Weak light image enhancement method fusing noise adaptive diffusion and illumination perception

Abstract

The invention discloses a weak light image enhancement method fusing noise adaptive diffusion and illumination perception, and the method comprises the steps: firstly designing a reflection-illumination decomposition module RID based on a Retinex theory, and extracting a reflectivity component and an illumination component of an input weak light image; then designing an ADRD (adaptive diffusion reflection denoising) module, performing denoising enhancement on the reflectivity component, and outputting an enhanced reflectivity image; introducing a noise estimation network, and dynamically adjusting the de-noising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity graph; designing an illumination perception enhancement module IEM, fusing an attention mechanism and a local enhancement block, enhancing the illumination image, and outputting an enhanced illumination component; and building a complete neural network framework, carrying out joint training on the provided module, finally testing the stored neural network model, and outputting an enhanced image. According to the method, the problems of detail loss, noise residue and uneven illumination of the enhanced image in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and digital image processing, and particularly relates to a low-light image enhancement method that integrates noise adaptive diffusion and illumination perception. Background Art

[0002] Images acquired in low-light environments are often accompanied by problems such as insufficient brightness, loss of details, severe noise, and color shift, which seriously affect the visual quality of the image and the robustness of subsequent visual tasks such as target detection, semantic segmentation, and autonomous driving perception. Traditional enhancement methods such as grayscale transformation and histogram equalization have certain effects in improving brightness and contrast, but they are prone to introducing artifacts, distortion, or noise amplification, and are difficult to adapt to complex non-uniform illumination distributions. Retinex theory provides an explanatory modeling framework for image enhancement by decomposing the image into two components: reflectance and illumination. However, existing Retinex-based enhancement methods generally have limitations in dealing with reflectance noise and illumination enhancement synergy. In particular, when faced with non-uniform noise distribution in reflectance, over-smoothing or detail loss is prone to occur.

[0003] In recent years, diffusion models, as a new generation of generative modeling technology, have demonstrated strong modeling capabilities and stability in tasks such as image generation and reconstruction. However, when directly applying diffusion models to low-light image enhancement, inaccurate noise estimation often leads to loss of texture detail or insufficient denoising. Therefore, a new method that combines Retinex decomposition and diffusion modeling is urgently needed. This method can effectively separate image reflectance and illumination information, while achieving noise-adaptive diffusion denoising of reflectance and structure-aware enhancement of illumination components, thereby improving the overall visual quality and structural fidelity of low-light images. Summary of the Invention

[0004] The purpose of the present invention is to provide a low-light image enhancement method that integrates noise adaptive diffusion and light perception, which solves the problems of detail loss, noise residual and uneven illumination in the enhanced image existing in the prior art.

[0005] The technical solution adopted by the present invention is a low-light image enhancement method that integrates noise adaptive diffusion and illumination perception, which is specifically implemented according to the following steps: Step 1: Design a reflection-illumination decomposition module RID based on Retinex theory to extract the reflectivity component of the input low-light image and lighting components ; Step 2: Design an adaptive diffuse reflection denoising module ADRD to denoise and enhance the reflectivity component and output the enhanced reflectivity image ; Step 3: Introduce the noise estimation network ,dynamically adjust the denoising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity map; Step 4: Design the illumination enhancement module IEM, integrate the attention mechanism with the local enhancement block, enhance the illumination image, and output the enhanced illumination component ; Step 5: Build a complete neural network framework, jointly train the proposed modules, and optimize multiple loss functions; Step 6: Test the neural network model saved in step 5 and output the enhanced image.

[0006] The present invention is also characterized in that: Step 1 is implemented as follows: Step 1.1. First, design a lightweight reflection-illumination decomposition network RID module. The overall structure of this module adopts a symmetrical U-shaped structure, which is composed of an input layer, a shallow convolution module, a downsampling module, an upsampling module and an output layer connected in sequence. The input layer is a 3×3 convolution layer, which is used to preliminarily extract the underlying features of the input image; the shallow feature extraction module consists of two consecutive convolution layers and ReLU activation functions, which are used to extract the image's texture, edge, local contrast, and color distribution, which are important visual features in low-light environments; the downsampling module uses three layers of convolution operations with a step size of 2 to compress the spatial resolution, enhance the channel dimension, and extract high-level semantic information; the upsampling module uses three layers of transposed convolution operations to restore the spatial size layer by layer, and at the same time fuse with the features of the corresponding downsampling layer through jump connections; the output layer includes two independent convolution branches, which output the reflectance image respectively. With lighting image , before network input, first from the low-light image Extract the maximum value of the three channels as the initial illumination map , and concatenated with the original RGB image in the channel dimension as the final input image of the RID module; Step 1.2: Take the low-light image in step 1.1 And the initial light map The input reflection-illumination decomposition network RID is spliced ​​in the channel dimension, and after deep decoupling of the reflection-illumination decomposition network RID, the original low-light image Decomposed into reflection components And the light component , where: reflection component Represents the inherent properties of the image, such as material, texture, and color reflection information; lighting component Representing the illumination information of the image, such as ambient brightness and lighting distribution, the decomposition process of the reflection-illumination decomposition network RID is shown in formula (1):

[0007] in, Indicates the use of RID for reflection-lighting decomposition. is the initial light map, Stands for element-wise multiplication.

[0008] Step 2 is implemented as follows: Step 2.1, the adaptive diffuse reflection denoising module ADRD is built based on the diffusion probability model architecture. It is divided into a forward diffusion process and a backward denoising process. In the forward diffusion stage, the input reflectivity image is gradually disturbed by adding Gaussian noise to simulate different degrees of noise pollution. Specifically, given an initial reflectivity image ( ) and time steps , using predefined noise scheduling parameters Perform noise sampling and addition to generate timing disturbance graphs , the process is shown in formula (2):

[0009] in, Indicates that at each time step The generated standard Gaussian noise is used to gradually perturb the original input image; is the noise retention coefficient of the current time step, and From time step 1 to The cumulative product of the retention coefficients, that is, the image is in front The proportion of the signal retained in the step, through this diffusion process, the noise perturbation of the reflectance image can be continuously modeled, providing the input state for the subsequent reverse denoising stage; Step 2.2: In the reverse denoising stage, the goal of the diffusion model is to gradually recover a clean reflectance image from the noise image sequence after forward diffusion. To this end, the low-light reflectance image obtained in step 1 is used. As a conditional input, at each time step Construct the conditional probability distribution from the current state Predict the previous image , the process is shown in formula (3): (3) in, is the conditional mean, is the variance, Represents the network at time step The predicted value of noise; Step 2.3: After introducing the noise guidance mechanism, the reverse denoising process is rewritten as follows:

[0010] in, 、 is the noise adjustment factor, ,

[0011] Step 2.4, will go through After back diffusion denoising, the enhanced reflectivity image is obtained , as the input for subsequent image fusion and reconstruction.

[0012] Step 3 is implemented as follows: Step 3.1: Build a noise estimation network , used to extract multi-scale features from the reflectance component image obtained in step 1 and estimate the noise intensity of each region, The design adopts an encoder-decoder structure. The encoder part consists of multiple convolutional layers with increasing channel number. It extracts multi-scale spatial features from the original reflectance image through gradual downsampling, which is used to model the local and global noise distribution at different resolutions. The decoder part consists of multiple upsampling layers and convolutional layers to gradually restore the spatial resolution of the image. At the same time, it uses skip connections between the corresponding layers of the encoder to fuse low-level details with high-level semantic features. Step 3.2: The features output by the encoder are input into two multi-scale convolution branches: one adopts Convolution kernel, another one uses Convolution kernels are used to capture the noise distribution characteristics in local and larger receptive fields respectively, and then use the channel attention mechanism SE Block to perform weighted fusion of multi-scale features to enhance the perception of key noise areas; Step 3.3: Input the above fused features into the decoder module for step-by-step upsampling and restoration, and pass a The convolutional layer outputs a noise intensity map with the same size as the input reflectivity map , where each pixel position corresponds to an estimated noise intensity value, reflecting the denoising amplitude that the position in the image should receive in diffusion modeling; Step 3.4: The estimated noise intensity map It is sent to the adaptive diffuse reflection denoising module ADRD in step 2 as a regulating factor to control the dynamic coefficient of each time step in the denoising process and , to achieve region-adaptive noise suppression and detail preservation.

[0013] Step 4 is implemented as follows: Step 4.1: Construct an illumination enhancement module (IEM) to dynamically enhance the illumination image obtained in step 1. The input of IEM is the illumination component image. , first through a The convolutional layer extracts initial features, which are then fed into multiple illumination enhancement blocks for local perception enhancement. Each illumination enhancement block consists of the following submodules connected in sequence: a standard convolutional layer, a ReLU activation function, a lightweight channel attention mechanism, and a depthwise separable convolutional layer for simultaneously modeling local details and illumination distribution. Step 4.2: In each illumination enhancement block, a lightweight channel attention mechanism is first introduced to perform channel recalibration on the input feature map. This mechanism extracts the response strength of each channel through global average pooling. Then, a channel attention path consisting of two fully connected layers and a ReLU activation function is used to calculate the channel weights and perform channel-level weighting on the original features. This improves the channel activation in areas with weak illumination response and enhances the brightness expression ability in low-light areas. Then, by setting the backbone residual connection path, the attention-enhanced features are added element-by-element to the original input features to achieve feature fusion; Step 4.3: Combine multiple enhancement blocks in cascade, and output feature maps through a The convolution layer is further optimized and fused, and the Sigmoid function is used to normalize it into a three-channel image, so that the output image is aligned with the original input image, and the enhanced illumination image is finally output. .

[0014] Step 5 is implemented as follows: Step 5.1: Based on the functional modules constructed in steps 1 to 4, a complete low-light image enhancement network framework is built. The network adopts an end-to-end multi-module series structure, which consists of a reflection-illumination decomposition module RID, a noise estimation network, and a noise estimation network. , the adaptive diffuse reflection denoising module ADRD and the illumination perception enhancement module IEM are connected in sequence: First, the low-light image is input Decompose through the RID module and output the reflectance image With lighting image ; Then, the reflectance image Input to the noise estimation network , generate a noise intensity map ;Will and Input the ADRD module together to complete the adaptive diffusion denoising of the reflectivity image and output the enhanced reflectivity image ; At the same time, the illumination image Input IEM module, after being processed by multiple illumination enhancement blocks, output enhanced illumination image ; Finally, we multiply and Fusion, generating the final enhanced image; Step 5.2, the loss function consists of three parts, namely 、 and , is the decomposition loss of the reflection-illumination decomposition module RID, is the loss of the adaptive diffuse reflection denoising module ADRD, is the loss of the illumination perception enhancement module IEM, the total loss function It is expressed as the following formula (5): (5) Step 5.3, calculation , the loss function consists of three parts: reconstruction loss , reflection consistency loss and lighting smoothing loss , input the illumination component of RID decomposition and the reflected component , respectively calculated 、 as well as ; The calculation formula is as follows: (6) in, 、 The input images are The reflection component and illumination component of The calculation formula is as follows: (7) in, 、 Represent the reflection components of normal light image and low light image respectively; The calculation formula is as follows:

[0015] in, Represents the gradient of the light map in the horizontal or vertical direction, and is the weight matrix, defined as the exponential decay form of the image derivative:

[0016] in, To adjust the parameters, is the original input image; Expressed as: (10) in, and To adjust the hyperparameters of the weights of different loss terms; Step 5.4, calculation , input enhanced reflection image and the reference reflection image , The definition is as follows: (11) represents the noise prediction loss, expressed as: (12) in, is the real noise, Noise predicted for the network; represents the reflection enhancement loss, which is expressed as: (13) in, is the enhanced reflection image, is the reference reflection image; Step 5.5, calculation , input enhanced illumination image and reference lighting image , The definition is as follows: (14) represents the illumination enhancement loss, expressed as: (15) in, is the enhanced illumination image, is the reference illumination image; Represents the illumination gradient loss, expressed as:

[0017] in, Represents the gradient of the light map in the horizontal or vertical direction; Step 5.6: Select a GPU with 16GB of video memory for training. During training, periodically save the model state within the set number of training rounds, including the current epoch number, network weight parameters, the internal state of the optimizer Adam, and the learning rate scheduler information to ensure the model's recoverability and stability for subsequent deployment. After training is complete, output the final converged neural network model for testing and practical application.

[0018] Step 5.7: During training, set the batch size to 2 and the image patch size to , the optimizer uses Adam, and the initial learning rate is , and adopt an exponential decay strategy, decaying the learning rate to 0.8 times the original after each iteration. The entire training process lasts for 2000 epochs. The hyperparameter weights used in the training are set to: =0.1, =0.001, =0.01.

[0019] Step 6 is implemented as follows: Load the neural network model trained and saved in step 5, feed the low-light image to be enhanced into the network as input, perform forward reasoning, output the enhanced image result, and then save the enhanced image to the specified path for subsequent evaluation and application.

[0020] The beneficial effect of the present invention is that it integrates adaptive noise diffusion with illumination perception in a low-light image enhancement method. It uses Retinex theory to decompose the input low-light image into two components: reflectance and illumination, and then performs enhancement processing on each. The reflectance component is modeled by the Adaptive Diffuse Reflectance Denoising Module (ADRD), which introduces forward perturbation and conditional reverse denoising mechanisms. Combined with the noise intensity map output by the noise estimation network, it achieves high-quality, regionally adaptive denoising to avoid detail loss. The illumination component is processed by the Illumination Enhancement Module (IEM), which uses enhancement blocks to increase brightness while maintaining local smoothness and natural transitions. During the training phase, a reconstruction loss, a reflection consistency loss, an illumination smoothness loss, and a perceptual loss are jointly used to guide the network to achieve a unified approach of structural consistency, texture fidelity, and brightness enhancement. The enhanced image exhibits natural brightness, low noise, and clear details, making it suitable for various low-light scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the overall structure of the low-light image enhancement method integrating noise adaptive diffusion and illumination perception of the present invention; Figure 2 2 is a schematic structural diagram of a reflection-illumination decomposition module RID of a low-light image enhancement method integrating noise adaptive diffusion and illumination perception according to the present invention; Figure 3 It is the noise estimation network in the adaptive diffuse reflection denoising module ADRD of the low-light image enhancement method that integrates noise adaptive diffusion and illumination perception of the present invention. Schematic diagram of the structure; Figure 4 This is a schematic diagram of the effect of the low-light image enhancement method of the present invention that integrates noise adaptive diffusion and light perception in low-light scenes. DETAILED DESCRIPTION

[0022] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] The low-light image enhancement method of the present invention integrates noise adaptive diffusion and illumination perception, and the flow chart is as follows: Figure 1 As shown, please follow the steps below: Step 1: Design a reflection-illumination decomposition module RID based on Retinex theory to extract the reflectivity component of the input low-light image and lighting components ; Combine Figure 2 , step 1 is implemented according to the following steps: Step 1.1. First, design a lightweight reflection-illumination decomposition network RID module. The overall structure of this module adopts a symmetrical U-shaped structure, which is composed of an input layer, a shallow convolution module, a downsampling module, an upsampling module and an output layer connected in sequence. The input layer is a 3×3 convolution layer, which is used to preliminarily extract the underlying features of the input image; the shallow feature extraction module consists of two consecutive convolution layers and ReLU activation functions, which are used to extract the image's texture, edge, local contrast, and color distribution, which are important visual features in low-light environments; the downsampling module uses three layers of convolution operations with a step size of 2 to compress the spatial resolution, enhance the channel dimension, and extract high-level semantic information; the upsampling module uses three layers of transposed convolution operations to restore the spatial size layer by layer, and at the same time fuse it with the features of the corresponding downsampling layer through skip connections; the output layer includes two independent convolution branches, which output the reflectance image respectively. With lighting image To improve the initial accuracy of illumination estimation, we first extract the low-light image from the network before inputting it. Extract the maximum value of the three channels as the initial illumination map , and concatenated with the original RGB image in the channel dimension as the final input image of the RID module; Step 1.2: Take the low-light image in step 1.1 And the initial light map The input reflection-illumination decomposition network RID is spliced ​​in the channel dimension, and after deep decoupling of the reflection-illumination decomposition network RID, the original low-light image Decomposed into reflection components And the light component , where: reflection component Represents the inherent properties of the image, such as material, texture, and color reflection information; lighting component Representing the illumination information of the image, such as ambient brightness and lighting distribution, the decomposition process of the reflection-illumination decomposition network RID is shown in formula (1):

[0024] in, Indicates the use of RID for reflection-lighting decomposition. is the initial light map, Stands for element-wise multiplication.

[0025] Step 2: Design an adaptive diffuse reflection denoising module ADRD to denoise and enhance the reflectivity component and output the enhanced reflectivity image ; An adaptive diffusion-based reflectance denoising module (ADRD) is designed to gradually model and denoise the reflectance components extracted in step 1. The noise generation and restoration mechanism is simulated through the forward perturbation and conditional reverse denoising process in the diffusion probability model, and regional adaptive noise suppression is performed in combination with the noise estimation map. Finally, an enhanced reflectance image with clear details and significant noise suppression effect is obtained. ; Combine Figure 1 、 Figure 3 , the step 2 is specifically implemented according to the following steps: Step 2.1, the adaptive diffuse reflection denoising module ADRD is built based on the diffusion probability model architecture. It is divided into a forward diffusion process and a backward denoising process. In the forward diffusion stage, the input reflectivity image is gradually disturbed by adding Gaussian noise to simulate different degrees of noise pollution. Specifically, given an initial reflectivity image ( ) and time steps , using predefined noise scheduling parameters Perform noise sampling and addition to generate timing disturbance graphs , the process is shown in formula (2):

[0026] in, Indicates that at each time step The generated standard Gaussian noise is used to gradually perturb the original input image; is the noise retention coefficient of the current time step, and From time step 1 to The cumulative product of the retention coefficients, that is, the image is in front The proportion of the signal retained in the step, through this diffusion process, the noise perturbation of the reflectance image can be continuously modeled, providing the input state for the subsequent reverse denoising stage; Step 2.2: In the reverse denoising stage, the goal of the diffusion model is to gradually recover a clean reflectance image from the noise image sequence after forward diffusion. To this end, the low-light reflectance image obtained in step 1 is used. As a conditional input, at each time step Construct the conditional probability distribution from the current state Predict the previous image , the process is shown in formula (3): (3) in, is the conditional mean, is the variance, Represents the network at time step The predicted value of noise; Step 2.3: To further improve the local adaptability of denoising, introduce the noise estimation network in step 3 (Introduced in step 3) Output noise intensity map (Introduced in step 3), this graph represents the noise intensity at each pixel in the reflectance image, allowing the model to perform adaptive denoising on different regions during the back-diffusion process. After introducing the noise guidance mechanism, the back-diffusion denoising process can be rewritten as follows:

[0027] in, 、 is the noise adjustment factor, ,

[0028] Step 2.4, will go through After back diffusion denoising, the enhanced reflectivity image is obtained , as the input for subsequent image fusion and reconstruction.

[0029] Step 3: Introduce the noise estimation network ,dynamically adjust the denoising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity map; Designing the noise estimation network , namely Noise Estimation Network, is used to extract multi-scale spatial features from the reflectivity component and estimate the noise intensity of each area. By constructing a multi-scale encoding-decoding structure with a channel attention mechanism, it outputs a spatially distributed noise intensity map , used to guide the diffusion denoising process in step 2 to achieve regional adaptive noise adjustment and fidelity enhancement; combined with Figure 1 , the step 3 is specifically implemented according to the following steps: Step 3.1: Build a noise estimation network , used to extract multi-scale features from the reflectance component image obtained in step 1 and estimate the noise intensity of each region, The encoder-decoder architecture is designed with multiple convolutional layers that gradually increase the number of channels. The encoder extracts multi-scale spatial features from the original reflectance map through progressive downsampling, which is used to model local and global noise distributions at different resolutions. The decoder consists of multiple upsampling layers (such as transposed convolution or upsampling + convolution) and convolutional layers to gradually restore the spatial resolution of the image. Skip connections are made between the corresponding encoder layers to fuse low-level details with high-level semantic features. Step 3.2: The features output by the encoder are input into two multi-scale convolution branches: one adopts Convolution kernel, another one uses Convolution kernels are used to capture the noise distribution characteristics within the local and larger receptive fields respectively. Then, the channel attention mechanism SE Block (Squeeze-and-Excitation block) is used to perform weighted fusion of multi-scale features to enhance the perception of key noise areas. Step 3.3: Input the above fused features into the decoder module for step-by-step upsampling and restoration, and pass a The convolutional layer outputs a noise intensity map with the same size as the input reflectivity map , where each pixel position corresponds to an estimated noise intensity value, reflecting the denoising amplitude that the position in the image should receive in diffusion modeling; Step 3.4: The estimated noise intensity map It is sent to the adaptive diffuse reflection denoising module ADRD in step 2 as a regulating factor to control the dynamic coefficient of each time step in the denoising process and , to achieve region-adaptive noise suppression and detail preservation.

[0030] Step 4: Design the illumination enhancement module IEM, integrate the attention mechanism with the local enhancement block, enhance the illumination image, and output the enhanced illumination component ; The illumination enhancement module (IEM) is designed to dynamically model and multi-scale enhance the illumination components decomposed in step 1. The module effectively improves the overall brightness of the illumination image while maintaining smooth and natural transitions of local structures by introducing multiple illumination enhancement blocks with attention mechanisms, and finally outputs the enhanced illumination image components. , used to compare with the enhanced reflectance image Jointly reconstruct the enhanced image; the step 4 is specifically implemented according to the following steps: Step 4.1: Construct an illumination enhancement module (IEM) to dynamically enhance the illumination image obtained in step 1. The input of IEM is the illumination component image. , first through a The convolutional layer extracts initial features, which are then fed into multiple illumination enhancement blocks for local perception enhancement. Each illumination enhancement block consists of the following submodules connected in sequence: a standard convolutional layer, a ReLU activation function, a lightweight channel attention mechanism, and a depthwise separable convolutional layer for simultaneously modeling local details and illumination distribution. Step 4.2: In each illumination enhancement block, a lightweight channel attention mechanism is first introduced to perform channel recalibration on the input feature map. This mechanism extracts the response strength of each channel through global average pooling. Then, a channel attention path consisting of two fully connected layers and a ReLU activation function is used to calculate the channel weights and perform channel-level weighting on the original features. This improves the channel activation in areas with weak illumination response and enhances the brightness expression ability in low-light areas. Subsequently, by setting the backbone residual connection path, the attention-enhanced features are added element-by-element to the original input features to achieve feature fusion, effectively alleviating the gradient vanishing problem caused by network depth, and enhancing information transfer during the brightness enhancement process while maintaining a natural transition; Step 4.3: Multiple enhancement blocks (usually 3) are cascaded together, and the output feature map is passed through a The convolution layer is further optimized and fused, and the Sigmoid function is used to normalize it into a three-channel image, so that the output image is aligned with the original input image, and the enhanced illumination image is finally output. .

[0031] Step 5: Build a complete neural network framework, jointly train the proposed modules, and optimize multiple loss functions; Based on the functional modules constructed in steps 1 to 4, a complete low-light image enhancement neural network framework is built, and multiple joint loss functions are designed to guide the network to achieve high-quality image decomposition, reflection enhancement, and illumination modeling. After the complete network structure is constructed, the parameters of each module are optimized through end-to-end training to ultimately obtain a network model for generating the optimal enhanced image. Step 5 is specifically implemented as follows: Step 5.1: Based on the functional modules constructed in steps 1 to 4, a complete low-light image enhancement network framework is built. The network adopts an end-to-end multi-module series structure, which consists of a reflection-illumination decomposition module RID, a noise estimation network, and a noise estimation network. , the adaptive diffuse reflection denoising module ADRD and the illumination perception enhancement module IEM are connected in sequence: First, the low-light image is input Decompose through the RID module and output the reflectance image With the lighting image ; Then, the reflectance image Input to the noise estimation network , generate a noise intensity map ;Will and Input the ADRD module together to complete the adaptive diffusion denoising of the reflectivity image and output the enhanced reflectivity image ; At the same time, the illumination image Input IEM module, after being processed by multiple illumination enhancement blocks, output enhanced illumination image ; Finally, we multiply and Fusion, generating the final enhanced image; Step 5.2, the loss function consists of three parts, namely 、 and , is the decomposition loss of the reflection-illumination decomposition module RID, is the loss of the adaptive diffuse reflection denoising module ADRD, is the loss of the illumination perception enhancement module IEM, the total loss function It is expressed as the following formula (5): (5) Step 5.3, calculation , the loss function consists of three parts: reconstruction loss , reflection consistency loss and lighting smoothing loss , input the illumination component of RID decomposition and the reflected component , respectively calculated 、 as well as ; The calculation formula is as follows: (6) in, 、 The input images are The reflection component and illumination component of The calculation formula is as follows: (7) in, 、 Represent the reflection components of normal light image and low light image respectively; The calculation formula is as follows:

[0032] in, Represents the gradient of the light map in the horizontal or vertical direction, and is the weight matrix, defined as the exponential decay form of the image derivative:

[0033] in, To adjust the parameters, is the original input image; Expressed as: (10) in, and To adjust the hyperparameters of the weights of different loss terms; Step 5.4, calculation , input enhanced reflection image and the reference reflection image , The definition is as follows: (11) represents the noise prediction loss, expressed as: (12) in, is the real noise, Noise predicted for the network; represents the reflection enhancement loss, which is expressed as: (13) in, is the enhanced reflection image, is the reference reflection image; Step 5.5, calculation , input enhanced illumination image and reference lighting image , The definition is as follows: (14) represents the illumination enhancement loss, expressed as: (15) in, is the enhanced illumination image, is the reference illumination image; Represents the illumination gradient loss, expressed as:

[0034] in, Represents the gradient of the light map in the horizontal or vertical direction; Step 5.6: Select a GPU with 16GB of video memory for training. During training, periodically save the model state within the set number of training rounds, including the current epoch number, network weight parameters, the internal state of the optimizer Adam, and the learning rate scheduler information to ensure the model's recoverability and stability for subsequent deployment. After training is complete, output the final converged neural network model for testing and practical application.

[0035] Step 5.7: During training, set the batch size to 2 and the image patch size to , the optimizer uses Adam, and the initial learning rate is , and adopt an exponential decay strategy, decaying the learning rate to 0.8 times the original after each iteration. The entire training process lasts for 2000 epochs. The hyperparameter weights used in the training are set to: =0.1, =0.001, =0.01.

[0036] Step 6: Test the neural network model saved in step 5 and output the enhanced image.

[0037] Step 6 is implemented as follows: Load the neural network model trained and saved in step 5, feed the low-light image to be enhanced into the network as input, perform forward reasoning, output the enhanced image result, and then save the enhanced image to the specified path for subsequent evaluation and application.

[0038] Example 1 The low-light image enhancement method of the present invention integrates noise adaptive diffusion and illumination perception, and the flow chart is as follows: Figure 1 As shown, please follow the steps below: Step 1: Design a reflection-illumination decomposition module RID based on Retinex theory to extract the reflectivity component of the input low-light image and lighting components ; Step 2: Design an adaptive diffuse reflection denoising module ADRD to denoise and enhance the reflectivity component and output the enhanced reflectivity image ; Step 3: Introduce the noise estimation network ,dynamically adjust the denoising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity map; Step 4: Design the illumination enhancement module IEM, integrate the attention mechanism with the local enhancement block, enhance the illumination image, and output the enhanced illumination component ; Step 5: Build a complete neural network framework, jointly train the proposed modules, and optimize multiple loss functions; Step 6: Test the neural network model saved in step 5 and output the enhanced image.

[0039] Example 2 The low-light image enhancement method of the present invention integrates noise adaptive diffusion and illumination perception, and the flow chart is as follows: Figure 1 As shown, please follow the steps below: Step 1: Design a reflection-illumination decomposition module RID based on Retinex theory to extract the reflectivity component of the input low-light image and lighting components ; Step 1 is implemented as follows: Step 1.1. First, design a lightweight reflection-illumination decomposition network RID module. The overall structure of this module adopts a symmetrical U-shaped structure, which is composed of an input layer, a shallow convolution module, a downsampling module, an upsampling module and an output layer connected in sequence. The input layer is a 3×3 convolution layer, which is used to preliminarily extract the underlying features of the input image; the shallow feature extraction module consists of two consecutive convolution layers and ReLU activation functions, which are used to extract the image's texture, edge, local contrast, and color distribution, which are important visual features in low-light environments; the downsampling module uses three layers of convolution operations with a step size of 2 to compress the spatial resolution, enhance the channel dimension, and extract high-level semantic information; the upsampling module uses three layers of transposed convolution operations to restore the spatial size layer by layer, and at the same time fuse it with the features of the corresponding downsampling layer through skip connections; the output layer includes two independent convolution branches, which output the reflectance image respectively. With lighting image To improve the initial accuracy of illumination estimation, we first extract the low-light image from the network before inputting it. Extract the maximum value of the three channels as the initial illumination map , and concatenated with the original RGB image in the channel dimension as the final input image of the RID module; Step 1.2: Take the low-light image in step 1.1 And the initial light map The input reflection-illumination decomposition network RID is spliced ​​in the channel dimension, and after deep decoupling of the reflection-illumination decomposition network RID, the original low-light image Decomposed into reflection components And the light component , where: reflection component Represents the inherent properties of the image, such as material, texture, and color reflection information; lighting component Representing the illumination information of the image, such as ambient brightness and lighting distribution, the decomposition process of the reflection-illumination decomposition network RID is shown in formula (1):

[0040] in, Indicates the use of RID for reflection-lighting decomposition. is the initial light map, Stands for element-wise multiplication.

[0041] Step 2: Design an adaptive diffuse reflection denoising module ADRD to denoise and enhance the reflectivity component and output the enhanced reflectivity image ; Step 3: Introduce the noise estimation network ,dynamically adjust the denoising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity map; Step 4: Design the illumination enhancement module IEM, integrate the attention mechanism with the local enhancement block, enhance the illumination image, and output the enhanced illumination component ; Step 5: Build a complete neural network framework, jointly train the proposed modules, and optimize multiple loss functions; Step 6: Test the neural network model saved in step 5 and output the enhanced image.

[0042] Example 3 The low-light image enhancement method of the present invention integrates noise adaptive diffusion and illumination perception, and the flow chart is as follows: Figure 1 As shown, please follow the steps below: Step 1: Design a reflection-illumination decomposition module RID based on Retinex theory to extract the reflectivity component of the input low-light image and lighting components ; Step 1 is implemented as follows: Step 1.1. First, design a lightweight reflection-illumination decomposition network RID module. The overall structure of this module adopts a symmetrical U-shaped structure, which is composed of an input layer, a shallow convolution module, a downsampling module, an upsampling module and an output layer connected in sequence. The input layer is a 3×3 convolution layer, which is used to preliminarily extract the underlying features of the input image; the shallow feature extraction module consists of two consecutive convolution layers and ReLU activation functions, which are used to extract the image's texture, edge, local contrast, and color distribution, which are important visual features in low-light environments; the downsampling module uses three layers of convolution operations with a step size of 2 to compress the spatial resolution, enhance the channel dimension, and extract high-level semantic information; the upsampling module uses three layers of transposed convolution operations to restore the spatial size layer by layer, and at the same time fuse it with the features of the corresponding downsampling layer through skip connections; the output layer includes two independent convolution branches, which output the reflectance image respectively. With lighting image To improve the initial accuracy of illumination estimation, we first extract the low-light image from the network before inputting it. Extract the maximum value of the three channels as the initial illumination map , and concatenated with the original RGB image in the channel dimension as the final input image of the RID module; Step 1.2: Take the low-light image in step 1.1 And the initial light map The input reflection-illumination decomposition network RID is spliced ​​in the channel dimension, and after deep decoupling of the reflection-illumination decomposition network RID, the original low-light image Decomposed into reflection components And the light component , where: reflection component Represents the inherent properties of the image, such as material, texture, and color reflection information; lighting component Representing the illumination information of the image, such as ambient brightness and lighting distribution, the decomposition process of the reflection-illumination decomposition network RID is shown in formula (1):

[0043] in, Indicates the use of RID for reflection-lighting decomposition. is the initial light map, Stands for element-wise multiplication.

[0044] Step 2: Design an adaptive diffuse reflection denoising module ADRD to denoise and enhance the reflectivity component and output the enhanced reflectivity image ; An adaptive diffusion-based reflectance denoising module (ADRD) is designed to gradually model and denoise the reflectance components extracted in step 1. The noise generation and restoration mechanism is simulated through the forward perturbation and conditional reverse denoising process in the diffusion probability model, and regional adaptive noise suppression is performed in combination with the noise estimation map. Finally, an enhanced reflectance image with clear details and significant noise suppression effect is obtained. ; Combine Figure 1 、 Figure 3 , the step 2 is specifically implemented according to the following steps: Step 2.1, the adaptive diffuse reflection denoising module ADRD is built based on the diffusion probability model architecture. It is divided into a forward diffusion process and a backward denoising process. In the forward diffusion stage, the input reflectivity image is gradually disturbed by adding Gaussian noise to simulate different degrees of noise pollution. Specifically, given an initial reflectivity image ( ) and time steps , using predefined noise scheduling parameters Perform noise sampling and addition to generate timing disturbance graphs , the process is shown in formula (2):

[0045] in, Indicates that at each time step The generated standard Gaussian noise is used to gradually perturb the original input image; is the noise retention coefficient of the current time step, and From time step 1 to The cumulative product of the retention coefficients, that is, the image is in front The proportion of the signal retained in the step, through this diffusion process, the noise perturbation of the reflectance image can be continuously modeled, providing the input state for the subsequent reverse denoising stage; Step 2.2: In the reverse denoising stage, the goal of the diffusion model is to gradually recover a clean reflectance image from the noise image sequence after forward diffusion. To this end, the low-light reflectance image obtained in step 1 is used. As a conditional input, at each time step Construct the conditional probability distribution from the current state Predict the previous image , the process is shown in formula (3): (3) in, is the conditional mean, is the variance, Represents the network at time step The predicted value of noise; Step 2.3: To further improve the local adaptability of denoising, introduce the noise estimation network in step 3 (Introduced in step 3) Output noise intensity map (Introduced in step 3), this graph represents the noise intensity at each pixel in the reflectance image, allowing the model to perform adaptive denoising on different regions during the back-diffusion process. After introducing the noise guidance mechanism, the back-diffusion denoising process can be rewritten as follows:

[0046] in, 、 is the noise adjustment factor, ,

[0047] Step 2.4, will go through After back diffusion denoising, the enhanced reflectivity image is obtained , as the input for subsequent image fusion and reconstruction.

[0048] Step 3: Introduce the noise estimation network ,dynamically adjust the denoising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity map; Step 4: Design the illumination enhancement module IEM, integrate the attention mechanism with the local enhancement block, enhance the illumination image, and output the enhanced illumination component ; Step 5: Build a complete neural network framework, jointly train the proposed modules, and optimize multiple loss functions; Step 6: Test the neural network model saved in step 5 and output the enhanced image.

[0049] Example 4 The low-light image enhancement method of the present invention integrates noise adaptive diffusion and illumination perception, and the flow chart is as follows: Figure 1 As shown, please follow the steps below: Step 1: Design a reflection-illumination decomposition module RID based on Retinex theory to extract the reflectivity component of the input low-light image and lighting components ; Step 1 is implemented as follows: Step 1.1. First, design a lightweight reflection-illumination decomposition network RID module. The overall structure of this module adopts a symmetrical U-shaped structure, which is composed of an input layer, a shallow convolution module, a downsampling module, an upsampling module and an output layer connected in sequence. The input layer is a 3×3 convolution layer, which is used to preliminarily extract the underlying features of the input image; the shallow feature extraction module consists of two consecutive convolution layers and ReLU activation functions, which are used to extract the image's texture, edge, local contrast, and color distribution, which are important visual features in low-light environments; the downsampling module uses three layers of convolution operations with a step size of 2 to compress the spatial resolution, enhance the channel dimension, and extract high-level semantic information; the upsampling module uses three layers of transposed convolution operations to restore the spatial size layer by layer, and at the same time fuse it with the features of the corresponding downsampling layer through skip connections; the output layer includes two independent convolution branches, which output the reflectance image respectively. With the lighting image To improve the initial accuracy of illumination estimation, we first extract the low-light image from the network before inputting it. Extract the maximum value of the three channels as the initial illumination map , and concatenated with the original RGB image in the channel dimension as the final input image of the RID module; Step 1.2: Take the low-light image in step 1.1 And the initial light map The input reflection-illumination decomposition network RID is spliced ​​in the channel dimension, and after deep decoupling of the reflection-illumination decomposition network RID, the original low-light image Decomposed into reflection components And the light component , where: reflection component Represents the inherent properties of the image, such as material, texture, and color reflection information; lighting component Representing the illumination information of the image, such as ambient brightness and lighting distribution, the decomposition process of the reflection-illumination decomposition network RID is shown in formula (1):

[0050] in, Indicates the use of RID for reflection-lighting decomposition. is the initial light map, Stands for element-wise multiplication.

[0051] Step 2: Design an adaptive diffuse reflection denoising module ADRD to denoise and enhance the reflectivity component and output the enhanced reflectivity image ; An adaptive diffusion-based reflectance denoising module (ADRD) is designed to gradually model and denoise the reflectance components extracted in step 1. The noise generation and restoration mechanism is simulated through the forward perturbation and conditional reverse denoising process in the diffusion probability model, and regional adaptive noise suppression is performed in combination with the noise estimation map. Finally, an enhanced reflectance image with clear details and significant noise suppression effect is obtained. ; Combine Figure 1 、 Figure 3 , the step 2 is specifically implemented according to the following steps: Step 2.1, the adaptive diffuse reflection denoising module ADRD is built based on the diffusion probability model architecture. It is divided into a forward diffusion process and a backward denoising process. In the forward diffusion stage, the input reflectivity image is gradually disturbed by adding Gaussian noise to simulate different degrees of noise pollution. Specifically, given an initial reflectivity image ( ) and time steps , using predefined noise scheduling parameters Perform noise sampling and addition to generate timing disturbance graphs , the process is shown in formula (2):

[0052] in, Indicates that at each time step The generated standard Gaussian noise is used to gradually perturb the original input image; is the noise retention coefficient of the current time step, and From time step 1 to The cumulative product of the retention coefficients, that is, the image is in front The proportion of the signal retained in the step, through this diffusion process, the noise perturbation of the reflectance image can be continuously modeled, providing the input state for the subsequent reverse denoising stage; Step 2.2: In the reverse denoising stage, the goal of the diffusion model is to gradually recover a clean reflectance image from the noise image sequence after forward diffusion. To this end, the low-light reflectance image obtained in step 1 is used. As a conditional input, at each time step Construct the conditional probability distribution from the current state Predict the previous image , the process is shown in formula (3): (3) in, is the conditional mean, is the variance, Represents the network at time step The predicted value of noise; Step 2.3: To further improve the local adaptability of denoising, introduce the noise estimation network in step 3 (Introduced in step 3) Output noise intensity map (Introduced in step 3), this graph represents the noise intensity at each pixel in the reflectance image, allowing the model to perform adaptive denoising on different regions during the back-diffusion process. After introducing the noise guidance mechanism, the back-diffusion denoising process can be rewritten as follows:

[0053] in, 、 is the noise adjustment factor, ,

[0054] Step 2.4, will go through After back diffusion denoising, the enhanced reflectivity image is obtained , as the input for subsequent image fusion and reconstruction.

[0055] Step 3: Introduce the noise estimation network ,dynamically adjust the denoising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity map; Designing the noise estimation network , namely Noise Estimation Network, is used to extract multi-scale spatial features from the reflectivity component and estimate the noise intensity of each area. By constructing a multi-scale encoding-decoding structure with a channel attention mechanism, it outputs a spatially distributed noise intensity map , used to guide the diffusion denoising process in step 2 to achieve regional adaptive noise adjustment and fidelity enhancement; combined with Figure 1 , the step 3 is specifically implemented according to the following steps: Step 3.1: Build a noise estimation network , used to extract multi-scale features from the reflectance component image obtained in step 1 and estimate the noise intensity of each region, The encoder-decoder architecture is designed with multiple convolutional layers that gradually increase the number of channels. The encoder extracts multi-scale spatial features from the original reflectance map through progressive downsampling, which is used to model local and global noise distributions at different resolutions. The decoder consists of multiple upsampling layers (such as transposed convolution or upsampling + convolution) and convolutional layers to gradually restore the spatial resolution of the image. Skip connections are made between the corresponding encoder layers to fuse low-level details with high-level semantic features. Step 3.2: The features output by the encoder are input into two multi-scale convolution branches: one adopts Convolution kernel, another one uses Convolution kernels are used to capture the noise distribution characteristics within the local and larger receptive fields respectively. Then, the channel attention mechanism SE Block (Squeeze-and-Excitation block) is used to perform weighted fusion of multi-scale features to enhance the perception of key noise areas. Step 3.3: Input the above fused features into the decoder module for step-by-step upsampling and restoration, and pass a The convolutional layer outputs a noise intensity map with the same size as the input reflectivity map , where each pixel position corresponds to an estimated noise intensity value, reflecting the denoising amplitude that the position in the image should receive in diffusion modeling; Step 3.4: The estimated noise intensity map It is sent to the adaptive diffuse reflection denoising module ADRD in step 2 as a regulating factor to control the dynamic coefficient of each time step in the denoising process and , to achieve region-adaptive noise suppression and detail preservation.

[0056] Step 4: Design the illumination enhancement module IEM, integrate the attention mechanism with the local enhancement block, enhance the illumination image, and output the enhanced illumination component ; Step 5: Build a complete neural network framework, jointly train the proposed modules, and optimize multiple loss functions; Step 6: Test the neural network model saved in step 5 and output the enhanced image.

[0057] Example 5 The low-light image enhancement method of the present invention integrates noise adaptive diffusion and illumination perception, and the flow chart is as follows: Figure 1 As shown, please follow the steps below: Step 1: Design a reflection-illumination decomposition module RID based on Retinex theory to extract the reflectivity component of the input low-light image and lighting components ; Step 2: Design an adaptive diffuse reflection denoising module ADRD to denoise and enhance the reflectivity component and output the enhanced reflectivity image ; Step 3: Introduce the noise estimation network ,dynamically adjust the denoising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity map; Step 4: Design the illumination enhancement module IEM, integrate the attention mechanism with the local enhancement block, enhance the illumination image, and output the enhanced illumination component ; The illumination enhancement module (IEM) is designed to dynamically model and multi-scale enhance the illumination components decomposed in step 1. The module effectively improves the overall brightness of the illumination image while maintaining smooth and natural transitions of local structures by introducing multiple illumination enhancement blocks with attention mechanisms, and finally outputs the enhanced illumination image components. , used to compare with the enhanced reflectance image Jointly reconstruct the enhanced image; the step 4 is specifically implemented according to the following steps: Step 4.1: Construct an illumination enhancement module (IEM) to dynamically enhance the illumination image obtained in step 1. The input of IEM is the illumination component image. , first through a The convolutional layer extracts initial features, which are then fed into multiple illumination enhancement blocks for local perception enhancement. Each illumination enhancement block consists of the following submodules connected in sequence: a standard convolutional layer, a ReLU activation function, a lightweight channel attention mechanism, and a depthwise separable convolutional layer for simultaneously modeling local details and illumination distribution. Step 4.2: In each illumination enhancement block, a lightweight channel attention mechanism is first introduced to perform channel recalibration on the input feature map. This mechanism extracts the response strength of each channel through global average pooling. Then, a channel attention path consisting of two fully connected layers and a ReLU activation function is used to calculate the channel weights and perform channel-level weighting on the original features. This improves the channel activation in areas with weak illumination response and enhances the brightness expression ability in low-light areas. Subsequently, by setting the backbone residual connection path, the attention-enhanced features are added element-by-element to the original input features to achieve feature fusion, effectively alleviating the gradient vanishing problem caused by network depth, and enhancing information transfer during the brightness enhancement process while maintaining a natural transition; Step 4.3: Multiple enhancement blocks (usually 3) are cascaded together, and the output feature map is passed through a The convolution layer is further optimized and fused, and the Sigmoid function is used to normalize it into a three-channel image, so that the output image is aligned with the original input image, and the enhanced illumination image is finally output. .

[0058] Step 5: Build a complete neural network framework, jointly train the proposed modules, and optimize multiple loss functions; Step 6: Test the neural network model saved in step 5 and output the enhanced image.

[0059] Example 6 like Figure 1 As shown, the embodiment of the present invention includes: a low-light image enhancement method that integrates noise adaptive diffusion and illumination perception, using a reflection-illumination decomposition module RID to decompose the input low-light image into a reflectance component and an illumination component; using a noise estimation network Generate a noise intensity map to guide the diffusion model to perform regional adaptive denoising; use the adaptive diffuse reflection denoising module ADRD to model the reflectance component, and achieve high-quality reflection enhancement through forward diffusion and conditional reverse denoising mechanism; use the illumination perception enhancement module IEM to perform multi-scale modeling of the illumination component to enhance brightness and maintain local smoothness; use element-by-element multiplication to fuse the enhanced reflection and illumination images to complete the final image reconstruction. During the training process, reconstruction loss, reflection consistency loss, illumination smoothness loss and perception loss are used to jointly supervise network learning. By dynamically adjusting the loss weights, the model is guided to achieve the optimal balance between brightness enhancement, detail preservation and structure restoration, thereby obtaining an image output with natural enhancement effect, clear details and significant noise suppression, such as Figure 4 shown.

[0060] This low-light image enhancement method integrates adaptive noise diffusion with illumination perception. It uses Retinex theory to decompose the input low-light image into two components: reflectance and illumination. These components are then enhanced separately. The reflectance component is modeled by the Adaptive Diffuse Reflectance Denoising Module (ADRD), which introduces forward perturbation and conditional backward denoising mechanisms. Combined with the noise intensity map output by the noise estimation network, this method achieves high-quality, region-adaptive denoising and avoids detail loss. The illumination component is processed by the Illumination Enhancement Module (IEM), which uses enhancement blocks to boost brightness while maintaining local smoothness and natural transitions. During the training phase, a reconstruction loss, a reflectance consistency loss, an illumination smoothness loss, and a perceptual loss are combined to guide the network to achieve a unified approach of structural consistency, texture fidelity, and brightness enhancement. The resulting enhanced image exhibits natural brightness, low noise, and sharp details, making it suitable for a variety of low-light scenarios. The enhanced image, obtained from the low-light image, boasts richer detail visibility, demonstrating practical significance.

Claims

1. A low-light image enhancement method integrating noise adaptive diffusion and illumination perception, characterized by: Please follow the steps below to implement: Step 1: Design a reflection-illumination decomposition module RID based on Retinex theory to extract the reflectivity component of the input low-light image and lighting components ; Step 2: Design an adaptive diffuse reflection denoising module ADRD to denoise and enhance the reflectivity component and output the enhanced reflectivity image ; Step 3: Introduce the noise estimation network ,dynamically adjust the denoising amplitude of the diffusion model according to the noise intensity of the region in the reflectivity map; Step 4: Design the illumination enhancement module IEM, integrate the attention mechanism with the local enhancement block, enhance the illumination image, and output the enhanced illumination component ; Step 5: Build a complete neural network framework, jointly train the proposed modules, and optimize multiple loss functions; Step 6: Test the neural network model saved in step 5 and output the enhanced image.

2. The low-light image enhancement method integrating noise adaptive diffusion and illumination perception according to claim 1, characterized in that: The step 1 is specifically implemented according to the following steps: Step 1.1: First, design a lightweight reflectance-illumination decomposition network (RID) module. This module adopts a symmetrical U-shaped structure, consisting of an input layer, a shallow convolution module, a downsampling module, an upsampling module, and an output layer. The input layer is a 3×3 convolution layer, which is used to initially extract the underlying features of the input image. The shallow feature extraction module consists of two consecutive convolution layers and a ReLU activation function, which is used to extract the image's texture, edges, local contrast, and color distribution, which are important visual features in low-light environments. The downsampling module uses a three-layer convolution operation with a step size of 2 to compress the spatial resolution, improve the channel dimension, and extract high-level semantic information; the upsampling module uses a three-layer transposed convolution operation to restore the spatial size layer by layer, and at the same time fuses the features of the corresponding downsampling layer through jump connections; the output layer includes two independent convolution branches, which output the reflectance image respectively. With lighting image , before network input, first from the low-light image Extract the maximum value of the three channels as the initial illumination map , and concatenated with the original RGB image in the channel dimension as the final input image of the RID module; Step 1.2: Take the low-light image in step 1.1 And the initial light map The input reflection-illumination decomposition network RID is spliced ​​in the channel dimension, and after deep decoupling of the reflection-illumination decomposition network RID, the original low-light image Decomposed into reflection components And the light component , where: reflection component Represents the inherent properties of the image, such as material, texture, and color reflection information; lighting component Representing the illumination information of the image, such as ambient brightness and lighting distribution, the decomposition process of the reflection-illumination decomposition network RID is shown in formula (1): in, Indicates the use of RID for reflection-lighting decomposition. is the initial light map, Stands for element-wise multiplication.

3. The low-light image enhancement method integrating noise adaptive diffusion and illumination perception according to claim 2, characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1, the adaptive diffuse reflection denoising module ADRD is built based on the diffusion probability model architecture. It is divided into a forward diffusion process and a backward denoising process. In the forward diffusion stage, the input reflectivity image is gradually disturbed by adding Gaussian noise to simulate different degrees of noise pollution. Specifically, given an initial reflectivity image ( ) and time steps , using predefined noise scheduling parameters Perform noise sampling and addition to generate timing disturbance graphs , the process is shown in formula (2): in, Indicates that at each time step The generated standard Gaussian noise is used to gradually perturb the original input image; is the noise retention coefficient of the current time step, and From time step 1 to The cumulative product of the retention coefficients, that is, the image is in front The proportion of the signal retained in the step, through this diffusion process, the noise perturbation of the reflectance image can be continuously modeled, providing the input state for the subsequent reverse denoising stage; Step 2.2: In the reverse denoising stage, the goal of the diffusion model is to gradually recover a clean reflectance image from the noise image sequence after forward diffusion. To this end, the low-light reflectance image obtained in step 1 is used. As a conditional input, at each time step Construct the conditional probability distribution from the current state Predict the previous image , the process is shown in formula (3): (3) in, is the conditional mean, is the variance, Represents the network at time step The predicted value of noise; Step 2.3: After introducing the noise guidance mechanism, the reverse denoising process is rewritten as follows: in, 、 is the noise adjustment factor, , Step 2.4, will go through After back diffusion denoising, the enhanced reflectivity image is obtained , as the input for subsequent image fusion and reconstruction.

4. The low-light image enhancement method integrating noise adaptive diffusion and illumination perception according to claim 3 is characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1: Build a noise estimation network , used to extract multi-scale features from the reflectance component image obtained in step 1 and estimate the noise intensity of each region, The encoder-decoder architecture is designed. The encoder consists of multiple convolutional layers with increasing channel counts. The encoder extracts multi-scale spatial features from the original reflectivity map by progressive downsampling, which is used to model local and global noise distribution at different resolutions. The decoder consists of multiple upsampling layers and convolutional layers, which are used to gradually restore the spatial resolution of the image. At the same time, jump connections are made between the corresponding layers of the encoder to fuse low-level details with high-level semantic features. Step 3.2: The features output by the encoder are input into two multi-scale convolution branches: one adopts Convolution kernel, another one uses Convolution kernels are used to capture the noise distribution characteristics in local and larger receptive fields respectively, and then use the channel attention mechanism SE Block to perform weighted fusion of multi-scale features to enhance the perception of key noise areas; Step 3.3: Input the above fused features into the decoder module for step-by-step upsampling and restoration, and pass a The convolutional layer outputs a noise intensity map with the same size as the input reflectivity map , where each pixel position corresponds to an estimated noise intensity value, reflecting the denoising amplitude that the position in the image should receive in diffusion modeling; Step 3.4: The estimated noise intensity map It is sent to the adaptive diffuse reflection denoising module ADRD in step 2 as a regulating factor to control the dynamic coefficient of each time step in the denoising process and , to achieve region-adaptive noise suppression and detail preservation.

5. The low-light image enhancement method integrating noise adaptive diffusion and illumination perception according to claim 4 is characterized in that: The step 4 is specifically implemented according to the following steps: Step 4.1: Construct an illumination enhancement module (IEM) to dynamically enhance the illumination image obtained in step 1. The input of IEM is the illumination component image. , first through a The convolutional layer extracts initial features, which are then fed into multiple illumination enhancement blocks for local perception enhancement. Each illumination enhancement block consists of the following submodules connected in sequence: a standard convolutional layer, a ReLU activation function, a lightweight channel attention mechanism, and a depthwise separable convolutional layer for simultaneously modeling local details and illumination distribution. Step 4.2: In each illumination enhancement block, a lightweight channel attention mechanism is first introduced to perform channel recalibration on the input feature map. This mechanism extracts the response strength of each channel through global average pooling. Then, a channel attention path consisting of two fully connected layers and a ReLU activation function is used to calculate the channel weights and perform channel-level weighting on the original features. This improves the channel activation in areas with weak illumination response and enhances the brightness expression ability in low-light areas. Then, by setting the backbone residual connection path, the attention-enhanced features are added element-by-element to the original input features to achieve feature fusion; Step 4.3: Combine multiple enhancement blocks in cascade, and output feature maps through a The convolution layer is further optimized and fused, and the Sigmoid function is used to normalize it into a three-channel image, so that the output image is aligned with the original input image, and the enhanced illumination image is finally output. .

6. The low-light image enhancement method integrating noise adaptive diffusion and illumination perception according to claim 5, characterized in that: The step 5 is specifically implemented according to the following steps: Step 5.1: Based on the functional modules constructed in steps 1 to 4, a complete low-light image enhancement network framework is built. The network adopts an end-to-end multi-module series structure, which consists of a reflection-illumination decomposition module RID, a noise estimation network, and a noise estimation network. , the adaptive diffuse reflection denoising module ADRD and the illumination perception enhancement module IEM are connected in sequence: First, the low-light image is input Decompose through the RID module and output the reflectance image With lighting image ; Then, the reflectance image Input to the noise estimation network , generate a noise intensity map ;Will and Input the ADRD module together to complete the adaptive diffusion denoising of the reflectivity image and output the enhanced reflectivity image ; At the same time, the illumination image Input IEM module, after being processed by multiple illumination enhancement blocks, output enhanced illumination image ; Finally, we multiply and Fusion, generating the final enhanced image; Step 5.2, the loss function consists of three parts, namely 、 and , is the decomposition loss of the reflection-illumination decomposition module RID, is the loss of the adaptive diffuse reflection denoising module ADRD, is the loss of the illumination perception enhancement module IEM, the total loss function It is expressed as the following formula (5): (5) Step 5.3, calculation , the loss function consists of three parts: reconstruction loss , reflection consistency loss and lighting smoothing loss , input the illumination component of RID decomposition and the reflected component , respectively calculated 、 as well as ; The calculation formula is as follows: (6) in, 、 The input images are The reflection component and illumination component of The calculation formula is as follows: (7) in, 、 Represent the reflection components of normal light image and low light image respectively; The calculation formula is as follows: in, Represents the gradient of the light map in the horizontal or vertical direction, and is the weight matrix, defined as the exponential decay form of the image derivative: in, To adjust the parameters, is the original input image; Expressed as: (10) in, and To adjust the hyperparameters of the weights of different loss terms; Step 5.4, calculation , input enhanced reflection image and the reference reflection image , The definition is as follows: (11) represents the noise prediction loss, expressed as: represents the reflection enhancement loss, which is expressed as: (13) in, is the enhanced reflection image, is the reference reflection image; Step 5.5, calculation , input enhanced illumination image and reference lighting image , The definition is as follows: (14) represents the illumination enhancement loss, expressed as: (15) in, is the enhanced illumination image, is the reference illumination image; Represents the illumination gradient loss, expressed as: (16) in, Represents the gradient of the light map in the horizontal or vertical direction; Step 5.6: Select a GPU with 16GB of video memory for training. During training, periodically save the model state within the set number of training rounds, including the current epoch number, network weight parameters, the internal state of the optimizer Adam, and the learning rate scheduler information to ensure the model's recoverability and stability for subsequent deployment. After training is complete, output the final converged neural network model for testing and practical application.

7. The low-light image enhancement method integrating noise adaptive diffusion and illumination perception according to claim 6, characterized in that: During the training process of step 5.7, the batch size is set to 2 and the image patch size is , the optimizer uses Adam, and the initial learning rate is , and adopt an exponential decay strategy, decaying the learning rate to 0.8 times the original after each iteration. The entire training process lasts for 2000 epochs. The hyperparameter weights used in the training are set to: =0.1, =0.001, =0.

01.

8. The low-light image enhancement method integrating noise adaptive diffusion and illumination perception according to claim 7, characterized in that: The step 6 is specifically implemented according to the following steps: Load the neural network model trained and saved in step 5, feed the low-light image to be enhanced into the network as input, perform forward reasoning, output the enhanced image result, and then save the enhanced image to the specified path for subsequent evaluation and application.

Images