Layered decomposition and adaptive fusion-based low-illumination expressway image enhancement method

Abstract

A low-illumination expressway image enhancement method based on hierarchical decomposition and adaptive fusion belongs to the field of image processing, and comprises the following steps: S1, based on an expressway video cloud platform, collecting a video stream of continuous low-illumination segments; s2, constructing an image enhancement model of multi-stage cooperative processing; s3, training an image enhancement model by using the data set; and S4, carrying out detection on a real-time monitoring camera by utilizing the trained image enhancement model, considering objective indexes and subjective evaluation, and verifying the practicability of the algorithm for the expressway low-illumination core task. Compared with a traditional method which only depends on global histogram equalization or a compromise optimization mode of single brightness adjustment, the method provided by the invention effectively solves the contradiction between dark area detail loss and global illumination imbalance in the traditional method through double-branch cooperative processing in the pre-enhancement stage.

Description

Technical Field

[0001] This invention belongs to the field of image processing, specifically a method for enhancing low-light highway images based on hierarchical decomposition and adaptive fusion. Background Technology

[0002] In low-light highway scenarios, images often exhibit typical defects such as underexposure, localized overexposure, strong light interference, and imbalanced contrast due to the complex ambient lighting conditions. Especially when the overall road illumination is insufficient, images simultaneously contain large areas of low-brightness background and areas of strong light interference from multiple sources, including vehicle headlights, road lighting, and traffic signs, creating complex lighting distribution characteristics. This type of lighting degradation not only significantly reduces the visual discernibility of road surface details, traffic targets, and road markings, but also leads to a significant performance degradation in downstream computer vision tasks such as object detection and behavior analysis.

[0003] Traditional image enhancement methods often employ global brightness enhancement or histogram adjustment strategies, but these fail to effectively distinguish the essential differences between various illuminated areas in an image, easily leading to contradictory problems such as overexposure in dark areas and loss of detail in bright areas. Under low-light or extreme lighting conditions, image illumination distribution exhibits highly uneven characteristics, with underexposed areas and areas of strong light interference often coexisting within the same frame, further exacerbating the technical challenges of image enhancement. In particular, areas of "light effects" such as glare and flooding caused by direct light sources exhibit strong brightness interference and low texture contrast, making it difficult for conventional enhancement strategies to effectively suppress them.

[0004] Furthermore, the illumination degradation mechanism of low-light highway images is complex, influenced not only by ambient lighting conditions but also by the nonlinear operational effects of imaging processes such as camera white balance, color mapping, and gamma correction. This complex degradation mechanism makes it difficult for a single exposure adjustment method to simultaneously restore details in dark areas and suppress highlight interference, often resulting in enhancements accompanied by texture blurring, color distortion, or inconsistent area enhancement.

[0005] In summary, existing technologies still lack a method for enhancing low-light images of highways that can simultaneously address both local structure and global illumination adjustment capabilities. There is an urgent need to propose an image processing scheme that can effectively handle both strong light interference and low-light areas. Summary of the Invention

[0006] This invention provides a low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion to address the shortcomings of existing technologies.

[0007] This invention is achieved through the following technical solution: A low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion includes the following steps: S1: Based on the highway video cloud platform, capture video streams of continuous low-light segments; S2: Construct a multi-stage collaborative image enhancement model; S3: Train an image enhancement model using the dataset; S4: Using the trained image enhancement model, detection is performed on real-time monitoring cameras, taking into account both objective indicators and subjective evaluation, and verifying the practicality of the algorithm for the core task of low-light highways.

[0008] The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion described above, specifically involves the following steps in S1: acquiring continuous low-light video streams during low-light periods using a highway video cloud platform. This process covers various typical low-light scenarios to ensure the comprehensiveness of the dataset, including different weather conditions such as sunny, rainy, foggy, and snowy conditions. It also covers scenarios such as low-light environments without moonlight, areas with uneven street lighting, and glare interference from oncoming vehicle lights. The original video streams are then denoised, frame-selected, and have their resolution unified to ultimately form a low-light image dataset containing 10,000 frames.

[0009] The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion described above involves the following steps in constructing the image enhancement model S2: A pre-enhancement module performs preliminary brightness and color correction on the input low-light highway image to obtain a pre-enhanced image that balances local details and global exposure. Subsequently, the pre-enhanced image is input into a hierarchical decomposition module, which decomposes it into a reflection layer, a shadow layer, and a lighting effect layer. Next, a hierarchical optimization module selectively enhances and suppresses the reflection layer, shadow layer, and lighting effect layer, respectively. Finally, an adaptive fusion module performs adaptive weighted fusion of the optimized layers to output the final enhanced image.

[0010] The low-light highway image enhancement method described above, based on hierarchical decomposition and adaptive fusion, includes a pre-enhancement module that performs initial enhancement on the input low-light highway image to ensure more stable hierarchical structure and more balanced feature distribution during subsequent decomposition. This pre-enhancement module employs parallel processing of local and global branches. The processing flow of the local branch of the pre-enhancement module is as follows: Step (1): Local feature extraction: Extract multi-scale feature maps F from the input image I using a depthwise separable convolutional network; Step (2): Mask weighted normalization: Obtain the road / guardrail mask Mr using the semantic segmentation network, and perform weighted normalization on the feature map F based on the mask to obtain the recalibrated feature map. : in, These are pixel coordinates; For channel indexing; For channel In pixels The original eigenvalues; The channel mean. and Represents the height and width of the image. This represents the total number of pixels in the image. The standard deviation of the channel; It is the numerical stability constant; A semantic mask for roads / guardrails; This is the mask magnification factor; This is the channel offset; Step (3): Brightness and color correction prediction: In feature map Predicting pixel-level brightness gain through convolution operations : in, These are 1×1 convolution kernels, and * indicates a convolution operation. For bias terms; In feature map Predict the three-channel color bias vector using convolution operations. : in These are 1×1 convolution kernels, and * indicates a convolution operation. For bias terms; In key areas such as license plates and lanes, masks generated by license plate detection are used. To suppress color bias, the calculation formula is: in The color bias suppression coefficient for the mask region; Indicates pixel-by-pixel multiplication; The input image I is pixel-level corrected based on the prediction results according to the following formula to obtain the locally enhanced image. : ; The processing flow of the global branch of the pre-enhancement module is as follows: Step (1): Global feature extraction: Lightweight convolutional network is used to extract features from the input image I, and global descriptive vector f is obtained through global average pooling; Step (2): Initial color correction matrix regression: Based on vector f, the initial color correction matrix is ​​obtained through regression using a fully connected network. Its calculation form is: in This indicates that the matrix can be expanded into a 9-dimensional column vector. This is the weight matrix. It is the bias vector; Step (3): Matrix constraint optimization: In the initial matrix Based on this, nonnegativity constraints and near-diagonal sparsity constraints are applied to obtain the final color correction matrix. ; in, The color matrix obtained by sampling the key regions of the input image; The corresponding target color is estimated from white balance or a high-quality reference. Let be a nonnegative and norm-restricted feasible region. For the upper realm, Based on global description vector The initial matrix obtained from the regression; A is a non-diagonal element of A; , The regularization coefficient; Step (4): Gamma-parameter regression: based on vector Simultaneously, regression yields the global gamma parameter γ; Step (5): Global calibration: using a color calibration matrix The input image is then globally corrected using the gamma coefficient γ according to the following formula to obtain the globally enhanced image. : According to the formula Fusion of locally enhanced images With global enhancement image Obtain the pre-enhanced image .

[0011] The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion, as described above, specifically includes the following operation of the hierarchical decomposition module: Step (1): Pre-enhanced image It can be broken down into three parts: Where R(x) represents the reflection layer, which preserves the texture and structure of the object; S(x) represents the shadow layer, which represents the distribution of dark areas in the scene; and G(x) represents the lighting layer, which includes glare and halo. Step (2): Decompose the network hierarchically and implement unsupervised loss constraints. arrive The decomposition is performed with the following orthogonality and energy distribution conditions: The first term is the reconstruction constraint; the second term is the decoupling constraint. The third term represents the inner product correlation; the third term represents the energy allocation constraint. The desired energy ratio; , These are the weighting coefficients.

[0012] The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion, as described above, wherein the hierarchical customized optimization module optimizes the shadow layer S, the light effect layer G, and the reflection layer; The shadow layer S optimization operation is as follows: Apply adaptive gamma highlighting: in, For pixel-adaptive gain, a soft film is used to reduce its value in the headlight area; Pixel-adaptive gamma value; This is the upper limit for brightness, used to prevent whitening. in , For vehicle headlights; in, As a baseline gamma, tanh is the gradient sensitivity coefficient; ) is the hyperbolic tangent function; This is the scaling factor; For the input image In pixels Normalized gradient magnitude at point, These are the sample mean and standard deviation, respectively. The optimization operation of the light effect layer G is as follows: For the light effect layer G at scale set The description vector is obtained by upsampling, downsampling, and channel average pooling. in, This represents the light-effect layer G at scale e; This is the set of pixels used for pooling at scale e; This represents the value of channel c at pixel x at scale e; It is the channel description vector for scale e; Channel compression weights are obtained from a shared MLP: in,[ ] indicates vector concatenation; For ReLU functions; for function; To share the first layer weights and biases of the MLP; To share the second layer weights and biases of the MLP; Introducing a region sparse smoothing regularization: in This is related to glare suppression losses; Non-light zone indicators have soft weights; This represents the sum of the channel dimensions; This represents the composite magnitude of the cross-channel gradient; , These are the initial hyperparameters; The optimization operation of the reflective layer is as follows: Constraints are imposed by combining structural consistency and high-frequency consistency loss: in, One loss in preserving structure and detail in the reflective layer; This indicates the pre-trained visual network's first... Feature mapping of layers; For the selected set of layers; Layer weights; For high-frequency operators, Standard deviation Gaussian filtering; This is a high-frequency consistency weight.

[0013] The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion described above, wherein the specific operation of the adaptive fusion module is as follows: An initial mask is generated based on the vehicle headlight monitoring frame. The PDE is solved by anisotropic diffusion. And take steady state Used for subsequent prior and fusion. in The mask function evolves over time during the diffusion process, with an initial value of Ultimately The steady-state value at time t; Divergence operator, used to calculate the divergence of flow rate; Represents the gradient vector Projection component in the direction parallel to the edge; Represents the gradient vector Projection component in the direction perpendicular to the edge; express The gradient vector; This is a function of the edge-sensitive diffusion coefficient; This is an anisotropy intensity adjustment factor; Fusion optimization objective function: in , , This represents the pixel contribution of the three layers. Indicates scalar Expanded to three channels; weighting and satisfy ; For total variation regularization; fusion weights are induced by mask energy and smoothing terms: Potential function , ,in The value of the soft mask for the vehicle headlight after anisotropic diffusion. These are calibration coefficients.

[0014] The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion described above involves the following training operation for the S3 image enhancement model: the ratio of training set, test set, and validation set in the dataset is 7:2:1. By comparing the training and testing effects of the image enhancement model with other models, the model with the best performance is selected. The loss function adopts a multi-task joint loss function, using the Adam optimizer, and training is performed on an NVIDIA GeForce RTX 2080 Ti GPU. The validation set loss is used as the monitoring metric, and an early stopping strategy is set to terminate training early when the validation loss does not decrease for 10 consecutive rounds.

[0015] The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion described above includes the following steps in S4: The trained image enhancement model was deployed to a real-time highway monitoring camera to perform real-time enhancement processing on low-light video streams. The model performance was comprehensively tested through a three-pronged approach: objective quantitative evaluation, subjective visual evaluation, and downstream task validation. The objective metrics mentioned are the peak signal-to-noise ratio and structural similarity of the images before and after enhancement, which are used to quantitatively evaluate the enhancement effect. The subjective evaluation involved organizing 50 testers to conduct blind evaluations of the images before and after enhancement. The scoring dimensions included "clarity of details in dark areas", "overexposure suppression effect of headlights" and "overall visual comfort", and the average score was taken as the subjective indicator. The downstream task verification involves running highway core task models, such as lane detection models, traffic sign recognition models, and pedestrian detection models, on the enhanced images. The task completion rates before and after enhancement are compared to verify the algorithm's practicality improvement for actual business.

[0016] The advantages of this invention are: This invention, through dual-branch collaborative processing in the pre-enhancement stage, effectively resolves the contradiction between the loss of dark area details and global illumination imbalance in traditional methods, compared to the compromise optimization mode that relies solely on global histogram equalization or single brightness adjustment. In the hierarchical decomposition stage, an unsupervised three-branch network explicitly decouples the pre-enhanced image into a reflection layer R, a shadow layer S, and a lighting effect layer G, and specifically implements: adaptive gamma brightening of the shadow layer S to restore dark area details; cross-scale channel attention dynamic compression of the lighting effect layer G to accurately suppress headlight glare and streetlight spotting; and applying structural consistency and high-frequency consistency constraints to the reflection layer R to maintain the geometric structure of road edges, lane lines, etc. The continuity and sharpness of the structure are maintained; the adaptive fusion stage uses an anisotropic diffusion-generated headlight mask to dynamically reduce the weight of the reflection layer and suppress the light effect layer in the headlight area, avoiding the problem of local over-enhancement or under-enhancement caused by the global uniform weight in traditional fusion methods; through the unsupervised learning mechanism of the hierarchical decomposition module and the lightweight design of the adaptive transformation module, the number of model parameters is controlled within 8.7M, and the inference latency is reduced to 15ms / frame, which fully meets the real-time requirements of highway monitoring cameras. It can be directly deployed in existing monitoring systems without additional hardware upgrades, providing an efficient and reliable image enhancement solution for low-light traffic safety monitoring. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the present invention; Figure 2 This is a flowchart of the pre-enhancement module of the present invention; Figure 3 This is a flowchart of the hierarchical decomposition module of the present invention; Figure 4 This is a flowchart of the adaptive fusion process of the present invention; Figure 5 This is a flowchart illustrating the process of generating the automotive headlight soft mask according to the present invention. Figure 6 This is a flowchart of the image enhancement model training process of the present invention; Figure 7 This is a preview image before the dark lighting was enhanced. Figure 8 This is an image showing the effect after enhancing the low-light effect. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, a low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion includes the following steps: S1: Based on the highway video cloud platform, capture video streams of continuous low-light segments; S2: Construct a multi-stage collaborative image enhancement model; S3: Train an image enhancement model using the dataset; S4: Using the trained image enhancement model, detection is performed on real-time monitoring cameras, taking into account both objective indicators and subjective evaluation, and verifying the practicality of the algorithm for the core task of low-light highways.

[0021] Preferably, the specific operation of collecting continuous low-light video streams in S1 of this embodiment is as follows: Based on the highway video cloud platform, continuous video streams during low-light periods are collected, covering multiple typical low-light scenarios to ensure the comprehensiveness of the dataset, including different meteorological environments such as sunny, rainy, foggy, and snowy, covering low-light environments without moonlight, areas with uneven street lighting, and scenarios with glare interference from oncoming vehicle lights, and the original video stream is denoised, frame-filtered, and resolution-unified, finally forming a low-light image dataset containing 10,000 frames.

[0022] Preferably, the specific operation of constructing the image enhancement model S2 described in this embodiment is as follows: the pre-enhancement module performs preliminary brightness and color correction on the input low-light highway image to obtain a pre-enhanced image that balances local details and global exposure; then, the pre-enhanced image is input into the layer decomposition module to decompose it into a reflection layer, a shadow layer, and a light effect layer; next, the layer optimization module performs targeted enhancement and suppression on the reflection layer, the shadow layer, and the light effect layer respectively; finally, the adaptive fusion module performs adaptive weighted fusion of the optimized images of each layer to output the final enhanced image.

[0023] like Figure 2 As shown, preferably, the pre-enhancement module described in this embodiment is used to perform preliminary enhancement on the input low-light highway image to ensure more stable layering and more balanced feature distribution in the subsequent decomposition process. The pre-enhancement module adopts parallel processing of local branches and global branches: The processing flow of the local branch of the pre-enhancement module is as follows: Step (1): Local feature extraction: Extract multi-scale feature maps F from the input image I using a depthwise separable convolutional network; Step (2): Mask weighted normalization: Obtain the road / guardrail mask Mr using the semantic segmentation network, and perform weighted normalization on the feature map F based on the mask to obtain the recalibrated feature map. : in, These are pixel coordinates; For channel indexing; For channel In pixels The original eigenvalues; The channel mean. and Represents the height and width of the image. This represents the total number of pixels in the image. The standard deviation of the channel; It is the numerical stability constant; A semantic mask for roads / guardrails; This is the mask magnification factor; This is the channel offset; Step (3): Brightness and color correction prediction: In feature map Predicting pixel-level brightness gain through convolution operations : in, These are 1×1 convolution kernels, and * indicates a convolution operation. For bias terms; In feature map Predict the three-channel color bias vector using convolution operations. : in These are 1×1 convolution kernels, and * indicates a convolution operation. For bias terms; In key areas such as license plates and lanes, masks generated by license plate detection are used. To suppress color bias, the calculation formula is: in The color bias suppression coefficient for the mask region; Indicates pixel-by-pixel multiplication; The input image I is pixel-level corrected based on the prediction results according to the following formula to obtain the locally enhanced image. : ; The processing flow of the global branch of the pre-enhancement module is as follows: Step (1): Global feature extraction: Lightweight convolutional network is used to extract features from the input image I, and global descriptive vector f is obtained through global average pooling; Step (2): Initial color correction matrix regression: Based on vector f, the initial color correction matrix is ​​obtained through regression using a fully connected network. Its calculation form is: in This indicates that the matrix can be expanded into a 9-dimensional column vector. This is the weight matrix. It is the bias vector; Step (3): Matrix constraint optimization: In the initial matrix Based on this, nonnegativity constraints and near-diagonal sparsity constraints are applied to obtain the final color correction matrix. ; in, The color matrix is ​​obtained by sampling the input image in key regions (such as near headlights or road areas); The corresponding target color is estimated from white balance or a high-quality reference. Let be a nonnegative and norm-restricted feasible region. For the upper realm, Based on global description vector The initial matrix obtained from the regression; A is a non-diagonal element of A; , The regularization coefficient; Step (4): Gamma-parameter regression: based on vector Simultaneously, regression yields the global gamma parameter γ; Step (5): Global calibration: using a color calibration matrix The input image is then globally corrected using the gamma coefficient γ according to the following formula to obtain the globally enhanced image. : According to the formula Fusion of locally enhanced images With global enhancement image Obtain the pre-enhanced image .

[0024] like Figure 3 As shown, preferably, the specific operation of the hierarchical decomposition module described in this embodiment is as follows: Step (1): Pre-enhanced image It can be broken down into three parts: Where R(x) represents the reflection layer, which preserves the texture and structure of the object; S(x) represents the shadow layer, which represents the distribution of dark areas in the scene; and G(x) represents the lighting layer, which includes glare and halo. Step (2): Decompose the network hierarchically and implement unsupervised loss constraints. arrive The decomposition is performed with the following orthogonality and energy distribution conditions: The first term is the reconstruction constraint; the second term is the decoupling constraint. The third term represents the inner product correlation; the third term represents the energy allocation constraint. The desired energy ratio; , These are the weighting coefficients.

[0025] Preferably, the layered customization optimization module described in this embodiment optimizes the shadow layer S, the light effect layer G, and the reflection layer; like Figure 5 As shown, the optimization operation of the shadow layer S is as follows: Apply adaptive gamma highlighting: in, For pixel-adaptive gain, a soft film is used to reduce its value in the headlight area; Pixel-adaptive gamma value; This is the upper limit for brightness, used to prevent whitening. in , For vehicle headlights; in, As a baseline gamma, tanh is the gradient sensitivity coefficient; ) is the hyperbolic tangent function; This is the scaling factor; For the input image In pixels Normalized gradient magnitude at point, These are the sample mean and standard deviation, respectively. The optimization operation of the light effect layer G is as follows: For the light effect layer G at scale set The description vector is obtained by upsampling, downsampling, and channel average pooling. in, This represents the light-effect layer G at scale e; This is the set of pixels used for pooling at scale e; This represents the value of channel c at pixel x at scale e; It is the channel description vector for scale e; Channel compression weights are obtained from a shared MLP: in,[ ] indicates vector concatenation; For ReLU functions; for function; To share the first layer weights and biases of the MLP; To share the second layer weights and biases of the MLP; Introducing a region sparse smoothing regularization: in This is related to glare suppression losses; Non-light zone indicators have soft weights; This represents the sum of the channel dimensions; This represents the composite magnitude of the cross-channel gradient; , These are the initial hyperparameters; The optimization operation of the reflective layer is as follows: Constraints are imposed by combining structural consistency and high-frequency consistency loss: in, One loss in preserving structure and detail in the reflective layer; This indicates the pre-trained visual network's first... Feature mapping of layers; For the selected set of layers; Layer weights; For high-frequency operators, Standard deviation Gaussian filtering; This is a high-frequency consistency weight.

[0026] like Figure 4 As shown, preferably, the specific operation of the adaptive fusion module described in this embodiment is as follows: An initial mask is generated based on the vehicle headlight monitoring frame. The PDE is solved by anisotropic diffusion. And take steady state Used for subsequent prior and fusion. in The mask function evolves over time during the diffusion process, with an initial value of Ultimately The steady-state value at time t; Divergence operator, used to calculate the divergence of flow rate; Represents the gradient vector Projection component in the direction parallel to the edge; Represents the gradient vector Projection component in the direction perpendicular to the edge; express The gradient vector; This is a function of the edge-sensitive diffusion coefficient; This is an anisotropy intensity adjustment factor; Fusion optimization objective function: in , , This represents the pixel contribution of the three layers. Indicates scalar Expanded to three channels; weighting and satisfy ; For total variation regularization; fusion weights are induced by mask energy and smoothing terms: Potential function , ,in The value of the soft mask for the vehicle headlight after anisotropic diffusion. These are calibration coefficients.

[0027] like Figure 6 As shown, preferably, the training operation of the image enhancement model S3 described in this embodiment is as follows: the ratio of the training set, test set and validation set of the dataset is 7:2:1. By comparing the training and testing effects of the image enhancement model with other models, the model with the best performance is selected. The loss function adopts a multi-task joint loss function, and the Adam optimizer is used. The training is carried out on an NVIDIA GeForce RTX 2080 Ti GPU. The validation set loss is used as the monitoring indicator, and an early stopping strategy is set. When the validation loss does not decrease for 10 consecutive rounds, the training is terminated in advance.

[0028] Preferably, the model training and testing in S4 of this embodiment includes the following steps: The trained image enhancement model was deployed to a real-time highway monitoring camera to perform real-time enhancement processing on low-light video streams. The model performance was comprehensively tested through a three-pronged approach: objective quantitative evaluation, subjective visual evaluation, and downstream task validation. The objective metrics mentioned are the peak signal-to-noise ratio and structural similarity of the images before and after enhancement, which are used to quantitatively evaluate the enhancement effect. The subjective evaluation involved organizing 50 testers to conduct blind evaluations of the images before and after enhancement. The scoring dimensions included "clarity of details in dark areas", "overexposure suppression effect of headlights" and "overall visual comfort", and the average score was taken as the subjective indicator. The downstream task verification involves running highway core task models, such as lane detection models, traffic sign recognition models, and pedestrian detection models, on the enhanced images. The task completion rates before and after enhancement are compared to verify the algorithm's practicality improvement for actual business.

[0029] Effect Verification: To verify the effectiveness of this invention, a low-light monitoring image of a highway containing severely dark areas and strong glare was selected for testing. Subjective effect: such as Figure 7 As shown, the original input image has blurred dark areas and overexposed headlight areas forming obvious halos; after enhancement by the method of this invention (e.g. Figure 8 As shown in the image, details in the dark areas are clearly discernible, the halo range of the headlights is significantly reduced and the brightness is lowered, key structures are not deformed, and overall visual comfort is high.

[0030] Objective metrics: As shown in Table 1, compared with representative methods in the field such as multi-scale Retinex, RetinexDIP, and SCI, the method proposed in this invention performs best in both peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) metrics. Compared with the original image, the PSNR and structural similarity are improved by 10.423 dB and 0.537, respectively. Compared with the well-performing SCI, the PSNR and structural similarity are improved by 1.894 dB and 0.061, respectively, thus improving the brightness performance and structural fidelity of the image.

[0031] Table 1 Objective indicators Original image Multiscale Retinex RetinexDIP SCI This invention Peak signal-to-noise ratio (in dB) 9.718 14.326 16.872 18.247 20.141 Structural similarity 0.191 0.454 0.612 0.667 0.728 In summary, this invention, through multi-stage collaborative processing of pre-enhancement, hierarchical decomposition and optimization, and adaptive fusion, combined with unsupervised learning mechanisms and lightweight design, significantly improves the details of dark areas in low-light images of highways while effectively suppressing glare interference, thus significantly improving image quality and providing an efficient and reliable image enhancement solution for low-light traffic safety monitoring.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A low-illumination highway image enhancement method based on hierarchical decomposition and adaptive fusion, characterized by: Includes the following steps: S1: Based on the highway video cloud platform, capture video streams of continuous low-light segments; S2: Construct a multi-stage collaborative image enhancement model; S3: Train an image enhancement model using the dataset; S4: Using the trained image enhancement model, detection is performed on real-time monitoring cameras, taking into account both objective indicators and subjective evaluation, and verifying the practicality of the algorithm for the core task of low-light highways. 2.The low-illumination highway image enhancement method based on hierarchical decomposition and adaptive fusion of claim 1, wherein: The specific operation of acquiring continuous low-light video streams in S1 is as follows: Based on the highway video cloud platform, continuous video streams during low-light periods are acquired, covering multiple typical low-light scenarios to ensure the comprehensiveness of the dataset. This includes different meteorological environments such as sunny, rainy, foggy, and snowy conditions, as well as scenarios such as low-light environments without moonlight, areas with uneven street lighting, and glare interference from oncoming vehicle lights. The original video streams are then denoised, frame-filtered, and resolution-unified to ultimately form a low-light image dataset containing 10,000 frames. 3.The low-illumination highway image enhancement method based on hierarchical decomposition and adaptive fusion of claim 1, characterized in that: The specific operation of constructing the image enhancement model of S2 is as follows: the pre-enhancement module performs preliminary brightness and color correction on the input low-light highway image to obtain a pre-enhanced image that takes into account both local details and global exposure balance; then, the pre-enhanced image is input into the layer decomposition module to decompose it into a reflection layer, a shadow layer, and a light effect layer; next, the layer optimization module performs targeted enhancement and suppression on the reflection layer, the shadow layer, and the light effect layer respectively; finally, the adaptive fusion module performs adaptive weighted fusion of the optimized images of each layer to output the final enhanced image.

4. The low-illumination highway image enhancement method based on hierarchical decomposition and adaptive fusion of claim 3, characterized in that: The pre-enhancement module is used to perform initial enhancement on the input low-light highway image, ensuring more stable layering and more balanced feature distribution in subsequent decomposition processes. The pre-enhancement module employs parallel processing of local and global branches. The processing flow of the local branch of the pre-enhancement module is as follows: Step (1): Local feature extraction: Extract multi-scale feature maps F from the input image I using a depthwise separable convolutional network; Step (2): Mask weighted normalization: obtain the road / barrier mask Mrusing the semantic segmentation network, and perform weighted normalization processing on the feature map F based on the mask to obtain the re-calibration feature map : wherein, is the pixel coordinate; is the channel index; is the channel at the pixel original feature value; is the channel mean, and represent the height and width of the image, is the total number of pixels of the image; is the channel standard deviation; is the numerical stability constant; is the road / barrier semantic mask; is the mask amplification coefficient; is the channel offset; Step (3): Brightness and color correction prediction: In the feature map Predicting pixel-level luminance gain by convolution operation : wherein, are 1x1 convolution kernels, respectively, * denotes a convolution operation, is a bias term; In the feature map Predicting a three-channel color bias vector by a convolution operation : in These are 1×1 convolution kernels, and * indicates a convolution operation. For bias terms; In key areas such as license plates and lanes, masks generated by license plate detection are used. To suppress color bias, the calculation formula is: in The color bias suppression coefficient for the mask region; Indicates pixel-by-pixel multiplication; The input image I is pixel-level corrected based on the prediction results according to the following formula to obtain the locally enhanced image. : ； The processing flow of the global branch of the pre-enhancement module is as follows: Step (1): Global feature extraction: Lightweight convolutional network is used to extract features from the input image I, and global descriptive vector f is obtained through global average pooling; Step (2): Initial color correction matrix regression: Based on vector f, the initial color correction matrix is ​​obtained through regression using a fully connected network. Its calculation form is: in This indicates that the matrix can be expanded into a 9-dimensional column vector. This is the weight matrix. It is the bias vector; Step (3): Matrix constraint optimization: In the initial matrix Based on this, nonnegativity constraints and near-diagonal sparse constraints are applied to obtain the final color correction matrix. ; in, The color matrix obtained by sampling the key regions of the input image; The corresponding target color is estimated from white balance or a high-quality reference. Let be a nonnegative and norm-restricted feasible region. For the upper realm, Based on global description vector The initial matrix obtained from the regression; A is a non-diagonal element of A; , The regularization coefficient; Step (4): Gamma-parameter regression: based on vectors Simultaneously, regression yields the global gamma parameter γ; Step (5): Global calibration: using a color calibration matrix The input image is then globally corrected using the gamma coefficient γ according to the following formula to obtain the globally enhanced image. : According to the formula Fusion of locally enhanced images With global enhancement image Obtain the pre-enhanced image .

5. The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion according to claim 3, characterized in that: The specific operation of the hierarchical decomposition module is as follows: Step (1): Pre-enhanced image It can be broken down into three parts: Where R(x) represents the reflection layer, which preserves the texture and structure of the object; S(x) represents the shadow layer, which represents the distribution of dark areas in the scene; and G(x) represents the lighting layer, which includes glare and halo. Step (2): Decompose the network hierarchically and implement unsupervised loss constraints. arrive The decomposition is performed with the following orthogonality and energy distribution conditions: The first term is the reconstruction constraint; the second term is the decoupling constraint. The third term represents the inner product correlation; the third term represents the energy allocation constraint. The desired energy ratio; , These are the weighting coefficients.

6. The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion according to claim 1, characterized in that: The layered customization optimization module optimizes the shadow layer S, the light effect layer G, and the reflection layer. The shadow layer S optimization operation is as follows: Apply adaptive gamma highlighting: in, For pixel-adaptive gain, a soft film is used to reduce its value in the headlight area; Pixel-adaptive gamma value; This is the upper limit for brightness, used to prevent whitening. in , For vehicle headlights; in, As a baseline gamma, tanh is the gradient sensitivity coefficient; ) is the hyperbolic tangent function; This is the scaling factor; For the input image In pixels Normalized gradient magnitude at point, These are the sample mean and standard deviation, respectively. The optimization operation of the light effect layer G is as follows: For the light effect layer G at scale set The description vector is obtained by upsampling, downsampling, and channel average pooling. in, This represents the light-effect layer G at scale e; This is the set of pixels used for pooling at scale e; This represents the value of channel c at pixel x at scale e; It is the channel description vector for scale e; Channel compression weights are obtained from a shared MLP: in,[ ] indicates vector concatenation; For ReLU functions; for function; To share the first layer weights and biases of the MLP; To share the second layer weights and biases of the MLP; Introducing a region sparse smoothing regularization: in This is related to glare suppression losses; Non-light zone indicators have soft weights; This represents the sum of the channel dimensions; This represents the composite magnitude of the cross-channel gradient; , These are the initial hyperparameters; The optimization operation of the reflective layer is as follows: Constraints are imposed by combining structural consistency and high-frequency consistency loss: in, One loss in preserving structure and detail in the reflective layer; This indicates the pre-trained visual network's first... Feature mapping of layers; For the selected set of layers; Layer weights; For high-frequency operators, Standard deviation Gaussian filtering; This is a high-frequency consistency weight.

7. The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion according to claim 3, characterized in that: The specific operation of the adaptive fusion module is as follows: An initial mask is generated based on the vehicle headlight monitoring frame. The PDE is solved by anisotropic diffusion. And take steady state Used for subsequent prior and fusion. in The mask function evolves over time during the diffusion process, with an initial value of Ultimately The steady-state value at time t; Divergence operator, used to calculate the divergence of flow rate; Represents the gradient vector Projection component in the direction parallel to the edge; Represents the gradient vector Projection component in the direction perpendicular to the edge; express The gradient vector; This is a function of the edge-sensitive diffusion coefficient; This is an anisotropy intensity adjustment factor; Fusion optimization objective function: in , , This represents the pixel contribution of the three layers. Indicates scalar Expanded to three channels; weighting and satisfy ; For total variation regularization; fusion weights are induced by mask energy and smoothing terms: Potential function , ,in The value of the soft mask for the vehicle headlight after anisotropic diffusion. These are calibration coefficients.

8. The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion according to claim 1, characterized in that: The training operation of the S3 image enhancement model is as follows: the ratio of training set, test set and validation set of the dataset is 7:2:

1. By comparing the training and testing effects of the image enhancement model with other models, the model with the best performance is selected. The loss function adopts a multi-task joint loss function, and the Adam optimizer is used. The training is carried out on an NVIDIA GeForce RTX 2080 Ti GPU. The validation set loss is used as the monitoring indicator. An early stopping strategy is set. When the validation loss does not decrease for 10 consecutive rounds, the training is terminated in advance.

9. The low-light highway image enhancement method based on hierarchical decomposition and adaptive fusion according to claim 1, characterized in that: The model training and testing in S4 includes the following steps: The trained image enhancement model was deployed to a real-time highway monitoring camera to perform real-time enhancement processing on low-light video streams. The model performance was comprehensively tested through a three-pronged approach: objective quantitative evaluation, subjective visual evaluation, and downstream task validation. The objective metrics mentioned are the peak signal-to-noise ratio and structural similarity of the images before and after enhancement, which are used to quantitatively evaluate the enhancement effect. The subjective evaluation involved organizing 50 testers to conduct blind evaluations of the images before and after enhancement. The scoring dimensions included "clarity of details in dark areas", "overexposure suppression effect of headlights" and "overall visual comfort", and the average score was taken as the subjective indicator. The downstream task verification involves running highway core task models, such as lane detection models, traffic sign recognition models, and pedestrian detection models, on the enhanced images. The task completion rates before and after enhancement are compared to verify the algorithm's practicality improvement for actual business.

Images