Low-illumination image enhancement method based on Retinex theory

Abstract

The invention discloses a low-illumination image enhancement method based on a Retinex theory, and belongs to the field of low-illumination image enhancement. According to the method, on the basis of the Retinex theory, three aspects of innovations are developed: a double-branch decomposition network DecomNet is established, and precise decoupling of illumination and reflection components is realized; a diffusion model is introduced to denoise reflection components, and noise suppression and detail retention are both considered by means of a self-constraint consistency loss function; reLumenNet is constructed, an LIT module and a CBAM mechanism are fused, and illumination is regulated and controlled in a self-adaptive mode. By means of a series of innovative designs, the low-illumination image enhancement effect is greatly improved, and the problems of brightness distortion, noise interference, detail missing and the like are effectively solved. The method has high practical value in the scenes of night monitoring, automatic driving and the like, and a novel low-illumination image enhancement method is provided on the technical level.

Description

Technical Field

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

[0002] In the field of digital image processing, low-light images have a wide range of application scenarios, covering several key areas such as security monitoring, autonomous driving, and medical imaging. In security monitoring, images taken in low-light environments are widely used for security monitoring in public places, transportation hubs, and other areas; in autonomous driving, images taken at night or in low-light scenarios are crucial for the vehicle's environmental perception; in the field of medical imaging, low-light images obtained from some examinations are of great significance for disease diagnosis. However, low-light images generally have problems such as insufficient brightness, blurred details, and noise interference, which seriously affect the actual application effect in these fields. In security monitoring, low-quality images make it difficult to recognize facial features and behavioral movements of people; in autonomous driving scenarios, this may lead to vehicle errors in identifying and judging obstacles; and in medical image analysis, it is easy for doctors to miss or misdiagnose lesions.

[0003] Currently, low-light image enhancement methods are mainly divided into traditional methods and deep learning-based methods. Traditional methods include histogram equalization and enhancement methods based on Retinex theory. Histogram equalization improves contrast through image processing, but it is prone to over-enhancement or loss of detail under complex lighting conditions. Retinex-based methods, such as single-scale Retinex and multi-scale Retinex, can improve the visibility of low-light images and preserve details to a certain extent. However, during the processing process, it is difficult to fully consider the denoising requirements after enhancement, and image details are easily lost when suppressing noise. In addition, they have poor adaptability to different lighting conditions.

[0004] With the development of deep learning technology, low-light image enhancement methods based on deep learning have gradually become mainstream. These methods, trained on large-scale data, can automatically learn complex image features, improving image quality to a certain extent. For example, RetinexNet combines Retinex theory with deep neural networks, Zero-DCE employs an unsupervised curve adjustment strategy, and EnlightenGAN uses a generative adversarial network to learn a brightness mapping function. However, these deep learning methods still have significant room for improvement in noise suppression, detail preservation, and natural enhancement. For example, some methods fail to effectively separate noise from detail when enhancing images, resulting in residual noise and blurred details in the enhanced images. Furthermore, the enhanced images may also exhibit unnatural lighting effects.

[0005] Against this backdrop, developing a low-light image enhancement method that effectively suppresses noise, preserves detail, and produces natural-looking enhancements is imperative. This has significant practical implications for improving image quality in various related fields, ensuring the accuracy of security monitoring, the safety of autonomous driving, and the reliability of medical diagnosis. The DT-Retinex method proposed in this paper aims to address key challenges in low-light image enhancement, providing a more efficient and reliable solution for this field. Summary of the Invention

[0006] To address the problems of insufficient brightness, blurred details, and noise interference in low-light images, as well as the shortcomings of existing enhancement methods in denoising, detail preservation, and enhancing naturalness, this paper proposes a DT-Retinex model. This model improves image quality through three stages: image decomposition, reflection denoising, and illumination enhancement. This involves using a network structure to decompose images, using a diffusion model for denoising and improving the loss function, and combining specific modules and attention mechanisms to achieve illumination enhancement. DT-Retinex performs well on multiple datasets, effectively improving the brightness, details, and naturalness of images, meeting the practical needs of low-light image enhancement in many aspects.

[0007] The present invention provides a low-light image enhancement method based on Retinex theory, comprising the following steps:

[0008] Step (1) Select a relatively classic data set and obtain low-light images and corresponding normal-light images;

[0009] Step (2) constructing the DT-Retinex model based on the Retinex theory;

[0010] Step (3) training the constructed DT-Retinex model;

[0011] Step (4) applies the trained DT-Retinex model to low-light images on multiple validation datasets to obtain enhanced results.

[0012] The dataset production mentioned above means that the datasets used for DT-Retinex model training and testing are widely sourced and carefully screened. It mainly includes the League of Legends (LOL) (v1 and v2) dataset, which contains 500 pairs of low-light and normal-light images shot in indoor environments with a resolution of 400×600 pixels, of which 485 pairs are used for training and 15 pairs are used for testing. The LOL-v2 dataset is further subdivided. LOL-v2-real captures low-light images by adjusting the exposure time and ISO, with 689 pairs of training images and 100 pairs of test images; LOL-v2-synthetic is synthesized based on the illumination distribution of the RAW image.

[0013] In addition, several advanced low-light image enhancement (LLIE) datasets, such as DICM, LIME, MEF, NPE, and VV, were selected for visual comparison experiments. Furthermore, to verify the model's effectiveness in real-world scenarios, the real-world LSRW dataset was used. This dataset consists of images captured in real-world scenes with a Nikon D7500 camera and a Huawei P40Pro phone. These datasets cover a wide range of scenes, lighting conditions, and image features, providing rich and diverse data support for the training, testing, and performance evaluation of the DT-Retinex model.

[0014] The low-light image enhancement model DT-Retinex includes: DecomNet adopts a dual-branch architecture and uses a unique design to optimize the extraction process of image components

[0015] In low-light image enhancement scenarios, traditional methods often struggle to effectively separate illumination and reflection components when dealing with varying brightness levels and complex lighting conditions, and can also lose underlying image structural features. DecomNet, on the other hand, uses a low-light image (Slow) and a normal-light image (Snormal) as input pairs and uses a weighted dual-branch architecture to extract the reflection component R and the illumination component I, respectively.

[0016] The reflection component branch constructs a feature pyramid using five cascaded 3×3 convolutional layers, each followed by a ReLU activation function. This process helps gradually extract image features. Furthermore, a cross-layer channel concatenation strategy is employed, such as fusing features from layers 2-6 with those from layers 3-5, to preserve high-frequency details. This allows the reflection component to more accurately represent the inherent material and texture characteristics of the object. Finally, after a 3×3 convolutional layer and a sigmoid activation function, the reflection component R is output.

[0017] The illumination component branch extracts basic features through a 3×3 convolutional layer and concatenates them with the features from the sixth layer of the reflectance branch. Next, two sets of 3×3 convolutions and ReLU modules are used to effectively extract low-frequency information, ultimately outputting the illumination component I, which accurately describes the impact of external lighting conditions on the image.

[0018] In the process of optimizing image decomposition, DecomNet uses an overall loss function that includes reconstruction loss, reflection consistency loss, illumination smoothness loss, and lighting consistency loss. The reconstruction loss ensures that the decomposed reflection and illumination components can effectively reconstruct the original image; the reflection consistency loss ensures that the reflection components of low-light images and normal-light images are consistent; the illumination smoothness loss maintains the smoothness of image illumination information to avoid loss of details; and the lighting consistency loss maintains the consistency of illumination information between low-light images and normal-light images. By adjusting the weights of these loss terms, DecomNet can achieve more accurate image decomposition, providing high-quality component data for subsequent reflection denoising and illumination enhancement, and improving the overall low-light image enhancement effect.

[0019] The low-light image enhancement model DT-Retinex includes: introducing a diffusion model and self-constrained consistency loss in the RDnoiseNet module to solve the problem of easy loss of details and difficulty in suppressing noise when denoising the reflective component of low-light images.

[0020] In low-light image denoising scenarios, traditional denoising methods often focus solely on single-dimensional features, making it difficult to capture the complex relationship between noise and detail in the image, resulting in a loss of detail or residual noise after denoising. The diffusion model, however, proposes a progressive denoising concept. It uses forward diffusion to gradually add noise to the reflectance component, and then uses backward diffusion to recover a clear reflectance component from the noise, thus solving the problem of separating noise and detail processing. It introduces cross-stage interactions in the denoising process, fully capturing the feature dependencies of the reflectance component at different noise levels.

[0021] The diffusion model designs a time step scheduling mechanism to transform the reflection component from the original state to the pure noise state by gradually increasing the noise level. low and highlight image I high Perform diffusion modeling and use their combination x0=R low ⊙L high As input data. The input data is processed by forward diffusion and gradually generates a noise image x through the diffusion process. T To complete this process, a predefined variance schedule {β1,β2,...,β T}, so that the input image x0 is gradually converted into noise at each time step. The formula is expressed as: Among them, x t Represents the noise data at time t. By adjusting the time step, the input image x0 can be guided to gradually evolve into noise x t Backward diffusion: After the forward diffusion is completed, the goal is to restore the clear image through the reverse denoising process. Using the editing and data synthesis capabilities of the conditional diffusion model, the random noise Start by gradually denoising to get the final clear image This process depends on the characteristics of the input image This ensures high fidelity of the recovery process. The mathematical formula for this process is:

[0022]

[0023] Unlike other denoising methods, the diffusion model extracts features and performs denoising operations at varying noise levels, effectively integrating the features of the reflective component under varying noise conditions and capturing both noise and detail at different scales. This comprehensive denoising strategy significantly enhances the network's ability to preserve reflective component detail and effectively handles complex noise and subtle details in low-light images.

[0024] At the same time, in order to ensure that the denoised reflection component is consistent with the original reflection component, a self-constrained consistency loss is introduced. The improved denoising loss function is: This loss function effectively smooths the impact of abnormal samples by introducing the expected operation, improving the stability and generalization ability of training. The self-constrained consistency loss is defined as The total loss function is L = L diff +λ1L SCC By balancing these two loss terms, the learning process of the denoising network is constrained, unnecessary bias is avoided during denoising, and the denoising effect is further improved.

[0025] The low-light image enhancement model DT-Retinex includes ReLumenNet as the core component of the DT-Retinex model. Through the deep integration of the encoder-decoder framework and the attention mechanism, it achieves adaptive illumination correction and detail enhancement for low-light images. Its core design is as follows:

[0026] The network architecture adopts a U-Net structure. The encoder consists of four groups of downsampling convolutional modules, each of which includes a 3×3 convolutional layer (stride 2), instance normalization, and a LeakyReLU activation function. The number of channels increases from 64 to 512, gradually extracting multi-level features. The decoder uses a transposed convolutional layer (stride 2) to reduce the number of channels to 64 and uses skip connections to fuse encoder features to restore the image's spatial resolution. Finally, a 3×3 convolution with a sigmoid activation function is used to generate an enhanced image, with pixel values ​​normalized to the range [0, 1].

[0027] The LIT module embeds position encoding in the Transformer calculation, the formula is X′=X+P, which enhances the model's ability to perceive the pixel spatial position. Next, the input feature X′ is split into k sub-blocks, denoted as X=[X1,X2,...,X k ], the size of each sub-block is HW×dk , where d k is the feature dimension of each attention head. Each sub-block is mapped to the query Q through the fully connected (FC) layer i 、Key(Key)K i Sum value (Value) V i , these matrices have dimensions HW×C and are calculated using the following formula:

[0028] Q i =X i W Q ,K i =X i W K ,V i =X i W V

[0029] Among them, W Q 、W K and W V is the learnable weight matrix and C is the dimension of the feature.

[0030] When calculating the relevance between the query and the key, the score is obtained by calculating their dot product and scaling. Specifically, for each head i, the attention score α is calculated i It is given by the following formula:

[0031]

[0032] Then, the scores are normalized by the softmax function:

[0033] α i =softmax(α i )

[0034] Finally, the attention score and value matrix V i Multiply them together to get the weighted output features:

[0035] Attention(Q i ,K i ,V i )=α i V i

[0036] Subsequently, the calculated enhanced features are further extracted through the feedforward network (FFN) and optimized through layer normalization (LN) and residual connection to ensure the stability of training and the continuity of information flow:

[0037] Output = LN(MSA(X′)+X′)

[0038] This design effectively ensures the stable flow of information in deep networks through residual connections and layer normalization, avoids the gradient vanishing problem, and accelerates the model training process.

[0039] LIT dynamically introduces illumination compensation during the Transformer calculation process, thereby enhancing the feature expression of low-light areas and reducing artifacts and over-enhancement problems. In addition, LIT uses local self-attention calculation to optimize the computational complexity to:

[0040] O(MSA)=2HWC 2 / k

[0041] Where k is the number of heads. Compared with the traditional global self-attention O((HW) 2 C), the computational efficiency of LIT is significantly improved, making it more suitable for high-resolution image processing.

[0042] CBAM enhances features through the synergy of channel attention and spatial attention. Channel attention generates weights through global pooling and MLP, and spatial attention generates spatial information through convolution. The final feature weighting formulas are:

[0043] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))),

[0044] M s (F′)=σ(Conv([AvgPool(F′);MaxPool(F′)]))

[0045] This module achieves a dynamic balance between brightness enhancement and detail preservation in low-light image enhancement by synergizing global illumination modeling with CBAM's local feature focus. Its innovative design effectively addresses the limitations of traditional methods in terms of naturalness and detail preservation, providing a new solution for low-light image enhancement.

[0046] Training the DT-Retinex model involves randomly cropping the input image to 96×96 pixels to improve training efficiency. Training was performed on an NVIDIA 4090 GPU using the PyTorch framework, Python version 3.8, and Cuda version 11.3. A phased training strategy was used, using the Adam optimizer with default parameters and a fixed learning rate of 0.0001. Training was repeated 200 times with a batch size of 128.

[0047] The results of the verification on real datasets refer to: model robustness and generalization evaluation on multiple datasets, and a comprehensive evaluation on the real dataset LSRW (including low-light images taken by Huawei P40Pro mobile phone and Nikon D7500 camera).

[0048] Compared with the existing technology, the above technical solution has the following beneficial effects:

[0049] (1) For low-light image enhancement tasks, DT-Retinex significantly improves performance through three-stage collaborative optimization. Image decomposition optimization: Using a dual-branch architecture and a cross-layer feature fusion strategy (DecomNet), the illumination component and the reflection component are accurately separated, retaining high-frequency details and low-frequency illumination characteristics, and solving the problems of residual noise and detail loss caused by component coupling in traditional methods. Diffusion denoising technology: The diffusion model (RDnoiseNet is introduced to perform progressive denoising on the reflection component, combined with an improved denoising loss function, to suppress noise while retaining complex texture structures. Adaptive illumination enhancement: The LIT module (global illumination modeling) and the CBAM attention mechanism (local feature enhancement) are integrated to achieve adaptive illumination adjustment and key area enhancement in ReLumenNet, solving the overexposure and artifact problems of methods such as EnlightenGAN.

[0050] (2) DT-Retinex effectively achieves multi-dimensional enhancement of low-light images, demonstrating excellent robustness, detail preservation, and natural enhancement, providing an innovative solution. Future research can further improve the model's performance in extreme low-light conditions and enhance its adaptability to lighting changes and noise. Combining advanced deep learning techniques to improve detail preservation and enhancement quality is also an important direction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is an overall flow chart of a low-light image enhancement method based on Retinex theory provided in an embodiment of the present application;

[0052] Figure 2 This is the overall structure diagram of DT-Retinex in the embodiment of this application;

[0053] Figure 3 It is the workflow diagram of the diffusion model in this application;

[0054] Figure 4 This is the ReLumenNet network structure diagram in this application;

[0055] Figure 5 This is the LIT module structure diagram of this application;

[0056] Figure 6This is the low-light enhancement result of this application. DETAILED DESCRIPTION

[0057] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. The examples provided are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments without creative effort are within the scope of protection of the present invention.

[0058] See also Figure 1 A low-light image enhancement method based on Retinex theory provided in an embodiment of the present invention specifically includes the following steps:

[0059] Step (1): Select a suitable dataset.

[0060] The present invention constructs a multi-dimensional dataset containing synthetic and real scenes, of which LOLv1 and v2 are used for model training, and other datasets are used for testing. The core training set is LOL, which includes two versions: LOLv1 consists of 500 pairs of indoor low-light and normal-light images (485 pairs for training and 15 pairs for testing); LOLv2 is divided into real scenes (LOLv2-real, 689 pairs for training and 100 pairs for testing) and synthetic data (LOLv2-syn, synthesized based on the illumination distribution of RAW images). The test set introduces benchmark datasets such as DICM, LIME, MEF, NPE, and VV, covering applications in multiple fields such as medical imaging, natural scenes, multi-exposure fusion, night environments, and vehicle vision.

[0061] Data screening focuses on clear samples from varying angles, distances, and lighting conditions to ensure coverage of the complexity of real-world scenarios. LOLv2-syn's synthetic design significantly improves the model's adaptability to complex environments by simulating extreme low light and dynamic lighting changes. All datasets combine real-world acquisition with synthetic data generation techniques, breaking through the limitations of traditional datasets with single lighting conditions and providing a comprehensive and reliable training and evaluation foundation for low-light image enhancement algorithms. The test dataset is used to verify the model's generalization capabilities and ensure the algorithm's robustness in both professional fields and everyday scenarios.

[0062] Step (2): Design the DT-Retinex model based on the Retinex theory.

[0063] First, the decomposition network DecomNet of the present invention realizes the decoupling of image components based on a dual-branch architecture

[22] . The network takes the low-light image Slow and the normal-light image Snormal as input pairs, and extracts the reflectance component R and the illumination component I respectively through a dual-branch structure with shared weights. The reflectance component branch constructs a feature pyramid by five groups of cascaded 3×3 convolutional layers (each layer is followed by ReLU activation), and retains high-frequency details through a cross-layer channel splicing strategy. Finally, after a 3×3 convolutional layer and a Sigmoid activation function, the reflectance component R is output. The illumination component branch extracts basic features through a 3×3 convolutional layer and splices them with the features of the 6th layer of the reflectance branch. Then, after two groups of 3×3 convolution and ReLU modules, low-frequency information is extracted and the illumination component I is finally output. After feature extraction, the network maps the features to the reflectance and illumination components through a reconstruction layer. The reflectance component mainly represents the inherent characteristics of the image that do not change with illumination, while the illumination component describes the influence of external lighting conditions on the image.

[0064] DecomNet uses multiple loss functions to optimize the image decomposition process. The overall loss function consists of three parts: reconstruction loss, reflection consistency loss, illumination smoothness loss, and lighting consistency loss.

[0065] Secondly, a diffusion model is introduced to denoise the reflection component, effectively removing noise and improving the quality of the restored image. Compared with traditional denoising methods, the diffusion model can retain image details and structural information while denoising. Through this method, the quality of the reflection component is significantly improved, noise is effectively suppressed, and the image restoration result is clearer and more natural. Its working principle is as follows Figure 2 shown.

[0066] The forward diffusion process first gradually contaminates the input low-light reflection component into a noisy image. Specifically, starting with the original clear image, a small amount of Gaussian noise is added at each time step. The amount of noise is dynamically controlled by a preset variance scheduling strategy. As the time step increases, the noise gradually accumulates, ultimately converting the original image into a pure noise signal. This process simulates the degradation of signals in a natural environment and provides training data for reverse denoising.

[0067] The reverse denoising process starts with a pure noise image and gradually recovers the clear reflection component through a neural network. The network learns the relationship between noise distribution and image content, predicting the clear image corresponding to the current noisy image at each time step. This prediction process is based on the inverse of forward diffusion, gradually removing noise and recovering details by adjusting the noise intensity and direction. This process uses a conditional generation strategy and incorporates the characteristic information of low-light images to ensure high fidelity of the restored results.

[0068] Loss function design: Denoising loss calculates the difference between the noise predicted by the network and the actual noise, guiding the network to learn accurate noise estimation capabilities. Self-constrained consistency loss: Ensures the structural consistency of the denoised image and the original low-light image through feature matching, avoiding detail loss caused by excessive denoising.

[0069] This paper proposes the ReLumenNet model. The core structure of this model adopts an encoder-decoder framework, combined with the LIT and CBAM modules to form an architecture for brightness enhancement.

[0070] ReLumenNet is an important module for low-light image enhancement. Its working principle involves multi-faceted collaboration to achieve high-quality image enhancement effects.

[0071] First, its overall architecture utilizes an encoder-decoder structure. The encoder consists of four groups of downsampling convolutional modules. After the input low-light image enters the encoder, it is processed by modules consisting of 3×3 convolutions, instance normalization, and LeakyReLU activation functions. During this process, the spatial size of the image is gradually reduced, while the number of channels increases from 64 to 512, thereby extracting multi-level features, including global image structure information and local texture details. The decoder restores the spatial resolution of the image through transposed convolutional layers, reducing the number of channels from 512 to 64. At the same time, skip connections are used to fuse the features of the corresponding encoder layers, integrating high-level semantic information with low-level detail information. Finally, a 3×3 convolution and sigmoid activation function are performed to output the enhanced image.

[0072] The LIT module plays a key role in illumination enhancement. It reshapes the feature map output by the encoder into a sequence, adds positional encoding to preserve spatial information, and then divides the image into local blocks. When calculating the value tensor, it incorporates the global average pooling features of the low-light component to form a conditional value. This allows for dynamic feature adjustment based on the image's illumination conditions, guiding the network to focus on low-light areas. Furthermore, it employs a local self-attention mechanism, performing attention calculations only within the local block, significantly reducing computational complexity and improving processing efficiency while ensuring accurate illumination modeling. The CBAM attention mechanism optimizes features along both channel and spatial dimensions. Channel attention first performs global average pooling and max pooling on the feature map to obtain channel-level statistics. A multi-layer perceptron is then used to generate channel weights to highlight important channels. Spatial attention then performs global pooling on the feature map processed by channel attention and generates spatial weights through convolution to focus on key areas. Finally, the channel and spatial weights are element-wise multiplied with the original feature map to achieve feature enhancement.

[0073] Step (3): Train the DT-Retinex model.

[0074] The low-light image enhancement method based on Retinex theory according to claim 1 is characterized in that the DT-Retinex model obtained after training includes: randomly cropping the input image to 96×96 pixels to improve training efficiency. Training is performed using an NVIDIA 4090 GPU based on the PyTorch framework, using Python version 3.8 and Cuda version 11.3. A phased training strategy is adopted, using the Adam optimizer with default parameter settings and a fixed learning rate of 0.0001. Training is performed for a total of 200 iterations, with a batch size of 128. To evaluate algorithm performance, paired datasets typically use established metrics such as peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), and learned perceptual image patch similarity (LPIPS) to quantitatively compare enhanced images with their corresponding ground truth reference images. Unpaired datasets use natural image quality evaluation (NIQE) and the no-reference image spatial quality evaluation (BRISQUE) as evaluation methods.

[0075] PSNR measures the ratio between the maximum possible power of a signal and the noise power present in the signal. It is often used to evaluate image quality. The formula for PSNR is:

[0076]

[0077] Among them, MAX I represents the maximum pixel value of the image (e.g., 255 for an 8-bit image), and MSE is the mean squared error between the enhanced image and the ground truth reference image.

[0078] SSIM is a metric that measures the similarity between two images. It takes into account the brightness, contrast, and structure of the images. The formula for SSIM is:

[0079]

[0080] Where x and y represent the enhanced reference image and the true reference image respectively. x and μ y is the average of x and y, and is the variance of x and y, σ xy is the covariance of x and y, and c1 and c2 are small constants that stabilize the division.

[0081] LPIPS is a perceptual similarity metric that measures the similarity between two images based on human perception. It takes into account color, texture, and structure to capture the perceptual differences between images and is calculated by comparing image patches extracted from two images using a deep neural network.

[0082] NIQE is an objective evaluation metric that extracts features from natural images to test the test image. These features are fitted into a multivariate Gaussian model. The formula is:

[0083]

[0084] x is the input image, f(x) is the image's eigenvector (natural image statistical features extracted from the image). μ is the mean of the eigenvector, which comes from the training set of natural images. Σ is the covariance matrix of the eigenvector, which comes from the training set of natural images. Σ -1 is the inverse of the covariance matrix.

[0085] BRISQUE measures the perceptual quality of images without requiring a high-quality reference image. Based on Natural Image Statistics (NSS), it analyzes the image's local contrast, brightness normalization properties, and statistical distribution to assess the degree of image distortion, such as blur, noise, and compression artifacts.

[0086] The DT-Retinex model was then trained using the aforementioned training parameters. The training results were compared across evaluation metrics, taking all indicators into consideration. Based on the training results and comparisons, the module structure and parameters were continuously refined to achieve the optimal model suitable for production use.

[0087] Step (4) applies the model to the real dataset LSRW to test the effect.

[0088] To further verify the effectiveness of the proposed method in real-world scenarios, we test the method on the real-world dataset LSRW, which includes images captured in real scenes using a Nikon D7500 camera and a Huawei P40Pro phone.

[0089] DT-Retinex is an innovative deep learning model for low-light image enhancement. It proposes a multi-stage collaborative optimization framework based on Retinex theory. The model decomposes the image into illumination and reflectance components using the DecomNet module, preserving underlying structural features through skip connections and branching. The RDnoiseNet module introduces a diffusion model to progressively denoise the reflectance component, achieving a balance between detail restoration and noise suppression through an improved denoising loss function. The ReLumenNet module combines the Light-Illumination Transformer (LIT) and the Convolutional BlockAttention Module (CBAM) attention mechanism to construct an encoder-decoder network architecture, achieving adaptive adjustment of illumination features and local contrast enhancement. In the decomposition stage, the model preserves high-frequency details through multi-scale feature extraction. In the denoising stage, it uses the inverse diffusion process of the diffusion model to restore texture structure. In the enhancement stage, the attention mechanism highlights key areas, effectively addressing the problems of insufficient brightness, noise interference, and detail loss in low-light images. This method demonstrates significant advantages in illumination modeling, detail preservation, and perceptual quality optimization, particularly in enhancing small objects and suppressing artifacts in complex lighting environments. Future research will focus on improving the model's adaptability in extreme low-light conditions, exploring lightweight design to optimize computing efficiency, and promoting the technology's cross-scenario application potential in areas such as autonomous driving and medical imaging.

Claims

1. A low-light image enhancement method based on Retinex theory, characterized in that: The following steps are involved: Step (1) Select a relatively classic data set and obtain low-light images and corresponding normal-light images; Step (2) Based on the Retinex theory, a DT-Retinex model is constructed; Step (3) training the constructed DT-Retinex model; Step (4) applies the trained DT-Retinex model to low-light images on multiple validation datasets to obtain enhanced results.

2. The low-light image enhancement method based on Retinex theory according to claim 1, characterized in that: Use a camera or other image acquisition device to collect low-light images and their corresponding normal-light images in different scenes. At the same time, select some data that meets the low-light and normal-light characteristics from the public image dataset as the training dataset.

3. The low-light image enhancement method based on Retinex theory according to claim 1, characterized in that: The League of Legends v1 and League of Legends v2 datasets and group images from different scenes are used for model training. The training dataset covers a variety of indoor scenes, providing rich information for model learning. During the model verification phase, the trained model was applied to multiple datasets such as DICM and LIME. The model performance and adaptability were verified by comparing indicators such as PSNR and SSIM of the images before and after enhancement. In actual scene evaluation, the model was applied to the real datasets LSRW-HUAWEI and LSRW-Nikon. After enhancing the low-light images, visual comparisons were made in terms of brightness, color, and details to determine the effectiveness and reliability of the model in real scenes.

4. The low-light image enhancement method based on Retinex theory according to claim 1, characterized in that: The DecomNet part used in building the DT-Retine model includes: introducing a dual-branch structure in the decomposition model architecture to replace the original image decomposition method. The dual branches share weights, one is responsible for extracting the reflection component, and the other focuses on extracting the illumination component; during the decomposition process, DecomNet uses a cross-layer channel splicing strategy to splice feature maps of different levels according to the channel dimension; for example, shallow feature maps contain rich detail information, and deep feature maps have strong semantic information. The two are spliced ​​to retain high-frequency details; at the same time, DecomNet uses multiple loss functions such as reconstruction loss, reflection consistency loss, illumination smoothing loss, and illumination consistency loss to optimize the image decomposition process, ensuring that while accurately separating the illumination and reflection components, the original image can be effectively reconstructed to maintain the stability and consistency of the image illumination information.

5. The low-light image enhancement method based on Retinex theory according to claim 4, characterized in that: The RDnoiseNet part used in building the DT-Retinex model includes: The RDnoiseNet part used in building the DT-Retinex model includes: introducing a diffusion model to denoise the reflection component; traditional denoising methods are difficult to balance noise suppression and detail preservation. The diffusion model learns the noise distribution of the reflection component by gradually adding and removing noise. In the loss function design, the original loss function is improved, and expectation is introduced to consider the sample distribution, thereby enhancing training stability and model generalization ability; at the same time, self-constrained consistency loss is introduced to ensure that the recovered features are consistent with the intrinsic information of the input low-light image, and the denoising loss and self-constrained consistency loss are combined to form a total loss function, thereby guiding the model to retain image details during denoising; RDnoiseNet first allows the reflection component to be processed by the diffusion model, adjusts the parameters according to the total loss function, and continuously optimizes the denoising effect; finally, it provides high-quality, low-noise and detail-rich reflection components for subsequent image enhancement steps, thereby improving the quality of overall image enhancement.

6. The low-light image enhancement method based on Retinex theory according to claim 1, characterized in that: The ReLumenNet architecture used in the DT-Retinex model includes: an encoder-decoder ReLumenNet network structure within the model architecture; the encoder utilizes a progressively downsampling convolutional module, paired with instance normalization and the LeakyReLU activation function, to efficiently extract multi-level detailed features from low-light input images. Within the encoder, a LIT module is introduced. This module adds positional encoding before performing multi-head self-attention calculations, helping the model accurately capture information at different image locations. In the value branch, an illumination compensation path is cleverly integrated, significantly improving illumination modeling capabilities and computational efficiency. Furthermore, the CBAM attention mechanism is introduced, which collaboratively optimizes across both channel and spatial dimensions. Channel attention focuses on the importance of features across different channels, while spatial attention focuses on key information at different spatial locations. During the illumination enhancement phase, it adaptively adjusts and enhances illumination features across different regions, effectively improving the brightness, color, and detail of low-light images.

7. The low-light image enhancement method based on Retinex theory according to claim 1, characterized in that: The improved DT-Retinex model includes the following: To improve training efficiency, the input image is randomly cropped to 96×96 pixels, and training is performed on an NVIDIA 4090 GPU based on the PyTorch framework, using Python version 3.8 and Cuda version 11.

3. A phased training strategy is adopted, using the Adam optimizer with default parameter settings, a fixed learning rate of 0.0001, a total of 200 training iterations, and a batch size of 128.

8. The low-light image enhancement method based on Retinex theory according to claim 1, characterized in that: The trained model is applied to actual scenes to obtain results. The specific operations are as follows: first, the robustness and generalization evaluation of the trained model in claim 7 is carried out on multiple verification data sets such as DICM, LIME, MEF, NPE, VV, and real data sets such as LSRW-HUAWEI and LSRW-Nikon; with the help of indicators such as PSNR, SSIM, LPIPS, NIQE, BRISQUE, etc., the performance of the model in various low-light scenes is measured from different dimensions to judge its stability and adaptability in processing different images. After the evaluation is completed, the model is directly applied to the actual low-light image to obtain the enhanced image result.

Images