction result is obtained, and the expression is shown in equation (13); (13); Where || represents the feature concatenation operation; Step 3.5: After performing illumination residual prediction at low resolution, a refined low-resolution illumination map is obtained. L Fixed lr Using guided filtering L Fixed lr Upsampling is performed, and the guide image is the brightness channel of the original low-light image. The guide filter can ensure the smoothness of the illumination transition while maintaining the structural boundary. The calculation formula is shown in Equation (14). (14); Where r is the filtering radius and ε is the regularization parameter; Step 3.6: Restore the illumination components to their original resolution. L Fixed The reflection component output by the latent decomposition module R ^ Perform element-wise multiplication to obtain the final illumination correction result, as shown in equation (15); (15); Step 4: Extract the original low-light image I The input multi-scale structure enhancement auxiliary branch employs an encoder-decoder structure, introducing channel attention and spatial attention mechanisms at different spatial resolutions to explicitly model high-frequency structures such as edges and textures, generating structure-enhanced images. I MSA The expression is shown in equation (1); (1); Step 5: Design a brightness alignment module, using the main branch output as a brightness reference, and perform global brightness normalization on the auxiliary branch output; Specifically, the luminance component of the RGB image is calculated. For any RGB image, the luminance component is obtained by weighted summation of each color channel, as shown in equation (2). (2); in, I (c) This represents the c-th color channel of the image. oh c This refers to the brightness coefficient. Calculate the illumination correction images of the main branches separately. I Fixed luminance component Y Fixed Image enhancement with auxiliary branch structures I MSA luminance component Y MSA The global scaling factor s is estimated by matching the spatial average brightness of the two, as shown in equation (3). (3); in, m(·) This indicates that a spatial averaging operation is performed on the image plane, which calculates the average brightness of all pixels; e It is a small constant used to ensure numerical stability; The auxiliary branch output is rescaled as a whole using the global scaling factor to obtain the brightness-aligned structural image, as shown in equation (4). (4); Step 6: Obtain the structure enhancement result after brightness alignment. I BA Then, a lightweight fusion network is designed to process the illumination-corrected image. I Fixed Structure image aligned with brightness I BA Adaptive fusion is performed, the fusion network predicts pixel-level weight mapping, and the contributions of illumination correction and structural enhancement are dynamically balanced according to spatial location to output the final enhanced image. The mathematical expression is shown in Equation (5). (5); in, F Fuse ( By predicting pixel-level weight mappings W∈[0,1] H×W The weighted fusion of the two results is achieved, and the mathematical expression is shown in equation (6). (6).

[0022] Example 6 The low-light image enhancement method proposed in this embodiment, based on retinal theory and implicit neural expression, employs a network structure as follows: Figure 1 As shown, it includes the following steps: Using a low-light image I∈R from the LOL dataset H×W×3 The spatial resolution is H=400, W=600, and the three color channels correspond to red, green, and blue, respectively. Pixel values ​​are normalized to the [0,1] interval. First, the low-light image I is input into a latent decomposition module based on retinal theory, which decomposes the image into reflectance and illumination components in the latent feature space. Specifically, I is mapped to the feature space F(I)∈R through a convolutional layer. H×W×3 The total number of feature channels, C, is set to 32. Then, a coarse estimate of the illumination intensity is extracted from the feature response using a channel-by-channel maximization operation. The calculation formula is as follows: ; in, t =10 -6 It is a small constant that ensures numerical stability. Under extremely low illumination conditions, illumination changes are highly coupled with texture structure. To address the structure leakage problem, the module introduces a feature interaction mechanism between the reflection branch and the illumination branch. For example... Figure 2 As shown, a cross-attention mechanism is employed. CA (·) Modeling reflection features R With light characteristics L The dependencies between them result in a more clearly defined reflection representation. R′ Simultaneously, a self-attention mechanism is used to extract the embedded structural response from the illumination features, resulting in the structural residual component L': ; After feature interaction is completed, the reconstruction process of the reflection component and the illumination component is represented as follows: ; in, Convs (·) indicates the convolution reconstruction operation. By... L' join in R' It can recover structural information that was mistakenly absorbed by the illumination branch, from L Subtract L' This can remove structural contamination and ultimately output the reflection component. R ^ and light component L ^ ,in L ^ The space is smooth and free of structural artifacts. R ^ Preserve complete texture and edge details.

[0023] The light components obtained from the decompositionL ^ Input a light intensifier based on implicit neural expression. First, for... L ^ A downsampling process is performed with a downsampling factor of 4 to obtain a low-resolution illumination map. L lr ∈R H / 4×W / 4×3 This step reduces computational complexity on one hand, and suppresses high-frequency noise interference by leveraging the predominantly low-frequency nature of illumination on the other. For any pixel position in the low-resolution illumination map... = ( , Extract the k×k local context neighborhood centered on it. N ( P i ), where k=3. The refined illumination is defined as the sum of the initial estimate and the predicted residual, i.e. ; Where Θ(·) represents the implicit mapping function for predicting illumination residuals. This implicit mapping is parameterized using a sinusoidal representation network, which is a multilayer perceptron with a sinusoidal function as the activation function. Each layer is defined as follows: ; in, W For the weight matrix, b This is the bias term. The implicit mapping network consists of three parts: a local image patch encoder... F p With neighborhood N ( P i The spatial coordinate encoder extracts local contextual features from the input. F s In pixel position P i Fourier position coding c ( P i To enhance the high-frequency expressiveness of the input, the number of hidden layer channels in both encoders was set to 256, and compressed to 128 dimensions before feature concatenation. Subsequently, the two feature streams were concatenated along the channel dimension and input into the output head. F o The final residual prediction result is output, expressed as: ; Here, || represents the feature concatenation operation. The output layer uses the Sigmoid activation function to constrain the magnitude of the predicted residual, ensuring it falls within the [0,1] interval. After completing the illumination residual prediction at low resolution, a refined low-resolution illumination map is obtained. L Fixed lr To restore the original image resolution, guided filtering is used to re-encode the image. L Fixed lr Upsampling is performed to restore the illumination image to its original resolution. The mathematical expression for the guided filter is: ; Among them, the filter radius r =8, regularization parameter e =10 -4 Guided filtering can maintain structural boundaries while ensuring smooth illumination transitions. Finally, the illumination components are restored to their original resolution. L Fixed The reflection component output by the latent decomposition module R ^ Perform element-wise multiplication to obtain the illumination-corrected image: ; like Figure 3 As shown, the image becomes brighter and more uniform after passing through the illumination enhancer based on implicit neural expression. Simultaneously, the original low-light image I is input into a multi-scale structure enhancement auxiliary branch. This branch employs an encoder-decoder structure, introducing channel attention and spatial attention mechanisms at different spatial resolutions to explicitly model high-frequency structures such as edges and textures, generating a structure-enhanced image. I MSA Since the auxiliary branch's primary objective is structural restoration, its output... I MSA The brightness level is often close to the original low-light input, and the overall brightness is dark. If directly compared with I Fixed Fusion reduces the brightness of the final result; therefore, a brightness alignment module is designed to perform global brightness normalization on the auxiliary branch output. First, the brightness components of the RGB image are calculated. For any RGB image, the brightness components... Y I The calculation formula is: ; Where I (c) ω represents the c-th color channel of the image. c This is the luminance coefficient. In this embodiment, the standard luminance conversion coefficient ω is used. r =0.299, ω g =0.587, ω b=0.114. The three coefficients correspond to the standard luminance conversion weights of the red, green, and blue channels, respectively. The illumination-corrected images of the main branches are calculated according to formula (9). I Fixed luminance component Y Fixed Image enhancement with auxiliary branch structures I MSA luminance component Y MSA The global scaling factor is then estimated by matching the spatial average brightness of the two. ; in m (·) indicates that a spatial averaging operation is performed on the image plane, i.e., the average brightness of all pixels is calculated. e =10 -6 A small constant is used to ensure numerical stability. Finally, this global scaling factor is used to rescale the auxiliary branch output to obtain the brightness-aligned structural image. ; Structural enhancement results after obtaining brightness alignment I BA Subsequently, a lightweight fusion network was designed to process the illumination-corrected image. I Fixed Structure image aligned with brightness I BA Adaptive fusion is performed. This fusion network predicts pixel-level weight mappings W∈[0,1]. H×W Based on the contributions of spatial location-based dynamic equilibrium illumination correction and structural enhancement, the fusion process is represented as follows: ; By predicting pixel-level weight mapping W Achieve weighted fusion of the two results. Final output image. I Final This refers to the enhanced normal light image. Experimental results on the LOL dataset show that the method of this invention has the most comfortable visual performance and the most accurate color information in terms of subjective visual effect, while effectively removing noise and preserving the texture details of low-light images; in terms of objective evaluation metrics, the method of this invention achieves the highest peak signal-to-noise ratio and structural similarity value, outperforming other existing methods in both subjective and objective evaluations.

[0024] Comparative Example 1 The method of this invention has been compared with several state-of-the-art methods, including RetinexNet, ZeroDCE, SCI, RUAS, EnGAN, and LIME. Figure 4As shown, a visualization comparison of different methods is performed on the LOL dataset. These images were obtained by testing the pre-trained models provided by the corresponding methods. "Our" refers to the results obtained by the method of this invention. RetinexNet shows the most significant brightness improvement, but has a slight tendency to overexpose. ZeroDCE, SCI, RUAS, EnGAN, and LIME, while achieving brightness enhancement, all suffer from varying degrees of detail loss, insufficient contrast, overall darkness, or residual noise. In summary, our method effectively enhances brightness while clearly restoring the detailed textures of the seats, stands, and ceiling, and avoids overexposure, distortion, and uneven brightness, presenting the most natural and comfortable visual effect. Overall, our enhancement performance is significantly better than other compared algorithms.

[0025] In summary, our method offers the most comfortable visual experience and the most accurate color information, while also effectively removing noise and preserving texture details in low-light images.

[0026] Table 1 provides objective metrics for the method of this invention and various other methods on the LOL and MIT datasets.

[0027] Table 1 Summary of Datasets

[0028] The method of this invention achieves the highest PSNR and SSIM values. Therefore, the method of this invention is superior to other methods in both subjective and objective evaluation.

Claims

1. A low-light image enhancement method based on retinal theory and implicit neural expression, characterized in that, Includes the following steps: Acquire the input low-light image; The low-light image is input into a latent decomposition module based on retinal theory, which decomposes the image into a reflection component and an illumination component. The decomposed light components are input into a light intensifier based on implicit neural expression to generate a normal lighting map; The normal illumination image and the reflection component are reconstructed to obtain the illumination-corrected image; Simultaneously, the original low-light image is input into the multi-scale structure enhancement auxiliary branch to generate a structure enhancement image, and the brightness of the auxiliary branch output is aligned using the illumination information of the main branch. A lightweight fusion network is used to adaptively fuse the illumination-corrected image and the brightness-aligned structure-enhanced image to output the final enhanced image.

2. The low-light image enhancement method based on retinal theory and implicit neural expression according to claim 1, characterized in that, The specific steps are as follows: Step 1: Input a low-light image I For digital images created in low-light environments, input a low-light image. I The spatial resolution is H×W; where H represents the height in pixels, W represents the width in pixels, and there are 3 color channels, corresponding to the red, green, and blue channels, i.e., I∈R. (H×W×3) Input low-light image I The pixel value range of each pixel is normalized to the interval [0,1], where 0 represents all black and 1 represents all white; Step 2: Input low-light image I Input a latent decomposition module based on retinal theory, which decomposes the image into reflectance components in the latent feature space. R ^ and light component L ^ ; Step 3: Decompose the obtained light components L ^ Input a light intensifier based on implicit neural expression, the intensifier for L ^ Downsampling is performed to obtain a low-resolution illumination map. L lr Using spatial coordinates and local context neighborhood as conditions, a sinusoidal representation network is employed to predict illumination residuals, generating spatially continuous and smoothly varying normal illumination maps. L Fixed lr Guided filtering is used to restore the illumination map to its original resolution, resulting in a corrected normal illumination map. L Fixed ; Step 4: Extract the original low-light image I The input multi-scale structure enhancement auxiliary branch employs an encoder-decoder structure, introducing channel attention and spatial attention mechanisms at different spatial resolutions to explicitly model high-frequency structures such as edges and textures, generating structure-enhanced images. I MSA The expression is shown in equation (1); （1）； Step 5: Design a brightness alignment module, using the main branch output as a brightness reference, and perform global brightness normalization on the auxiliary branch output; Specifically, the luminance component of the RGB image is calculated. For any RGB image, the luminance component is obtained by weighted summation of each color channel, as shown in equation (2). （2）； in, I (c) This represents the c-th color channel of the image. ω c This refers to the brightness coefficient; Calculate the illumination correction images of the main branches separately. I Fixed luminance component Y Fixed Image enhancement with auxiliary branch structures I MSA luminance component Y MSA The global scaling factor s is estimated by matching the spatial average brightness of the two, as shown in equation (3). （3）； in, μ(·) This indicates that a spatial averaging operation is performed on the image plane, which calculates the average brightness of all pixels; ε It is a small constant used to ensure numerical stability; The auxiliary branch output is rescaled as a whole using the global scaling factor to obtain the brightness-aligned structural image, as shown in equation (4). （4）； Step 6: Obtain the structure enhancement result after brightness alignment. I BA Then, a lightweight fusion network is designed to process the illumination-corrected image. I Fixed Structure image aligned with brightness I BA Adaptive fusion is performed. The fusion network predicts pixel-level weight mapping and dynamically balances the contributions of illumination correction and structural enhancement based on spatial location to output the final enhanced image. The mathematical expression is shown in Equation (5). （5）； in, F Fuse ( By predicting pixel-level weight mappings W∈[0,1] H×W The weighted fusion of the two results is achieved, and the mathematical expression is shown in equation (6). （6）。 3. The low-light image enhancement method based on retinal theory and implicit neural expression according to claim 2, characterized in that, The input low-light image mentioned in step 1 I Any low-light image captured in nighttime surveillance, indoor low-light, or backlit scenarios.

4. The low-light image enhancement method based on retinal theory and implicit neural expression according to claim 2, characterized in that, The input low-light image mentioned in step 1 I This includes degradation features such as insufficient brightness, low contrast, noticeable noise, and blurred details.

5. The low-light image enhancement method based on retinal theory and implicit neural expression according to claim 2, characterized in that, Step 2 includes: estimating coarse illumination by channel-wise maximum value operation, introducing a feature interaction mechanism between the reflection branch and the illumination branch, and redistributing structural information from the illumination component to the reflection component, thereby obtaining a structurally complete reflection component and a spatially smooth illumination component.

6. The low-light image enhancement method based on retinal theory and implicit neural expression according to claim 5, characterized in that, Step 2 is implemented in the following steps: Step 2.1: Transfer the low-light image I Modeled as reflection component R With light component L The element-wise product is expressed as shown in equation (7); （7）； Where ⊙ represents element-wise multiplication; R It represents the inherent content of the scene, including the inherent properties of texture, edges, and color; L This represents a smooth spatial distribution of light, under ideal conditions. R Preserve complete high-frequency details. L Maintain low-frequency, smooth, and continuous changes; Step 2.2: In order to achieve decomposition in the latent feature space, the input low-light image is... I By mapping to the feature space through convolutional layers, we obtain the feature representation F(I)∈R. H×W×C , where C is the total number of channels; the maximum value operation of each channel is used to extract a rough estimate of the illumination intensity from the feature response. This operation assumes that at each spatial location, the maximum response of the feature channel can reflect the dominant illumination intensity at that location, as shown in Equation (8); （8）； in, F c(I) Feature representation F ( I ) c Each channel response, max c(·) This indicates taking the maximum value along the channel dimension. τ It is a small constant that ensures numerical stability; Step 2.3: Introduce a feature interaction mechanism between the reflection branch and the illumination branch to realize the redistribution and constraint of structural information; Step 2.4: Introduce a cross-branch feature modulation strategy: On the one hand, use illumination features to guide and enhance reflection features; on the other hand, apply structural suppression constraints to the illumination features themselves to achieve separation and redistribution of structural information, as shown in Equation (9). （9）； in, CA ( ) represents the cross-attention mechanism, used to model reflection features. R With light characteristics L The dependencies between them are determined by dynamically weighting the reflection features guided by illumination characteristics, thereby enhancing the structure-related response and obtaining a clearer reflection representation of the structure. R′ ; SA ( () represents a self-attention mechanism used to model long-range dependencies within illumination features, extracting embedded structural responses from the illumination features to obtain structural residual components. L′ ; Step 2.5: After completing the feature interaction, the final reflection component and illumination component are reconstructed, and the expression is shown in Equation (10); （10）； in, Convs ( ) indicates the convolution reconstruction operation.

7. The low-light image enhancement method based on retinal theory and implicit neural expression according to claim 2, characterized in that, Step 3 is implemented in the following steps: Step 3.1: Analyze the illumination components output by the latent decomposition module based on retina theory. L ^ Perform downsampling to obtain a low-resolution illumination map. L Fixed lr This serves as the initial estimate for subsequent implicit refinement; Step 3.2: Obtain a low-resolution illumination map L Fixed lr Based on this, the illumination enhancement process is modeled as a continuous residual field conditioned on spatial coordinates and local context information; for any pixel position = ( , Extract the corresponding local context neighborhood. N ( P i The neighborhood is defined as... The illumination of the k×k image patch centered on the image is defined as the sum of the initial estimate and the prediction residual, as shown in equation (11). （11）； Where Θ(·) represents the position An implicit mapping function for predicting illumination residuals is used to compensate for local spatial inconsistencies while maintaining smooth changes in overall illumination. Step 3.3: Implicit mapping Θ(·) is parameterized using a sinusoidal representation network; the sinusoidal representation network is a multilayer perceptron with a sinusoidal function as the activation function, and the definition of each layer is shown in equation (12); （12）； in, W This is the weight matrix. b For bias terms, sin (·) is the sinusoidal activation function; The Sigmoid activation function is used in the output layer to constrain the magnitude of the predicted residual to fall within the [0,1] interval, thereby ensuring the numerical stability of the illumination refinement process and avoiding over-enhancement or oscillation.

8. The low-light image enhancement method based on retinal theory and implicit neural expression according to claim 2, characterized in that, Step 3 is further implemented according to the following steps: Step 3.4: The implicit mapping network Θ(·) is generated by the local image patch encoder. F p Spatial coordinate encoder F s and output head F o It consists of three parts; among them, the local image block encoder F p With neighborhood N ( P i The input is a convolutional neural network, which extracts local contextual features; the spatial coordinate encoder... F s In pixel position P i Fourier position coding γ ( P i The input is 256, which enhances the network's ability to express high-frequency location information. The number of hidden layer channels in both encoders is set to 256, and they are compressed to 128 dimensions before feature splicing, thereby controlling the computational complexity while ensuring expressive power. The two feature streams are concatenated along the channel dimension and then input into the output head. F o The final residual prediction result is obtained, and the expression is shown in equation (13); （13）； Where || represents the feature concatenation operation; Step 3.5: After performing illumination residual prediction at low resolution, a refined low-resolution illumination map is obtained. L Fixed lr Using guided filtering L Fixed lr Upsampling is performed, and the guide image is the brightness channel of the original low-light image. The guide filter can ensure the smoothness of the illumination transition while maintaining the structural boundary. The calculation formula is shown in Equation (14). （14）； Where r is the filtering radius and ε is the regularization parameter; Step 3.6: Restore the illumination components to their original resolution. L Fixed The reflection component output by the latent decomposition module R ^ Perform element-wise multiplication to obtain the final illumination correction result, as shown in equation (15); （15）。

Images