A low-light image enhancement method and system

Abstract

This invention discloses a low-light image enhancement method and system, belonging to the field of image processing technology. The method involves projecting the input image from the sRGB color space to the CIELAB color space through a color space transformation module in a constructed network model, decoupling it into physically independent luminance and chrominance components. A luminance stream encoder extracts multi-scale luminance features from the luminance component, and a chrominance stream encoder extracts multi-scale chrominance features from the chrominance component. A cross-attention module performs bidirectional information exchange between the multi-scale luminance and chrominance features to generate fused features. The luminance stream decoder and chrominance stream decoder use the fused features as input to reconstruct the enhanced luminance and chrominance channels, respectively, and output the enhanced image after inverse color space transformation. This method can solve the problems of color distortion, overexposure, and loss of detail encountered when processing complex lighting changes.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method and system for enhancing low-light images. Background Technology

[0002] Low-light image enhancement (LLIE) is a key technology for solving image quality degradation in scenarios such as nighttime surveillance, autonomous driving, and deep-sea exploration.

[0003] Currently, the most mainstream and widely used method in existing technologies is end-to-end deep learning enhancement based on the sRGB color space (such as enhancement networks based on Transformer or CNN). This method directly uses low-light sRGB images as input and leverages the powerful feature extraction capabilities of convolutional neural networks (CNNs) or visual Transformer architectures to extract multi-scale features in both spatial and channel dimensions. Driven by a large amount of paired or unpaired data, the network directly learns a non-linear mapping function from low-light to normal-light conditions, ultimately outputting the enhanced image directly in the sRGB color space.

[0004] Because the sRGB color space is a gamma-compressed non-linear space with high correlation and coupling among its R, G, and B channels, luminance and chrominance information are mixed together. When a deep network attempts to increase the brightness of dark areas (i.e., increase the absolute pixel value) through non-linear mapping, the lack of physical orthogonality makes it difficult for the network to strictly maintain the inherent proportional relationship between the R, G, and B channels while increasing brightness. This luminance-color coupling mechanism causes the brightness increase operation to directly disrupt color balance, leading to color shift, distortion, and local saturation overflow, thus failing to achieve high-quality visual restoration. Summary of the Invention

[0005] To address the problems existing in the aforementioned fields, this invention proposes a low-light image enhancement method and system. The constructed network model can solve the problems of color distortion, brightness overexposure, and detail loss faced by existing methods when dealing with complex lighting changes.

[0006] To address the aforementioned technical problems, this invention discloses a low-light image enhancement method, comprising the following steps: Acquire low-light RGB images; Input a low-light RGB image into a pre-trained network model, and output an enhanced image through forward inference; The network model is based on the traditional U-Net network structure, with a color space transformation module embedded before the encoder of the original U-Net network; and the encoder-decoder of the original U-Net network is transformed into a parallel luma stream encoder-decoder and a chroma stream encoder-decoder, with a cross attention module embedded between the corresponding layers of the luma stream encoder and the chroma stream encoder. The color space transformation module utilizes the physical orthogonality of the CIELAB color space to project the input image from the sRGB color space to the CIELAB color space, decoupling it into physically independent luminance and chrominance components. The luminance stream encoder extracts multi-scale luminance features of the luminance component, and the chrominance stream encoder extracts multi-scale chrominance features of the chrominance component. The cross-attention module performs bidirectional information exchange between the multi-scale luminance and multi-scale chrominance features, generating fusion features by establishing guiding constraints of luminance features on chrominance restoration and balancing constraints of chrominance features on luminance enhancement. The luminance stream decoder and chrominance stream decoder use the fusion features as input to reconstruct the enhanced luminance and chrominance channels respectively, and output the enhanced image after inverse color space transformation.

[0007] Preferably, the cross-attention module performs bidirectional information exchange between multi-scale luminance features and multi-scale chrominance features. By establishing guiding constraints of luminance features on chrominance recovery and balancing constraints of chrominance features on luminance enhancement, it generates fused features, specifically including: The cross-attention module, along the feature processing direction, sequentially includes an intensity-guided feature rectification component, a cross-attention component, and a multi-scale structure injection component, wherein: The intensity-guided feature rectification component uses a zero-initialization residual structure to adaptively purify multi-scale luminance features and multi-scale chrominance features to obtain purified luminance features and chrominance features. The cross-attention component employs a bidirectional attention mechanism to perform bidirectional information interaction between the purified luminance and chrominance features, thereby obtaining interactive features. The multi-scale structure injection component employs a channel-level adaptive weighting strategy to compensate for high-frequency texture details of interactive features, thereby obtaining fused features.

[0008] Preferably, the intensity-guided feature rectification component employs a zero-initialization residual structure to adaptively purify and clean multi-scale luminance and chrominance features, obtaining purified luminance and chrominance features, specifically including: When multi-scale luminance features are used as input features, multi-scale chrominance features are used as guiding features; when multi-scale chrominance features are used as input features, multi-scale luminance and chrominance features are used as guiding features. A dual-stream concurrent design is used to extract the global intensity information and edge structure information of the guiding features separately, and then they are fused to obtain the cleaned features. for: ; ; in, Let C, H, and W be the number of channels, height, and width of the input feature, respectively. For guiding features; and These correspond to 1×1 convolution and 3×3 convolution; To introduce nonlinearity; For projection; Based on guiding features The generated adaptive attention weight map; This represents a zero-initialized residual structure, where, This indicates element-wise multiplication; This is the residual scaling factor.

[0009] Preferably, the cross-attention component employs a bidirectional attention mechanism to perform bidirectional information interaction between the purified luminance and chrominance features to obtain interaction features, specifically including: Features after cleaning the current branch As query Q, the guiding feature Y of the other branch is used as key K and value V, for the purified features. Perform standard cross-attention operations: ; ; Obtain the global attention mechanism graph and perform feature fusion to obtain interaction features. for: in, For depthwise separable convolution, This is a similarity matrix. For normalization function, It is the temperature coefficient.

[0010] Preferably, the multi-scale structure injection component employs a channel-level adaptive weighting strategy to compensate for high-frequency texture details of interactive features to obtain fused features, specifically including: We use standard convolution and dilated convolution in parallel to extract local detail features and contextual structure features of interaction features, thereby capturing the multi-scale structural features of interaction features. Channel-level adaptive weighting is applied to interactive features and multi-scale structural features to generate fused features.

[0011] Preferably, the network model further includes an illumination-sensing modulation module, which is embedded in the bottleneck layer between the luminance stream encoder and the luminance stream decoder; The illumination sensing modulation module uses the original luminance component as a global guiding signal and performs spatial modulation on the deep features output by the luminance stream encoder through a gated attention mechanism to generate modulated deep features. The modulated deep features are input into the luminance stream decoder for reconstruction.

[0012] Preferably, the network model further includes multiple structural enhancement modules, which are respectively embedded in the skip connection paths of the luma stream encoder-decoder and the chroma stream encoder-decoder; The structure enhancement module employs a depthwise separable convolution residual structure to perform noise filtering and structural detail enhancement on the shallow jump features output by the luminance stream encoder and chrominance stream encoder in the jump connection path, generating refined jump features. The refined jump features are fused with the fusion features.

[0013] Preferably, the network model further includes a color space inverse transformation module and a thinning head module sequentially connected to the outputs of the luminance stream decoder and the chrominance stream decoder, wherein: The color space inverse transformation module converts the reconstructed luminance and chrominance channels into a coarse RGB image; the thinning head module performs global residual fusion between the coarse RGB image and the input image to generate the final enhanced image.

[0014] Preferably, the color space transformation module utilizes the physical orthogonality of the CIELAB color space to project the input image from the sRGB color space to the CIELAB color space, decoupling it into physically independent luminance and chrominance components, specifically including: Perform inverse gamma correction on the pixels of the input image to obtain a linear RGB image; A linear RGB image is mapped to the CIEXYZ space through a standard linear transformation to generate an image represented in the CIEXYZ color space. The components of the image represented in the CIEXYZ color space are normalized relative to the reference white point, and the physical orthogonal luminance and chrominance components are extracted by simulating the nonlinear response of the human eye to illumination through a piecewise nonlinear function. The numerical range of the luminance and chrominance components is normalized, the luminance component is scaled to the [0,1] interval, and the chrominance component is scaled to the [-1,1] interval, generating a feature tensor that adapts to the input of the network model.

[0015] Preferably, it further includes a low-light image enhancement system, comprising: The data acquisition module is used to acquire low-light RGB images; The image enhancement module is used to input low-light RGB images into a pre-trained network model and output enhanced images through forward inference. The network model is based on the traditional U-Net network structure, embedding a color space transformation module before the encoder of the original U-Net network; and transforming the original U-Net encoder-decoder into a parallel luminance stream encoder-decoder and chrominance stream encoder-decoder, embedding a cross-attention module between the corresponding layers of the luminance stream encoder and chrominance stream encoder; the color space transformation module utilizes the physical orthogonality of the CIELAB color space to project the input image from the sRGB color space to the CIELAB color space, decoupling it into physically independent luminance and chrominance components; the luminance stream encoder extracts multi-scale luminance features of the luminance component, and the chrominance stream encoder extracts multi-scale chrominance features of the chrominance component; the cross-attention module performs bidirectional information interaction between the multi-scale luminance and multi-scale chrominance features, generating fusion features by establishing guiding constraints of luminance features on chrominance restoration and balancing constraints of chrominance features on luminance enhancement; the luminance stream decoder and chrominance stream decoder use the fusion features as input to reconstruct the enhanced luminance and chrominance channels respectively, and output the enhanced image after inverse color space transformation.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The low-light image enhancement method proposed in this invention constructs a network model that, through a color space transformation module, utilizes the inherent physical orthogonality of the CIELAB color space to explicitly decouple the highly coupled luminance and chrominance information in the traditional sRGB color space. This physical separation ensures that when the network enhances the brightness of dark areas in subsequent processing, it does not mathematically interfere with the color ratio, fundamentally overcoming the common defect of existing end-to-end networks that result in color cast after brightening, and achieving excellent color fidelity. Through a parallel architecture of luminance and chrominance streams, feature modeling is performed separately for these two heterogeneous data types. Since the cross-attention module achieves deep luminance and chrominance information fusion during the encoding stage, the subsequent decoder can reconstruct based on high-quality fused features, effectively avoiding excessive amplification of noise in dark areas and clearly preserving the edge and texture details of the image. By establishing a cross-attention module, luminance features guide chroma restoration and chroma features balance luminance enhancement. On one hand, luminance features with rich texture structures guide chroma features to restore correct colors; on the other hand, chroma features balance the luminance enhancement process. This mutual constraint mechanism effectively avoids local highlight overexposure and artifacts caused by the network blindly boosting brightness, making the enhancement results more visually natural and harmonious. The network model constructed by this method adopts a modular embedding + dual-stream modification strategy, effectively controlling the number of parameters and computational complexity while ensuring long-distance dependency modeling capabilities and luminance-color interaction quality. This makes the network model more practical for real-world engineering applications. Attached Figure Description

[0017] Figure 1 This is a flowchart of the low-light image enhancement method proposed in this invention; Figure 2 The architecture of the network model constructed for this invention; Figure 3 A schematic diagram of the color space transformation module of the network model constructed in this invention; Figure 4 This is a schematic diagram of the two-stream branch structure of the network model constructed in this invention; Figure 5 A schematic diagram of the cross-attention module of the network model constructed in this invention; Figure 6 Enhanced images of the paired dataset LOL-v1 provided in this embodiment of the invention; Figure 7 Enhanced images of the pairwise dataset LOL-v2-Real provided in this embodiment of the invention; Figure 8 Enhanced images of the unpaired datasets LIME and DICM provided in this embodiment of the invention; Figure 9Enhanced images of ablation experiments on paired datasets provided in embodiments of the present invention; Figure 10 Enhanced images of ablation experiments on unpaired datasets provided in embodiments of the present invention. Detailed Implementation

[0018] The following will refer to the appendices in the embodiments of the present invention. Figures 1-10 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0019] Example like Figure 1 As shown, this invention proposes a low-light image enhancement method, which includes the following steps: S1: Acquire a low-light RGB image; S2: Input the low-light RGB image into the pre-trained network model and output the enhanced image through forward inference.

[0020] The constructed network model (LAB-Feature Rectification and Intensity Net, LAB-FRINet) is as follows: Figure 2 As shown, based on the traditional U-Net network structure, a color space transformation module is embedded before the encoder of the original U-Net network; and the encoder-decoder of the original U-Net network is transformed into a parallel luma stream encoder-decoder and a chroma stream encoder-decoder, with a cross attention module embedded between the corresponding layers of the luma stream encoder and the chroma stream encoder. The color space transformation module utilizes the physical orthogonality of the CIELAB color space to project the input image from the sRGB color space to the CIELAB color space, decoupling it into physically independent luminance and chrominance components. The luminance stream encoder extracts multi-scale luminance features of the luminance component, and the chrominance stream encoder extracts multi-scale chrominance features of the chrominance component. The cross-attention module performs bidirectional information exchange between the multi-scale luminance and multi-scale chrominance features, generating fused features by establishing guiding constraints of luminance features on chrominance restoration and balancing constraints of chrominance features on luminance enhancement. The luminance stream decoder and chrominance stream decoder use the fused features as input to reconstruct the enhanced luminance and chrominance channels, respectively, and output the enhanced image after inverse color space transformation.

[0021] Specifically, in step S2, to address the enhancement inconsistency problem caused by the high coupling between luminance and chrominance in the sRGB color space, the method proposed in this invention first introduces a fully differentiable color space transformation module (CST). This module transforms the input low-light RGB image... Projecting onto a perceptually uniform LAB space achieves explicit decoupling of physical properties at the data level. This process includes three steps: linearization, color space projection, and perceptual nonlinear mapping, such as... Figure 3 As shown.

[0022] linearization Because sRGB data undergoes gamma compression, it does not conform to the physical linear superposition property of light. Therefore, inverse gamma correction is first performed on the pixels of the input image to restore linear brightness and obtain a linear RGB image.

[0023] Color space projection A linear RGB image is mapped to the CIEXYZ color space using a standard linear transformation to generate an image represented in the CIEXYZ color space. for: ; in, This is a 3×3 transformation matrix based on the D65 standard white point. It is a linear RGB image.

[0024] Perceptual nonlinear mapping To simulate the nonlinear response of the human eye to illumination and extract orthogonal features, images represented in the CIEXYZ color space were... The components of each channel relative to the reference white point Perform normalization and use the normalized value as the independent variable. t Input to nonlinear activation function The defined nonlinear activation function for: ; Among them, the truncation threshold And slope factor The values ​​strictly adhere to CIE standards.

[0025] By simulating the nonlinear response of the human eye to illumination using a piecewise nonlinear activation function, physically orthogonal luminance components are extracted. With chromaticity components and : ; ; ; Finally, to adapt to the numerical range of the neural network and stabilize the training gradient, the numerical ranges of the luminance and chrominance components are normalized. The luminance component is scaled to the [0,1] interval, and the chrominance component is scaled to the [-1,1] interval, generating a feature tensor that adapts to the input of the network model. for: ; This design provides orthogonal and numerically stable feature priors for subsequent parallel luminance stream encoder-decoder and chrominance stream encoder-decoder.

[0026] Parallel luminance stream encoder and chrominance stream encoder To address the issue of illumination and color coupling in low-light image enhancement tasks, this invention constructs a network model that differs from the traditional U-Nnet network. In the encoding and decoding stages, it proposes parallel luminance stream encoder-decoder and chrominance stream encoder-decoder as Feature Rectification and Intensity (FRI) modules. The input low-light RGB image is first processed by a differentiable transformation module. Will Convert to LAB color space:

[0027] ; Subsequently, Decoupled into luminance components and chromaticity components ; These two components are fed into two parallel luma stream encoders and chroma stream encoders, respectively, to extract luma features. and chromaticity characteristics .

[0028] Although parallel luma stream encoders and chroma stream encoders can effectively extract multi-scale luma and chroma features, achieving efficient interaction between the two feature streams remains a challenge.

[0029] Although the standard cross-attention mechanism can capture global dependencies, it has two main drawbacks in low-light enhancement tasks: (1) High-frequency noise contained in the original features can easily interfere with the generation of the query, resulting in inaccurate calculation of the attention map and the model is very likely to get stuck in local optima; (2) Attention mechanism is essentially a low-pass filter, which can easily smooth out key high-frequency texture details in long-distance modeling.

[0030] To address the aforementioned problems, this invention proposes an improved lightweight cross-attention module (RectifiedIntensity-guided Cross-attention, RICA), such as... Figure 4As shown, RICA includes an Intensity-Guided Feature Rectification (IGFR) component, a cross-attention interaction component, and a Multi-Scale Structural Injection (MSSI) component. In other words, RICA introduces the Intensity-Guided Feature Rectification (IGFR) component and the Multi-Scale Structural Injection (MSSI) component into the standard Transformer architecture.

[0031] Intensity-guided characteristic rectifier component IGFR To generate clean query features and accelerate network convergence, this invention introduces the IGFR module before attention calculation. Unlike traditional single-path convolution, IGFR adopts a dual-stream concurrent design: (1) Intensity pass-through: Utilizing the low receptive field of 1×1 convolution, the feature intensity is quickly increased and the gradient is directly backpropagated, effectively avoiding training stagnation caused by gradient vanishing; (2) Texture denoising: Utilizing 3×3 convolution to capture local spatial context and filter out high-frequency noise in the input features.

[0032] Assume the input features are The guiding feature is Y, and the calculation process of IGFR can be described as follows: ; ; in, Let C, H, and W be the number of channels, height, and width of the input feature, respectively. and These correspond to 1×1 convolution and 3×3 convolution; To introduce nonlinearity; For projection; Based on guiding features The generated adaptive attention weight map; This represents a zero-initialized residual structure, where, This indicates element-wise multiplication; The learnable residual scaling factor is initialized to 0. This zero-initialization residual structure ensures a smooth start-up of the model in the early stages of training. As training progresses, IGFR gradually rectifies the input features X to obtain purified features. Used for subsequent query generation.

[0033] Cross-attention interaction components After obtaining purified characteristics Then, standard cross-attention calculation is performed. To fully utilize the complementarity of the two-stream information, this invention will perform the current branch's... The query uses the leading feature Y of the other branch as both the key and value:

[0034] ; ; Subsequently, the global attention mechanism graph is calculated and features are fused to obtain interaction features. for: ; in, For depthwise separable convolution, This is a similarity matrix. For normalization function, A learnable temperature coefficient is used to adjust the sharpness of the attention distribution. Through this design, the luminance stream can guide the chrominance stream to restore the correct colors, and the chrominance stream can also assist the luminance stream in distinguishing texture from noise.

[0035] Multiscale Structure Injection Component (MSSI) To address the issue of high-frequency detail loss caused by the attention mechanism in cross-attention interaction components, this invention designs MSSI at the end of the cross-attention interaction component. MSSI uses standard convolution in parallel. (Extracting local details) and dilated convolution (Extracting contextual structure) Process the guiding feature Y to capture multi-scale structural information.

[0036] To achieve refined feature fusion, this invention abandons traditional scalar weighting and instead adopts a channel-wise adaptive weighting strategy to obtain fused features. : ; ; in, It is a channel-level learnable parameter vector that allows the network to adaptively enhance or suppress structural information in different channels, thereby significantly improving the edge sharpness and texture details of the image while maintaining global consistency.

[0037] The luma stream decoder and chroma stream decoder take the fused features as input and reconstruct the enhanced luma and chroma channels respectively through a feedforward network (GDFN / IEL) with a gating mechanism. After inverse color space transformation, the enhanced image is output, such as... Figure 5 As shown.

[0038] Light sensing modulation module Standard U-Net network architectures typically suffer from the problem of global illumination prior degradation in deep semantic features. To alleviate this problem, this invention embeds an illumination-aware modulation (IAM) module in the bottleneck layer between the luminance stream encoder and luminance stream decoder.

[0039] The Illumination Aware Modulation (IAM) module uses the original luminance component as a global guiding signal and performs spatial modulation on the deep features output by the luminance stream encoder through a gated attention mechanism to generate modulated deep features: make This represents the input feature map of the bottleneck layer. This indicates the original brightness input. Because... Spatial resolution lower than First, the bilinear interpolation operator is applied. Perform spatial dimension alignment:

[0040] ; Spatial attention map The generation is modeled as a nonlinear mapping function. Specifically, a cascaded operation is employed, including a 3×3 convolution for extracting local context, a LeakyReLU activation function for introducing non-linearity, and a 1×1 convolution for channel projection. The process is defined as follows:

[0041] ; ; in, This represents the convolution operation. ( () represents the LeakyReLU activation function with a negative slope of 0.1. ( ) is the Sigmoid function, to ensure . , and , These represent the learnable weights and biases, respectively. Intermediate features The channel dimension was reduced to To reduce computational costs, the modulated deep features are ultimately obtained through residual affine transformation. :

[0042] ; in, This is the Hadamard product (element-by-element multiplication). This formula allows the network to adaptively enhance the feature responses of the regions that need to be recovered, based on the original illumination distribution.

[0043] The modulated deep features are input into the luminance stream decoder for reconstruction.

[0044] Structural reinforcement module Because skip connections in traditional U-Net directly pass shallow features to the decoder, they often introduce high-frequency noise. To suppress noise while preserving structural details, this invention proposes a Structure Enhancement Module (SEM). SEM is mathematically modeled as a residual structure-aware filter employing depthwise separable convolution.

[0045] Multiple structural enhancement modules are embedded in the skip connection paths of the luma stream encoder-decoder and chroma stream encoder-decoder, respectively. The structural enhancement modules employ a residual structure of depthwise separable convolution to perform noise filtering and structural detail enhancement on the shallow skip features output by the luma stream encoder and chroma stream encoder in the skip connection paths, generating refined skip features.

[0046] make This represents the shallow hop features from the encoder's hop connection path output. The SEM module learns a residual component. To refine Structure extraction, the extraction process specifically includes:

[0047] ; in, ( ) indicates that the number of groups is equal to the channel dimension C of a 3×3 depth convolution, and this operation processes the spatial information of each channel independently. ( ) represents batch normalization.

[0048] Subsequently, pointwise convolution is used to fuse channel information, followed by the PReLU activation function. ( ): ; Final refined jump features Refined jump features are obtained by adding the learned residual structural features back to the original input. for: ; This formula ensures It can capture clear high-frequency edge information, thus effectively acting as a learnable denoising filter before the feature fusion stage of the decoder.

[0049] The refined jump features are then fused with the fusion features.

[0050] Refine head module To eliminate potential color artifacts and spatial discontinuities in the reconstructed image, this invention connects an inverse color space transform module and a refinement head module sequentially to the outputs of the luminance stream decoder and the chrominance stream decoder. This module models the image reconstruction process as a residual learning problem between coarse prediction and the final high-quality output.

[0051] The inverse color space transformation module converts the reconstructed luminance and chrominance channels into a coarse RGB image; the thinning head module performs global residual fusion between the coarse RGB image and the input image to generate the final enhanced image. make A rough RGB image reconstructed by the decoder. The input image is the original image. First, the two images are concatenated along the channel dimension to form a hybrid tensor. :

[0052] ; The refinement head module approximates the residual mapping function R by stacking alternating convolutional layers and PReLU activation functions. ),make This represents the residual map of the i-th convolutional layer in the head. The calculation is as follows: ; The weights of the convolutional layer Initialize to zero to ensure identity mappings during the initial training phase. Final augmented image. for:

[0053] ; This global residual form allows the network to focus only on learning high-frequency residuals used to correct artifacts, thus ensuring training stability and faster convergence.

[0054] The present invention also proposes a low-light image enhancement system, comprising: The data acquisition module is used to acquire low-light RGB images; The image enhancement module is used to input low-light RGB images into a pre-trained network model and output enhanced images through forward inference. The network model is based on the traditional U-Net network structure, embedding a color space transformation module before the encoder of the original U-Net network. The original U-Net encoder-decoder is transformed into a parallel luminance stream encoder-decoder and chrominance stream encoder-decoder, with a cross-attention module embedded between corresponding layers of the luminance stream encoder and chrominance stream encoder. The color space transformation module utilizes the physical orthogonality of the CIELAB color space to project the input image from the sRGB color space to the CIELAB color space, decoupling it into physically independent luminance and chrominance components. The luminance stream encoder extracts multi-scale luminance features of the luminance component, and the chrominance stream encoder extracts multi-scale chrominance features of the chrominance component. The cross-attention module performs bidirectional information exchange between the multi-scale luminance and multi-scale chrominance features, generating fused features by establishing guiding constraints of luminance features on chrominance restoration and balancing constraints of chrominance features on luminance enhancement. The luminance stream decoder and chrominance stream decoder use the fused features as input to reconstruct the enhanced luminance and chrominance channels respectively, and output the enhanced image after inverse color space transformation.

[0055] Experimental Analysis To comprehensively evaluate the performance and generalization ability of the proposed method under different illumination distributions and real-world degradation scenarios, two types of benchmark datasets were used for experiments: paired supervised datasets for quantitative analysis and unpaired natural scene datasets for generalization testing. The acquisition process of these datasets was legal and reasonable. Specifically:

[0056] The supervised benchmark datasets were used for training and quantitative evaluation on two widely used datasets, both containing image pairs under low and normal lighting conditions. The LOL-v1 dataset contains 500 pairs of real-world captured images, with 485 pairs used for training and 15 pairs for testing. The images are primarily captured from indoor scenes, exhibiting strong light attenuation and noise interference, making them a core benchmark for evaluating the color fidelity of low-light enhancement algorithms. To further verify the model's robustness under complex lighting variations, we used the larger LOL-v2-Real dataset, containing 689 pairs of real-world scene images (600 pairs for training and 89 pairs for testing). Compared to the v1 version, its scenes are more diverse, covering a dynamic range from low light to strong light.

[0057] To verify the model's generalization ability in unseen real-world natural scenes, cross-dataset testing was conducted using LIME and DICM, two classic unpaired datasets. LIME contains challenging outdoor low-light scenes; DICM mainly covers night scenes and backlit images from commercial photography.

[0058] For the LOLv1 and LOLv2 datasets containing paired ground truth values, we employ Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). For LIME and DICM, which lack ground truth values, we introduce Average Gradient (AG), Information Entropy (IE), Natural Image Quality Evaluation (NIQE), and No-Reference Quality Evaluation (NIQMC) for contrast distortion.

[0059] The network model constructed in this invention is implemented based on the PyTorch framework and trained on a single NVIDIA RTX 4060 GPU. The Adam optimizer is used to update the network parameters, and the initial learning rate is set to 2×10⁻⁶. −4 To ensure the stability of the training process and avoid getting trapped in local optima, a cosine annealing strategy with preheating is adopted to adjust the learning rate. Specifically, the first X epochs are used for linear preheating, and then the learning rate gradually decays to 1×10⁻⁶ according to a cosine function. −6 Furthermore, to prevent gradient explosion, a gradient clipping mechanism was introduced, limiting the maximum norm of the gradient to 0.01. To improve the model's generalization ability and prevent overfitting, online data augmentation strategies were employed during training. In addition to conventional random rotation and flipping, random gamma correction was introduced, which simulates different lighting distributions by applying random nonlinear transformations to the input image, enhancing the network's robustness to lighting changes. The input image was randomly cropped into 256×256 image patches, with a batch size of 4. The total number of training epochs was 400.

[0060] To fully verify the performance advantages of the proposed method, ten representative advanced algorithms were selected for comparison. The comparison included traditional methods such as KinD, KinD++, Zero-DCE, Zero-DCE++, SCI, MBLLEN, DARKIR, SNR, HVI, and HVI-CIDNet. The experiments were divided into two parts: full-reference evaluation on paired datasets (LOL-v1, LOL-v2-Real), and no-reference and generalization evaluation on unpaired datasets (LIME, DICM).

[0061] In this evaluation, the LAB-FRINet network model constructed in this invention was comprehensively compared with the aforementioned traditional methods. Specific data and visual effects are shown in Tables 1-3 and... Figures 6-8 middle.

[0062] Table 1. Comparison results of paired datasets Table 1 shows the comparison results of each method on paired datasets. Figure 6As shown, this is an augmented image on the pairwise dataset LOL-v1, such as... Figure 7 The image shown is an augmented image on the pairwise dataset LOL-v2-Real.

[0063] By analyzing the data in Table 1, Figure 6 and Figure 7 It can be seen that the LAB-FRINet network model constructed in this invention has better image quality than traditional methods when obtaining enhanced images on paired datasets. Therefore, the LAB-FRINet network model constructed in this invention is more robust when dealing with complex lighting changes. The method of this invention not only effectively removes noise, but also better preserves the natural colors and structural details of the image.

[0064] Table 2 Comparison results on unpaired datasets Table 2 shows the comparison results of each method on unpaired datasets. Figure 8 As shown, the top image is the augmented image on the unpaired dataset LIME, and the bottom image is the augmented image on the unpaired dataset DICM.

[0065] By analyzing the data in Table 2 and Figure 8 On the unpaired datasets LIME and DICM, the network model constructed in this invention performs significantly better than traditional methods when evaluated using metrics such as NIQE. This demonstrates that the proposed method possesses strong generalization ability and can adapt to diverse natural scenarios in the real world.

[0066] In summary, the network model constructed by the method proposed in this invention seamlessly integrates traditional image processing methods with modern deep learning techniques. Through a direct and efficient two-stream interaction structure, it successfully achieves state-of-the-art experimental performance and highlights the synergistic effect between the prior physical model and the data-driven approach.

[0067] ablation experiment To verify the effectiveness of the three core innovations proposed in the LAB-FRINet network model constructed by the method of this invention—namely, the LAB module, the RICA module, and the FRI module—a series of ablation experiments were conducted on the LOL-v1 and LOL-v2-real datasets in this embodiment. To ensure fairness, all experimental models were trained under the same hyperparameter settings (400 epochs, Batch Size 4). Detailed quantitative comparison results are shown in Table 3, and visual effect comparisons can be found in [link to table]. Figure 9 and Figure 10 .

[0068] Table 3 Comparison Results of Ablation Experiments To explore the impact of color space on enhancement performance, this invention compares the model Model (c) trained in the traditional RGB space (i.e., removing LAB and retaining only RICA and FRI) with the complete model Model (e). Since the three channels in the RGB space are highly correlated, direct enhancement in the RGB domain often struggles to balance brightness adjustment and color preservation. In contrast, by introducing the LAB module, the network can explicitly decouple the illumination component (L) and color component (AB) using physical properties. As shown in Table 3, compared to the RGB baseline model Model (c), the complete model Model (e) with the LAB module improved the PSNR by 0.24 dB (from 25.83 to 26.07) and 0.55 dB (from 28.97 to 29.52) on LOL-v1 and LOL-v2-Real, respectively. This fully demonstrates that the LAB module is a key foundation for overcoming color distortion bottlenecks and improving the upper limit of recovery in extremely dark scenes.

[0069] Building upon the LAB module, the role of the RICA module is further validated. Model (d) removes the RICA module, using only simple feature concatenation to fuse the two-stream branches, which severely limits the information complementarity between different modalities. Model (e) is the complete model with the RICA module incorporated. The network in Model (e) can adaptively align and fuse luminance and color features through the pre-cleaning component and post-structural compensation component in the RICA module. Experimental results show that the addition of the RICA module significantly increases the PSNR of Model (e) from 25.76 dB to 26.07 dB on LOL-v1 and from 29.01 dB to 29.52 dB on LOL-v2-Real, validating the significant advantages of the RICA module in maintaining global consistency and efficient noise resistance.

[0070] Finally, the synergistic effect of the components in the FRI module was verified. Model (b) only includes the LAB and RICA modules, lacking multi-level feature optimization. Model (e) is the complete model with the FRI module. In contrast, Model (e) comprehensively constrains common problems in low-light images such as local overexposure, dark noise, and artifacts: chroma stream coding suppresses oversaturation in highlight areas through illumination perception; luminance stream coding filters and enhances the structural texture of dark areas. As shown in Table 3, after adding the FRI module, the model achieves optimal performance on both datasets (PSNR of 26.07 dB and SSIM of 0.8638 on the LOL-v1 dataset; PSNR of 29.52 dB and SSIM of 0.8970 on the LOL-v2-Real dataset), successfully achieving the best perceptual quality and objective metrics.

[0071] Experimental results on paired datasets such as LOL-v1 and LOL-v2-Real, as well as unpaired datasets such as LIME and MEF, demonstrate that the image enhancement performance of the proposed method significantly outperforms the aforementioned traditional methods. Ablation experiments further validate the effectiveness of the network model constructed using the proposed method, proving that the LAB-FRINet network model can generate high-quality images with natural visuals, clear textures, and accurate colors while maintaining computational efficiency. Therefore, the proposed method can solve the problems of color distortion, overexposure, and loss of detail faced by existing methods when dealing with complex lighting changes.

[0072] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0073] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A low-light image enhancement method, characterized in that, Includes the following steps: Acquire low-light RGB images; Input a low-light RGB image into a pre-trained network model, and output an enhanced image through forward inference; The network model is based on the traditional U-Net network structure, with a color space transformation module embedded before the encoder of the original U-Net network; and the encoder-decoder of the original U-Net network is transformed into a parallel luma stream encoder-decoder and a chroma stream encoder-decoder, with a cross attention module embedded between the corresponding layers of the luma stream encoder and the chroma stream encoder. The color space transformation module utilizes the physical orthogonality of the CIELAB color space to project the input image from the sRGB color space to the CIELAB color space, decoupling it into physically independent luminance and chrominance components. The luminance stream encoder extracts multi-scale luminance features of the luminance component, and the chrominance stream encoder extracts multi-scale chrominance features of the chrominance component. The cross-attention module performs bidirectional information exchange between the multi-scale luminance and multi-scale chrominance features, generating fusion features by establishing guiding constraints of luminance features on chrominance restoration and balancing constraints of chrominance features on luminance enhancement. The luminance stream decoder and chrominance stream decoder use the fusion features as input to reconstruct the enhanced luminance and chrominance channels, respectively, and output the enhanced image after inverse color space transformation.

2. The low-light image enhancement method according to claim 1, characterized in that, The cross-attention module performs bidirectional information exchange between multi-scale luminance features and multi-scale chrominance features. By establishing guiding constraints on chrominance recovery based on luminance features and balancing constraints on luminance enhancement based on chrominance features, it generates fused features, specifically including: The cross-attention module, along the feature processing direction, sequentially includes an intensity-guided feature rectification component, a cross-attention component, and a multi-scale structure injection component, wherein: The intensity-guided feature rectification component uses a zero-initialization residual structure to adaptively purify multi-scale luminance features and multi-scale chrominance features to obtain purified luminance features and chrominance features. The cross-attention component employs a bidirectional attention mechanism to perform bidirectional information interaction between the purified luminance and chrominance features, thereby obtaining interactive features. The multi-scale structure injection component employs a channel-level adaptive weighting strategy to compensate for high-frequency texture details in interactive features, thereby obtaining fused features.

3. The low-light image enhancement method according to claim 2, characterized in that, The intensity-guided feature rectification component employs a zero-initialization residual structure to adaptively purify multi-scale luminance and multi-scale chrominance features, obtaining purified luminance and chrominance features, specifically including: When multi-scale luminance features are used as input features, multi-scale chrominance features are used as guiding features; when multi-scale chrominance features are used as input features, multi-scale luminance and chrominance features are used as guiding features. A dual-stream concurrent design is used to extract the global intensity information and edge structure information of the guiding features separately, and then they are fused to obtain the cleaned features. for: ； ； in, Let C, H, and W be the number of channels, height, and width of the input feature, respectively. For guiding features; and These correspond to 1×1 convolution and 3×3 convolution; To introduce nonlinearity; For projection; Based on guiding features The generated adaptive attention weight map; This represents a zero-initialized residual structure, where, This indicates element-wise multiplication; This is the residual scaling factor.

4. The low-light image enhancement method according to claim 3, characterized in that, The cross-attention component employs a bidirectional attention mechanism to perform bidirectional information exchange between the purified luminance and chrominance features, obtaining interactive features, specifically including: Features after cleaning the current branch As query Q, the guiding feature Y of the other branch is used as key K and value V, for the purified features. Perform standard cross-attention operations: ； ； Obtain the global attention mechanism graph and perform feature fusion to obtain interaction features. for: in, For depthwise separable convolution, This is a similarity matrix. For normalization function, It is the temperature coefficient.

5. The low-light image enhancement method according to claim 2, characterized in that, The multi-scale structure injection component employs a channel-level adaptive weighting strategy to compensate for high-frequency texture details of interactive features, thereby obtaining fused features, specifically including: We use standard convolution and dilated convolution in parallel to extract local detail features and contextual structure features of interaction features, thereby capturing the multi-scale structural features of interaction features. Channel-level adaptive weighting is applied to interactive features and multi-scale structural features to generate fused features.

6. The low-light image enhancement method according to claim 1, characterized in that, The network model also includes an illumination-sensing modulation module, which is embedded in the bottleneck layer between the luminance stream encoder and the luminance stream decoder. The illumination sensing modulation module uses the original luminance component as a global guiding signal and performs spatial modulation on the deep features output by the luminance stream encoder through a gated attention mechanism to generate modulated deep features. The modulated deep features are input into the luminance stream decoder for reconstruction.

7. The low-light image enhancement method according to claim 1, characterized in that, The network model also includes multiple structural enhancement modules, which are respectively embedded in the jump connection paths of the luma stream encoder-decoder and the chroma stream encoder-decoder; The structure enhancement module employs a depthwise separable convolution residual structure to perform noise filtering and structural detail enhancement on the shallow jump features output by the luminance stream encoder and chrominance stream encoder in the jump connection path, generating refined jump features. The refined jump features are fused with the fusion features.

8. The low-light image enhancement method according to claim 1, characterized in that, The network model further includes a color space inverse transformation module and a thinning head module connected sequentially to the outputs of the luminance stream decoder and the chrominance stream decoder, wherein: The color space inverse transformation module converts the reconstructed luminance and chrominance channels into a coarse RGB image; the thinning head module performs global residual fusion between the coarse RGB image and the input image to generate the final enhanced image.

9. The low-light image enhancement method according to claim 1, characterized in that, The color space transformation module utilizes the physical orthogonality of the CIELAB color space to project the input image from the sRGB color space to the CIELAB color space, decoupling it into physically independent luminance and chrominance components. Specifically, it includes: Perform inverse gamma correction on the pixels of the input image to obtain a linear RGB image; A linear RGB image is mapped to the CIEXYZ space through a standard linear transformation to generate an image represented in the CIEXYZ color space. The components of the image represented in the CIEXYZ color space are normalized relative to the reference white point, and the physical orthogonal luminance and chrominance components are extracted by simulating the nonlinear response of the human eye to illumination through a piecewise nonlinear function. The numerical range of the luminance and chrominance components is normalized, the luminance component is scaled to the [0,1] interval, and the chrominance component is scaled to the [-1,1] interval, generating a feature tensor that adapts to the input of the network model.

10. A low-light image enhancement system, characterized in that, include: The data acquisition module is used to acquire low-light RGB images; The image enhancement module is used to input low-light RGB images into a pre-trained network model and output enhanced images through forward inference. The network model is based on the traditional U-Net network structure, embedding a color space transformation module before the encoder of the original U-Net network; and transforming the original U-Net encoder-decoder into a parallel luminance stream encoder-decoder and chrominance stream encoder-decoder, embedding a cross-attention module between the corresponding layers of the luminance stream encoder and chrominance stream encoder; the color space transformation module utilizes the physical orthogonality of the CIELAB color space to project the input image from the sRGB color space to the CIELAB color space, decoupling it into physically independent luminance and chrominance components; the luminance stream encoder extracts multi-scale luminance features of the luminance component, and the chrominance stream encoder extracts multi-scale chrominance features of the chrominance component; the cross-attention module performs bidirectional information interaction between the multi-scale luminance and multi-scale chrominance features, generating fusion features by establishing guiding constraints of luminance features on chrominance restoration and balancing constraints of chrominance features on luminance enhancement; the luminance stream decoder and chrominance stream decoder use the fusion features as input to reconstruct the enhanced luminance and chrominance channels respectively, and output the enhanced image after inverse color space transformation.

Images