An intelligent traffic scene image enhancement processing method and device

Abstract

The present application relates to the technical field of image defogging, in particular to a kind of enhancement processing method and device of fog image in intelligent traffic scene.The present application extracts image multi-scale basic feature, generates physical perception enhancement feature by introducing atmospheric scattering prior in shallow layer and fusing to obtain single-layer main feature, which is consistent with the imaging mechanism of fog and haze;Multi-scale aggregation and adaptive attention weighting are performed on each layer main feature, and hierarchical information is fused to obtain multi-layer enhancement side feature.Single-layer main feature is sampled and multi-scale corrected to obtain overall main feature, and the two types of features are fused to generate comprehensive features for image reconstruction.The present application fully utilizes the interaction between physical prior and multi-level features, enriches the multi-scale details of the image, optimizes the feature correction effect in the fog area, and reconstructs a clear fog-free traffic image to improve the execution accuracy of downstream traffic visual processing tasks.

Description

Technical Field

[0001] This invention relates to the field of image dehazing technology, and in particular to an enhancement processing method and apparatus for foggy images in intelligent transportation scenarios. Background Technology

[0002] In the transportation sector, the accurate acquisition and processing of visual information is crucial for the efficient operation of autonomous driving, intelligent transportation systems, and other applications. However, hazy weather reduces light propagation, significantly decreasing the clarity of images captured by cameras. This results in blurred outlines of key targets such as roads, vehicles, and pedestrians, directly interfering with downstream visual tasks such as target recognition, behavior analysis, and traffic flow monitoring.

[0003] Current mainstream image dehazing enhancement methods rely solely on network fitting of shallow texture information in images for modeling, without incorporating the physical laws of atmospheric scattering imaging as constraints, resulting in a lack of physical prior support for the models. At the same time, the interaction and correlation between features at different levels are weak during the feature extraction stage, the mining of multi-scale detail information in the image is insufficient, and the feature correction accuracy is inadequate during the decoding stage. As a result, the reconstructed fog-free images still suffer from defects such as loss of target details and obvious fog residue, which cannot meet the high-precision operation requirements of intelligent transportation visual processing tasks in foggy weather.

[0004] Therefore, overcoming the shortcomings of the existing technology is an urgent problem to be solved in this technical field. Summary of the Invention

[0005] The technical problem this invention aims to solve is that existing intelligent transportation fog image enhancement methods rely solely on network learning of shallow image textures without considering the physical laws of fog imaging. This results in insufficient feature interaction between layers, inadequate extraction of multi-scale information, poor decoding and correction accuracy, and ultimately, blurred target details in the reconstructed image, making it difficult to meet the accuracy requirements of downstream visual tasks such as traffic target recognition and traffic flow monitoring in foggy weather.

[0006] The present invention adopts the following technical solution: Firstly, a method for enhancing foggy images in intelligent transportation scenarios is provided, including: Acquire foggy images collected by traffic data acquisition equipment, and extract multi-scale basic image features from the foggy images in layers; in the shallow feature extraction stage, introduce atmospheric scattering physical priors to obtain physical perception enhancement features, and fuse the basic image features of the same level with the physical perception enhancement features to obtain single-layer backbone features of each level; Multi-scale feature aggregation and adaptive attention weighting are performed on the single-layer backbone features of each layer, and the weighting results of all layers are fused to obtain multi-layer enhanced side features. The single-layer backbone features of each level are sampled and multi-scale fusion correction is performed to obtain the corrected overall backbone features; the overall backbone features are fused with multi-layer enhanced side features to obtain comprehensive features. Based on the comprehensive features, a fog-free traffic image is reconstructed and output. The fog-free traffic image is then input into a downstream traffic vision processing device to perform traffic vision processing tasks.

[0007] Preferably, the step of introducing atmospheric scattering physical priors to obtain physical perception enhancement features in the shallow feature extraction stage, and fusing the basic image features at the same level with the physical perception enhancement features to obtain single-layer backbone features at each level, specifically includes: Construct at least three parallel feature extraction branches to generate approximate transmittance features, approximate atmospheric light features, and joint features of transmittance and atmospheric light, respectively. The physical perception enhancement features are obtained by performing pixel-by-pixel multiplication, pixel-by-pixel subtraction, and pixel-by-pixel addition on the transmittance approximation features, atmospheric light approximation features, and transmittance and atmospheric light joint features based on the atmospheric scattering model. The obtained physical perception enhancement features and the basic image features of the same level are input into the selective kernel fusion unit to complete the fusion of the two features to generate the single-layer backbone features of the corresponding level.

[0008] Preferred, approximate transmittance characteristics Represented as: ; Atmospheric light approximation characteristics Represented as: ; Combined characteristics of transmittance and atmospheric light Represented as: ; Physical perception enhancement features Represented as: ; in, As basic image features at the same level, Indicates lightweight convolution. Indicates nonlinear activation. This represents compressing the eigenvalue range to [0,1] using global average pooling. This indicates sampling of the global mean of the entire feature map. It represents the Hadamardi (or Hadama) stack.

[0009] Preferably, the step of performing multi-scale feature aggregation and adaptive attention weighting on the single-layer backbone features of each layer, and fusing the weighting results of all layers to obtain multi-layer enhanced side features, specifically includes: After batch regularization preprocessing of the single-layer backbone features of the current level, different convolution operations are used to extract texture sub-features of different granularities. The texture sub-features are added and fused pixel by pixel and then processed by convolution to generate aggregated features that aggregate multi-scale texture information, thus completing multi-scale feature aggregation. The single-layer backbone features of the original input are extracted into branches, and adaptive attention weights are generated through convolution and activation operations. The adaptive attention weights are multiplied pixel by pixel with the aggregated features to complete the feature filtering and weighting, and the weighted texture features are obtained. The weighted texture features are convolved to adjust the channels, and then added to the pixel-by-pixel residual of the original input single-layer backbone features to obtain the output features. The neighboring layer backbone features of the output of adjacent layers are obtained, and the neighboring layer backbone features are upsampled or downsampled to complete the scale alignment. The scale-aligned neighboring layer backbone features are added to the output features pixel by pixel and fused to generate the layer enhancement features corresponding to the current layer. After obtaining the corresponding layer enhancement features for each layer in sequence, all layer enhancement features are summed pixel by pixel according to preset weights to obtain the multi-layer enhancement side features.

[0010] Preferably, the output feature The formula for generating it is: ; ; ; ; in, For small-scale texture sub-features, It represents a large-scale texture sub-feature. As an aggregation feature, This represents the single-layer backbone features of the current level after batch regularization preprocessing. To add pixels one by one, To multiply pixel by pixel, To simplify convolution operations, This is the activation function.

[0011] Preferably, the multi-layered enhanced side feature Represented as: ; Where n is the total number of layers. For hyperparameters, This represents the layer-level enhancement feature of the i-th layer.

[0012] Preferably, the sampling and multi-scale fusion correction processing of the single-layer backbone features at each level to obtain the corrected overall backbone features specifically includes: The process is iterated layer by layer from deep to shallow, and the corrected backbone features obtained after the previous iteration are sent into the upsampling unit to complete the scale magnification. The scaled-up corrected backbone features are added and fused pixel by pixel with the single-layer backbone features of the current level, and then passed through two levels of residual units to complete texture detail enhancement. The enhanced features are subtracted pixel by pixel from the corrected backbone features obtained after the previous iteration to complete difference correction. After correction, the features are sent to the selective kernel fusion unit to dynamically fuse multi-scale channel information to obtain the corrected backbone features after the current level iteration. Complete all iterative correction operations at each level, integrate the corrected backbone features after all levels of iteration, and summarize to form the corrected overall backbone features.

[0013] Preferably, the method further includes: Construct teacher and student models; Collect real traffic scene foggy image samples, input them simultaneously into the teacher model and student model, and output the corresponding predicted feature map and fog-free reconstructed image respectively; The feature distribution output by the teacher model and the grayscale distribution of the reconstructed image are used as soft labels, and combined with hard labels composed of clear real-world images to construct a composite loss function. The weights of the convolutional and fusion units of the student model are iteratively updated by backpropagation using a composite loss function, continuously reducing the distribution difference between the output results of the student model and the teacher model, and completing the real distillation training. After training, the student model weights are fixed and used to process foggy images output by traffic data collection equipment in real time.

[0014] Secondly, an enhancement processing device for foggy images in intelligent transportation scenarios is provided. The enhancement processing device for foggy images in intelligent transportation scenarios includes: a processor and a memory for storing processor-executable instructions. The processor is configured to execute the enhancement processing method for foggy images in the intelligent transportation scenario.

[0015] Thirdly, a non-volatile computer storage medium is provided, the computer storage medium storing computer-executable instructions, which are executed by one or more processors to perform the enhancement processing method for foggy images in intelligent transportation scenarios described in the first aspect.

[0016] Fourthly, a chip is provided, comprising: a processor and an interface for calling and running a computer program stored in a memory to perform the enhancement processing method for foggy images in a smart transportation scenario as described in the first aspect.

[0017] Fifthly, a computer program product containing instructions is provided, which, when executed on a computer or processor, causes the computer or processor to perform the enhancement processing method for foggy images in a smart transportation scenario as described in the first aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention extracts multi-scale basic features from images, introduces atmospheric scattering priors at the shallow layer to generate physically enhanced features, and fuses them to obtain single-layer backbone features, which aligns with the imaging mechanism of haze. Multi-scale aggregation and adaptive attention weighting are applied to the backbone features at each layer, fusing hierarchical information to obtain multi-layer enhanced side features. The single-layer backbone features are sampled and corrected at multiple scales to obtain overall backbone features. These two types of features are fused to generate comprehensive features for image reconstruction. This invention fully utilizes physical priors and multi-level feature interactions, enriching multi-scale image details, optimizing the feature correction effect in foggy areas, and reconstructing clear, fog-free traffic images, thus improving the accuracy of downstream traffic visual processing tasks. Attached Figure Description

[0019] 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of a defogging network provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for enhancing foggy images in a smart transportation scenario, as provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a physical sensing multi-branch unit provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a process for obtaining single-layer backbone features according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a lightweight multi-scale feature aggregation block provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a process for obtaining multi-layered enhanced side features according to an embodiment of the present invention; Figure 7 This is a schematic diagram of a process for obtaining overall backbone features according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an SOS enhancement module provided in an embodiment of the present invention; Figure 9This is a schematic diagram of a distillation strategy provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of a distillation strategy provided in an embodiment of the present invention; Figure 11 This is a schematic diagram of an experimental environment provided in an embodiment of the present invention; Figure 12 This is a schematic diagram illustrating the quantitative comparison results of various defogging methods provided in an embodiment of the present invention; Figure 13 This is a schematic diagram illustrating the quantitative comparison results of various defogging methods provided in an embodiment of the present invention; Figure 14 This is an example diagram illustrating the qualitative comparison results of various defogging methods provided in this embodiment of the invention; Figure 15 This is a schematic diagram showing a local magnified comparison of results provided in an embodiment of the present invention; Figure 16 This is a schematic diagram illustrating the qualitative comparison results of various dehazing methods on the Dense-Haze dataset provided in an embodiment of the present invention; Figure 17 This is a schematic diagram illustrating the qualitative comparison results of various dehazing methods on the NH-Haze datasets provided in an embodiment of the present invention; Figure 18 This is a schematic diagram of an ablation experiment study with different loss functions provided in an embodiment of the present invention; Figure 19 This is a schematic diagram of an ablation experiment study using different modules provided in an embodiment of the present invention; Figure 20 This is a schematic diagram of the structure of an enhancement processing device for foggy images in a smart transportation scenario provided by an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as openly inclusive, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this disclosure. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples; that is, although they may be incorporated into embodiments or examples using the above terms for reasons such as order and position, it does not limit them to be incorporated in combination by a single embodiment or example.

[0023] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, for example, the description may use the prefix "A" or "B" to describe the same type of nouns as two independent entities. In this case, the corresponding features defined with "A" and "B" are used only to distinguish between similar entities and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features.

[0024] In describing some embodiments, the terms "coupled," "coupled," and "connected," and their derivative expressions, may be used. For example, the term "connected" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact with each other. Similarly, the term "coupled" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact. However, the terms "connected" or "coupled" may also refer to two or more components that do not have direct contact with each other but still cooperate or interact with each other, such as "optical coupling," "wireless connection," etc. The embodiments disclosed herein are not necessarily limited to the scope of this invention.

[0025] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] Example 1: In recent years, with the development of deep learning, most image dehazing algorithms are based on deep learning, among which CNN (Convolutional Neural Network) and Transformer are widely used and have achieved good results. CNN can effectively extract local features of an image through convolutional operations of convolutional layers, thus achieving image dehazing to a certain extent. However, the limitation of its local receptive field makes it difficult to capture global dependencies. When faced with large areas of haze, it is difficult to fully utilize global information for dehazing, resulting in insufficient detail and overall clarity in the dehazed image. Transformer introduces a self-attention mechanism, which can effectively capture global dependencies in an image, bringing a new approach to image dehazing. However, the quadratic time complexity of this mechanism leads to a sharp increase in computation when processing high-resolution traffic images, requiring extremely high hardware resources and resulting in low computational efficiency, which cannot meet the real-time requirements of traffic scenarios.

[0027] To address the shortcomings of existing technologies and solve the technical problems of traditional intelligent transportation fog image enhancement methods, which fail to consider the physical laws of fog imaging, have insufficient multi-scale feature mining, poor hierarchical feature interactivity, and low feature correction accuracy, resulting in loss of detail and significant fog residue in defogging images and failing to meet the high-precision requirements of traffic vision processing in foggy weather, this embodiment proposes an enhancement processing method for foggy images in intelligent transportation scenarios. Before detailing the specific steps of this enhancement processing method, this embodiment first describes the overall architecture of the defogging network used to execute the method.

[0028] In one implementation, the defogging network comprises at least five layers, each containing a different number of DehazeFormer blocks. However, due to the large number of modules and parameters in DehazeFormer, as well as the enormous computational load, DehazeFormer is difficult to apply to resource-constrained and real-time scenarios, such as vehicle applications.

[0029] To obtain a lightweight dehazing network, this embodiment first reduces the number of DehazeFormer blocks in each layer of the DehazeFormer-T network, which significantly reduces redundancy in model parameters. After reducing the number of blocks, the number of DehazeFormer blocks used per layer is 1, 2, 1, 1, and 1, respectively. The reason for retaining two blocks in the second layer is that it performs better in learning feature information from the teacher network than using a single DehazeFormer block. Furthermore, the feature information extracted using more parameters in the second layer has a significant impact on feature acquisition in subsequent layers. On the other hand, due to reducing the number of DehazeFormer blocks in each layer to alleviate the reduced network parameters, local content details of the image may not be fully explored. To enhance the performance of the proposed lightweight model, this embodiment designs a PMU and integrates it into the DehazeFormer blocks of the first three layers, achieving a combination of CNN and Vision Transformer, extracting local-global dependencies, and generating richer content details. Furthermore, to promote the interaction of visual features across different layers and improve the utilization of multi-scale features, this embodiment also proposes MFEM to improve the feature utilization rate of each layer and enhance the overall performance of the network.

[0030] like Figure 1 As shown, the dehazing network proposed in this embodiment mainly consists of a main network and side networks. The main network is based on a lightweight baseline model and mainly includes PMU and DehazeFormer modules. It uses Selective Kernel Fusion (SKF) and an improved SOS enhancement strategy for feature fusion and connection. The side networks are constructed using MFEM and mainly include LMFAB. The working principles of each module in the dehazing network will be introduced below.

[0031] Based on the specific structure of the defogging network described above, in one implementation, such as Figure 2 As shown, the method for enhancing foggy images in the intelligent transportation scenario includes: Step 101: Acquire foggy images collected by traffic data acquisition equipment, extract multi-scale basic image features from the foggy images in layers; introduce atmospheric scattering physical priors to obtain physical perception enhancement features in the shallow feature extraction stage, and fuse the basic image features of the same level with the physical perception enhancement features to obtain single-layer backbone features of each level.

[0032] First, foggy images captured by traffic acquisition equipment in a smart transportation scenario are collected. Then, multi-layer feature extraction is performed on the acquired foggy images to mine the basic texture and contour information of different scales and depths of the images, resulting in multi-scale basic image features. The basic image features are the original visual basic information of the road, vehicles, pedestrians and other targets in the image.

[0033] To overcome the problem of incomplete dehazing caused by traditional networks relying solely on data learning and lacking physical constraints, this embodiment integrates PMU (Physical Awareness Multi-Branch Unit) in the first three DehazeFormer blocks. At the shallow feature extraction stage, atmospheric scattering imaging physical priors are introduced, and physical awareness enhancement features are calculated based on the real imaging mechanism of haze, thus making up for the deficiency of basic image features lacking physical logic support.

[0034] Finally, the basic image features and physical perception enhancement features at the same level are fused together. The feature that combines image texture information and haze physical imaging information is the single-layer backbone feature. Each feature extraction level generates a corresponding single-layer backbone feature, thus constructing a complete multi-layer backbone feature system.

[0035] Step 102: Perform multi-scale feature aggregation and adaptive attention weighting on the single-layer backbone features of each layer, and fuse the weighting results of all layers to obtain multi-layer enhanced side features.

[0036] This embodiment performs refined feature enhancement processing on the single-layer backbone features corresponding to each level, and completes multi-scale feature aggregation and adaptive attention weighting operations one by one. Among them, multi-scale feature aggregation is used to extract and fuse image texture details of different granularities, enriching the multi-scale information of single-layer features; adaptive attention weighting can automatically filter effective target features, suppress fog noise features, and enhance the feature representation capabilities of key traffic targets.

[0037] After the above processing, the single-layer backbone features of each layer generate corresponding hierarchical enhancement features, effectively solving the problems of insufficient interaction of hierarchical features and insufficient detail extraction in traditional techniques. Finally, the hierarchical enhancement features of all layers are fused and summarized according to preset weights to obtain multi-layer enhanced side features that integrate full-level detail information. The multi-layer enhanced side features mainly retain the local fine texture and multi-scale detail information of the image.

[0038] Step 103: Sample and perform multi-scale fusion correction on the single-layer backbone features of each level to obtain the corrected overall backbone features; fuse the overall backbone features with the multi-layer enhanced side features to obtain the comprehensive features.

[0039] Using an iterative processing logic from deep to shallow layers, the single-layer backbone features of each level are sequentially upsampled and scaled and then fused for correction. Through layer-by-layer feature correction and bias correction, the feature offset and distortion caused by fog are eliminated. After all layers have been iterated, all corrected layer features are integrated to obtain the corrected overall backbone features with global fog correction capabilities. These features focus on the defogging correction effect of the global image.

[0040] To balance global dehazing effects with local detail information, this embodiment deeply integrates the overall backbone features that emphasize global correction with the multi-layered enhanced side features that emphasize local detail enhancement. By integrating global correction information with multi-layered refined detail information, a comprehensive feature with complete information is finally obtained.

[0041] Step 104: Based on the comprehensive features, reconstruct and output a fog-free traffic image, and input the fog-free traffic image into a downstream traffic vision processing device to perform traffic vision processing tasks.

[0042] Among them, based on the high-dimensional complete comprehensive features obtained by fusion, the foggy image is reconstructed and restored to remove fog interference information in the image, restore clear and realistic scene visual information, and finally output a fog-free traffic image with clear texture and no fog residue.

[0043] The reconstructed fog-free traffic images are input into various downstream traffic vision processing devices, providing high-quality image data support for core intelligent traffic vision tasks such as traffic target recognition, pedestrian behavior analysis, and traffic flow monitoring, ensuring the accuracy and stability of intelligent transportation system operation in foggy scenarios.

[0044] The following section will provide a detailed explanation of each step in the method for enhancing foggy images in the intelligent transportation scenario.

[0045] Many dehazing algorithms are based on atmospheric scattering models, which can be expressed by the following formula: (1) in, These are observed images with fog. It is a single pixel in the haze-free image to be reconstructed. The value at that location, It is the transmittance in the atmosphere. It is direct transmission, and This represents atmospheric light. However, using formula (1) directly applies to atmospheric light... and transmission diagram This can lead to accumulated errors. Imposing physical priors in the feature space can promote interpretability consistent with the fog process, without relying on... and The true value.

[0046] This embodiment proposes a Physical Sensing Multi-Branch Unit (PMU), such as Figure 3 As shown, branches are used to estimate separately. and ,pass and The features considered take into account the synergistic effect between the two. However, this embodiment believes that this does not extract the characteristics. and Based on the collaborative characteristics, this embodiment considers adding branches. By extracting collaborative features and improving each branch, the system can better focus on the joint features of atmospheric light and transmission maps, further extracting physical feature information and obtaining more useful image content information. To make the PMU in this embodiment more lightweight, group convolution and depthwise convolution are used in the PMU to reduce the number of parameters. The PMU exists in the first three layers of the backbone network and is connected together through SKF and DehazeFormer blocks. The fusion of the CNN-based structure and the Transformer-based architecture allows the extracted features to complement each other, further improving the performance of the entire network. Based on the above analysis, in one embodiment, such as... Figure 3 and Figure 4 As shown, the process of introducing atmospheric scattering physics priors to obtain physics-enhanced features in the shallow feature extraction stage, and fusing the basic image features at the same level with the physics-enhanced features to obtain single-layer backbone features at each level, specifically includes: Step 1011: Construct at least three parallel feature extraction branches to generate approximate transmittance features, approximate atmospheric light features, and joint transmittance and atmospheric light features, respectively.

[0047] Specifically, at least three parallel feature extraction branches are constructed, and the three branches operate independently and in parallel to mine physical information from shallow basic features, thereby generating approximate transmittance features, approximate atmospheric light features, and joint features of transmittance and atmospheric light. This not only retains the independent feature information of the two core physical quantities, but also supplements the joint feature information of their coupling relationship, thus improving the physical feature representation system.

[0048] Step 1012: Based on the atmospheric scattering model, perform pixel-by-pixel multiplication, pixel-by-pixel subtraction, and pixel-by-pixel addition fusion operations on the transmittance approximation feature, the atmospheric light approximation feature, and the joint transmittance and atmospheric light feature to obtain the physical perception enhancement feature.

[0049] Based on the physical operation logic of the atmospheric scattering imaging model, the extracted approximate transmittance features, approximate atmospheric light features, and joint features of transmittance and atmospheric light are subjected to multi-dimensional pixel-level fusion operations of pixel-by-pixel multiplication, pixel-by-pixel subtraction, and pixel-by-pixel addition. This fully integrates independent physical features and joint physical features, mines deep physical prior information, and finally calculates physical perception enhancement features that are both independent and coupled.

[0050] Step 1013: Input the obtained physical perception enhancement features and the basic image features of the same level into the selective kernel fusion unit to complete the fusion of the two features to generate the single-layer backbone features of the corresponding level.

[0051] The physically enhanced features obtained from the computation are input together with the basic image features extracted at the current level into the selective kernel fusion unit. Through adaptive channel filtering and weighted fusion, the deep complementary fusion of shallow texture basic features and deep physically enhanced features is completed, generating a single-layer backbone feature corresponding to the coding level.

[0052] To obtain a clear, fog-free image, the above formula (1) can be restated as: (2) Using convolution kernels Simultaneously convolving the signals on both sides of formula (2) yields formula (3).

[0053] (3) in, and Let represent the convolution operation and the Hadamard product, respectively. Then, equation (3) is introduced into matrix-vector form, i.e. Using matrices and vectors respectively Equation (3) can be expressed as: (4) Next, the matrix can be... Decomposed into two matrices and The product of, since , and It is unknown; achieving this decomposition can be represented as solving an undetermined equation, and it can be guaranteed that... The existence of allows us to obtain the following formula: (5) Will Represented as the characteristics corresponding to atmospheric light The approximation will Represented as Approximate. Furthermore. and This can be viewed as extracting features from a foggy image and its corresponding clear image, respectively. and Based on formula (5), the physical perception features can be calculated. : (6) Here, 1 represents a matrix whose elements are all 1s. This embodiment further considers... and Due to the synergistic effect between them, this embodiment further improves formula (6) to obtain formula (7): (7) in, These are hyperparameters used to balance the terms of the equation. Therefore, the PMU in this embodiment mainly has three branches, which are used for estimation. , and That is: approximate characteristics of transmittance Represented as: (8) Atmospheric light approximation characteristics Represented as: (9) Combined characteristics of transmittance and atmospheric light Represented as: (10) Physical perception enhancement features Represented as: ; in, As basic image features at the same level, Indicates lightweight convolution. Indicates nonlinear activation. This represents compressing the eigenvalue range to [0,1] using global average pooling. This indicates sampling of the global mean of the entire feature map. It represents the Hadamardi (or Hadama) stack.

[0054] Since atmospheric light is usually assumed to be uniform, therefore, in estimating... Use global average pooling This is to eliminate unnecessary information in the feature space. This embodiment... The estimates further strengthened and The synergistic effect between them results in more comprehensive information extracted from the enhanced physical perception features.

[0055] During downsampling, details gradually disappear. To make fuller use of the feature information extracted by the U-Net-like backbone network, this embodiment proposes a new module, MFEM, which retains more spatial details and generates richer multi-scale features. For example... Figure 1 As shown, the proposed multi-layer feature enhancement module MFEM mainly includes a lightweight multi-scale feature aggregation block LMFAB and a sigmoid layer. Its main function is to process the features of the current layer and the features of the two adjacent layers to gradually and adaptively improve the dehazing performance.

[0056] In one implementation, such as Figure 5 and Figure 6 As shown, the process involves performing multi-scale feature aggregation and adaptive attention weighting on the single-layer backbone features of each layer, and fusing the weighting results of all layers to obtain multi-layer enhanced side features, specifically including: Step 1021: After batch regularization preprocessing of the single-layer backbone features of the current level, different convolution operations are used to extract texture sub-features of different granularities. The texture sub-features are added and fused pixel by pixel and then processed by convolution to generate aggregated features that aggregate multi-scale texture information, thus completing the multi-scale feature aggregation.

[0057] In this approach, the single-layer backbone features output from the current layer are used as the input to the LMFAB module. They are first fed into the Batch Normalization (BN) unit for normalization preprocessing, unifying feature distribution and improving network training stability. In one implementation, the BN-processed features are divided into three parallel convolutional branches: the first branch uses 1×1 convolutions to extract minute, fine-grained texture sub-features; the second branch uses two sets of cascaded 3×3 convolutions to extract medium-scale texture sub-features; and the third branch uses 5×5 convolutions to extract large-scale global texture sub-features.

[0058] The 1×1 branch output, the two sets of cascaded 3×3 branch outputs, and the 5×5 branch output are each fused within the branch by adding pixels one by one in the image. Then, 3×3 convolution smoothing is applied to the three fused results. The three convolution outputs are then summed up by adding pixels one by one to complete the full-scale information aggregation, generating aggregated features that fuse all granular texture information of 1×1, 3×3, and 5×5. This completes the multi-scale feature aggregation operation.

[0059] Step 1022: Extract branch features from the original input single-layer backbone features, generate adaptive attention weights through convolution and activation operations, multiply the adaptive attention weights with the aggregated features pixel by pixel to complete feature filtering and weighting, and obtain weighted texture features.

[0060] A separate branch is drawn from the leftmost original Input (single-layer backbone features) of LMFAB, without going through the three multi-scale convolution branches: this branch first performs a 1×1 convolution to compress the channel dimension, and then sends it into the Sigmoid activation unit to constrain the feature values ​​to the range of 0~1, generating pixel-level adaptive attention weights.

[0061] The aggregated features obtained in step 1021 are retrieved and weighted at the pixel level by multiplying the adaptive attention weights output by this branch with the pixel-wise multiplication sign. This automatically amplifies the effective target features such as roads, vehicles, and pedestrians, while suppressing meaningless fog noise features. This completes the feature filtering and weighting, and the weighted texture features are obtained after the operation.

[0062] Step 1023: The weighted texture features are convolved to adjust the channels, and then added to the pixel-by-pixel residual of the original input single-layer backbone features to obtain the output features. The neighboring layer backbone features of the adjacent layer output are obtained, and the neighboring layer backbone features are upsampled or downsampled to complete the scale alignment. The scale-aligned neighboring layer backbone features are added to the output features pixel by pixel and fused to generate the layer enhancement features corresponding to the current layer.

[0063] The weighted texture features are fed into a 1×1 convolutional unit to adjust the channel dimensions. The adjusted features are then connected to the original single-layer backbone features through a pixel-by-pixel summation unit at the end to obtain the output features of the LMFAB module itself.

[0064] Image features at different scales possess different semantic information. To further utilize the rich semantic information extracted by the backbone network PMU and DehazeFormer blocks, this embodiment proposes a lightweight multi-scale feature aggregation block, the specific structure of which is as follows: Figure 5 As shown. Specifically, assuming the input feature map is used, this embodiment first uses BatchNorm for normalization, that is... Then, a convolution kernel is used to extract highly correlated blurred features to further enrich the texture features. The output features The formula for generating it is: ; ; ; ; in, For small-scale texture sub-features, It represents a large-scale texture sub-feature. As an aggregation feature, This represents the single-layer backbone features of the current level after batch regularization preprocessing. To add pixels one by one, To multiply pixel by pixel, To simplify convolution operations, This is the activation function.

[0065] It should be noted that, in order to make the overall structure of this embodiment more lightweight, the LMFAB... and The convolutional layers use depthwise convolution, which significantly reduces the number of parameters and computational complexity, while maintaining the LMFAB feature extraction capability. Combined with the multi-layer feature enhancement module proposed in this embodiment, the overall network performance is improved.

[0066] Finally, the neighboring layer features output by the MFEM module of the adjacent layer of the network are retrieved. Based on the size difference between the current layer and the neighboring layer, the features of the previous layer are downsampled and the features of the next layer are upsampled to complete the scale alignment. The neighboring layer features after scale unification are added and fused with the LMFAB output features of the current layer pixel by pixel, introducing the correlation information between the upper and lower layers, eliminating the limitations of single-layer feature information, and obtaining the layer enhancement features exclusive to the current layer after fusion.

[0067] Step 1024: After obtaining the corresponding layer enhancement features for each layer in sequence, add all the layer enhancement features pixel by pixel according to the preset weight to obtain the multi-layer enhancement side features.

[0068] In this process, steps 1021 to 1023 are repeated to process the single-layer backbone features output by the five DehazeFormer blocks from Layer 1 to Layer 5 independently, thereby obtaining the one-to-one corresponding layer-enhanced features of the five layers.

[0069] As can be seen from the overall network diagram, the five-layer MFEM modules output their respective layer-level enhancement features. The five-layer features are sequentially added pixel by pixel through the continuous nodes in the diagram, and combined with the preset weight coefficients of each layer to complete the weighted accumulation and summarization. This integrates all the shallow fine-grained texture and deep global contour information, and finally summarizes and generates multi-layer enhanced side features that fuse all-layer detail information. These features are then sent to the backend to be fused with the corrected overall backbone features.

[0070] Specifically, given a certain layer Output feature map and adjacent layers Output feature map This embodiment first feeds the input into the LMFAB and Sigmoid layers. LMFAB is a multi-scale feature aggregation module that can make the feature information of each layer more saturated. The outputs are denoted as follows: , Meanwhile, for adjacent layers The characteristics of the layer, in this embodiment, are obtained through upsampling or downsampling and the first layer. After the feature information of each layer has the same scale, it is also input into the Sigmoid layer, and the output is... Finally the The hierarchical enhancement features output by the MFEM of a layer can be expressed by the following formula: (11) It should be noted that since the network in this embodiment mainly consists of five layers, in the fifth layer, It consists only of the first item. The main function of the MFEM in this embodiment is to further extract multi-scale features from each layer of the backbone network, promoting the interaction of visual features at different levels, improving the utilization rate between multi-scale features, and thus enhancing the performance of the backbone network. Therefore, the final output of the side network is a linear combination of the MFEM output features of each layer, representing the features of local details and global semantics in the image. The multi-layer enhanced side features... Represented as: Where n is the total number of floors. For hyperparameters, This represents the layer-level enhancement feature of the i-th layer.

[0071] The presence of the Sigmoid layer enables MFEM to have an attention mechanism, reducing feature redundancy and adaptively learning weights at the feature map level. This results in compact and effective feature information, making features relevant to the dehazing task more prominent, thereby further helping the backbone network to recover images with clearer texture details and more vivid colors.

[0072] In one implementation, such as Figure 7 As shown, the sampling and multi-scale fusion correction of the single-layer backbone features at each level to obtain the corrected overall backbone features specifically includes: Step 1031: Iterate layer by layer from deep to shallow, and send the corrected backbone features obtained after the previous iteration into the upsampling unit to complete the scale magnification.

[0073] The decoding phase of this network follows an iterative processing logic from the deep Layer 5 to the shallow Layer 3, as shown in the diagram J. n+1 This represents the corrected backbone feature output after the iteration of the previous layer. This feature is input into the UP upsampling unit, and the resolution and scale of the feature map are amplified through upsampling operation so that its size matches the feature size of the current layer, thus completing cross-layer scale alignment and providing feature data with uniform size for subsequent feature fusion.

[0074] Step 1032: The scaled-up corrected backbone features are added and fused with the single-layer backbone features of the current level pixel by pixel, and then passed through two levels of residual units to complete the texture detail enhancement.

[0075] The deep-corrected backbone features, after being upsampled and amplified by UP, are transmitted upwards to the pixel-by-pixel summation unit in the image. The single-layer backbone features of the current level correspond to the attached input I. n The two features are superimposed and fused pixel by pixel in the addition unit, combining the deep global dehazing correction information with the original texture information of the current layer; After fusion, the features flow upwards through two RSB residual blocks. The two-level residual units extract and enhance the texture details of key targets such as roads, vehicles, and pedestrians in the image layer by layer, making up for the feature detail attenuation problem caused by fog, and outputting the enhanced fused features.

[0076] Step 1033: Subtract the enhanced features from the corrected backbone features obtained after the previous iteration pixel by pixel to complete the difference correction. After the correction is completed, send it to the selective kernel fusion unit to dynamically fuse multi-scale channel information to obtain the corrected backbone features after the current level iteration.

[0077] The enhanced texture features output from the two-stage residual blocks are fed upwards into the pixel-by-pixel subtraction unit in the image, and the original corrected backbone features J from the previous layer, transmitted back from the other branch, are taken. n+1 As a reference, the two are interpolated pixel by pixel to offset the feature shift and numerical distortion caused by fog, thus completing pixel-level difference correction. The corrected features are input upwards into the SKF selective kernel fusion unit. This unit adaptively selects effective feature channels, suppresses fog noise channels, dynamically fuses multi-scale and multi-channel feature information, and outputs iteratively corrected backbone features adapted to the current level.

[0078] Step 1034: Complete all iterative correction operations at each level, integrate the corrected backbone features after all levels of iteration, and summarize to form the corrected overall backbone features.

[0079] Repeat steps 1031 to 1033 to sequentially complete the iterative correction of the three SOS modules of Layer 5, Layer 4, and Layer 3, and obtain the corresponding iterative correction backbone features of each of the three layers. Integrate and summarize all the correction backbone features of the three layers in hierarchical order, and fuse the deep global fog correction information and the shallow fine texture correction information to finally form the overall corrected backbone features with global defogging correction capability and complete hierarchical information.

[0080] Combining the above process, DehazeFormer uses the Selective Kernel Fusion (SKF) module based on SKNet (Selective Kernel Networks) to achieve multi-scale feature map fusion. The SKF module can dynamically fuse feature maps from different paths through channels, offering better performance compared to other feature fusion methods. On the other hand, the SOS enhancement module is used to enhance the multi-scale characteristics of the decoder. This module employs an effective network design enhancement strategy, which has been proven to improve the performance of image dehazing networks. To further enhance these characteristics, this embodiment uses a novel approach to combine the SOS enhancement module and the SKF module, such as... Figure 8As shown, the SOS enhancement module consists of three residual blocks. However, this embodiment found that in the proposed lightweight network, using only two residual blocks results in better model performance and further reductions in model parameters. Therefore, this embodiment uses only two residual blocks in the SOS enhancement module.

[0081] Since reducing the number of parameters may degrade the performance of the proposed lightweight model, this embodiment employs a knowledge distillation strategy for training, aiming to suppress performance degradation as much as possible. In one implementation, such as Figure 9 and Figure 10 As shown, the method further includes: Step 201: Construct teacher and student models.

[0082] Among them, reference Figure 9 There are two parallel network links, one above the other, representing the teacher model and the other below, representing the lightweight student model in this embodiment.

[0083] The teacher model adopts a standard complete dehazing network architecture. In the encoding and decoding stages, only SKF selective kernel fusion units are configured, without a lightweight SOS correction structure. The network has a larger number of parameters and stronger feature fitting ability, and can learn complete fog feature mapping rules. The student model is the lightweight hazy image enhancement network described in the aforementioned embodiments. The first three layers of the encoding layer integrate lightweight PMU units, and the decoding module is replaced with an SKF+SOS composite correction module, which significantly reduces the overall number of network parameters. This is the target network used for online inference in this scheme. The input ports of the two models are independent of each other, and the network layers correspond one-to-one. The Layer1~Layer5 encoding layers, the terminal 3×3 convolution, and the SR reconstruction module are completely aligned, forming a layer-matched teacher-student distillation pairing structure.

[0084] Step 202: Collect real traffic scene foggy image samples, input them into the teacher model and student model simultaneously, and output the corresponding predicted feature map and fog-free reconstructed image respectively.

[0085] Hazy collects foggy image samples from real roads, intersections, and other intelligent transportation scenarios, and simultaneously sends the same foggy image to the input ports of both the teacher model and the student model.

[0086] At the feature level, after the image undergoes 3×3 convolution and Layer 1~Layer 3 encoding to extract shallow features, the teacher model Layer 3 outputs intermediate features, and the student model Layer 3 outputs lightweight intermediate features. The two sets of features correspond one-to-one and serve as the basis for supervision of intermediate layer distillation, corresponding to the dashed line L in the figure. KDl3Loss-constrained branches; at the image reconstruction level, after the image flows completely through all encoding and decoding modules, it is sent to its respective terminal SR reconstruction module. The teacher network outputs a high-precision, fog-free reconstructed image, and the student network outputs a fog-free reconstructed image generated by the lightweight model. The two sets of reconstructed images form an output layer distillation supervision pair, corresponding to the dashed line L in the figure. KDout Loss-constrained branches.

[0087] Step 203: Use the feature distribution output by the teacher model and the grayscale distribution of the reconstructed image as soft labels, and combine them with the hard labels composed of clear real-world images to construct a composite loss function.

[0088] The first type is hard-label loss, which collects clear, fog-free real-world images that match the input foggy image one-to-one as ground truth hard labels, constraining the student model to reconstruct images that closely match the real, clear scene; the second type is intermediate-layer feature soft-label loss L. KDl3 Using the teacher network's Layer 3 output features as soft labels, the distribution of student network's output features at the same level is constrained to maintain consistency with the teacher's feature distribution. Figure 9 The dashed loss link between Layer 3; the third type is the soft label loss L of the output image. KDout The haze-free reconstructed image output by the teacher network's SR module is used as a soft label to constrain the grayscale and texture distribution of the student network's final reconstructed image to align with the teacher's output. Figure 9 The dashed loss link between the mid-to-end SR modules.

[0089] The three types of losses are weighted and combined to form a composite loss function that simultaneously constrains intermediate features, terminal images, and real-world images.

[0090] Step 204: Use the composite loss function to backpropagate and iteratively update the weights of the convolution and fusion units of the student model, continuously reduce the distribution difference between the output results of the student model and the teacher model, and complete the real distillation training.

[0091] Using the overall error of the composite loss function as the optimization objective, backpropagation updates are performed only on the student model, while the teacher model's weights are frozen and not used in training. During network backpropagation, L is synchronously propagated back. KDl3 L KDout The gradient of the hard-label reconstruction loss is updated synchronously with all weight parameters of the PMU physical perception unit, LMFAB lightweight aggregation block, SKF+SOS correction module, convolutional layers, and feature fusion units in the student network. The distribution deviation between the intermediate layer features of the student model and the final reconstructed image and the output of the teacher model is continuously reduced iteratively. The prior knowledge of the large-parameter teacher network is used to make up for the insufficient feature extraction ability of the lightweight student network due to parameter reduction, and to suppress the degradation of the dehazing effect.

[0092] Step 205: After training, solidify the student model weights for real-time processing of foggy images output by traffic data collection equipment.

[0093] The knowledge distillation process is terminated after continuous iterative training until the composite loss converges and the number of iterations reaches a preset threshold. The weights of the trained student model are then permanently saved, while the teacher model, with its large number of parameters, is discarded. During online use, only a lightweight student network is deployed to receive real-time foggy images from traffic data collection devices such as road cameras. Based on the aforementioned feature extraction, side-channel enhancement, decoding correction, and image reconstruction processes, clear, fog-free images are generated for downstream visual tasks such as traffic target recognition and flow monitoring.

[0094] In one implementation, this embodiment uses a version of DehazeFormer, DehazeFormer-S, as the teacher network. The DehazeFormer-S network is a minor variant of DehazeFormer without excessive redundancy, so the student network can efficiently receive useful feature information with less training time.

[0095] Some researchers employ knowledge distillation strategies to train image deblurring networks, but they use clear photos as input to the teacher network. Because lightweight networks lack the capacity to extract feature information compared to the teacher network, the features extracted by the teacher network can differ significantly when given a hazy image as input. Consequently, the student network cannot accurately extract useful knowledge from the teacher network, resulting in poor quality hazy-free images at the final output. Therefore, both the teacher and student networks select the same hazy image as input to reconstruct clean, hazy-free images.

[0096] In many image restoration tasks, training models using the L1 loss function can achieve better performance than L2 in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). Therefore, this embodiment also uses L1 to construct the distillation loss function. In the lightweight network and DehazeFormer-S network of this embodiment, the features of the third layer play a role in continuing and enriching the texture features extracted by the first two layers. Since the latter two layers are multilayer perceptron structures, the data stream generated by the third layer plays an important role in the final reconstruction of the haze-free image. Therefore, this embodiment constructs a distillation loss function in the third layer of the teacher-student network. In the output section, this embodiment also constructs a distillation loss. This is used to monitor and guide the final output. Therefore, the final total knowledge distillation loss is... Depend on and Composition can be represented as: ; ; ; in, and These represent the feature maps of the third layer and the final output of the teacher network, respectively. and These represent the feature maps of the third layer and the final output of the student network, respectively. This represents the L1 norm.

[0097] This embodiment uses L1 loss. Multi-scale structural similarity (MS-SSIM) loss Perceived loss and distillation loss A linear combination of these factors is used to balance visual perception and quantitative assessment, resulting in an overall loss function. Described as: (20) (twenty one) (twenty two) (twenty three) in, These are hyperparameters used to balance each loss function, taking values ​​of 1, 1, and 0.001 respectively; assuming and These refer to the dehazed image recovered via the network and the corresponding original clear image, respectively. This is a function representing the features extracted from layers 3, 8, and 15 of the VGG model. The SSIM loss integrates variations in resolution and visualization conditions to account for structural differences. Therefore, the L1 loss... MS-SSIM loss Perceived loss This can be expressed by the formula above. This represents the L2 norm.

[0098] In summary, this embodiment extracts multi-scale basic features from the image, introduces atmospheric scattering priors at the shallow layer to generate physically enhanced features, and fuses them to obtain single-layer backbone features, which aligns with the imaging mechanism of haze. Multi-scale aggregation and adaptive attention weighting are applied to the backbone features at each layer, fusing hierarchical information to obtain multi-layer enhanced side features. Single-layer backbone features are sampled and corrected at multiple scales to obtain overall backbone features. These two types of features are fused to generate comprehensive features for image reconstruction. This invention fully utilizes physical priors and multi-level feature interactions, enriches multi-scale image details, optimizes the feature correction effect in foggy areas, reconstructs clear, fog-free traffic images, and improves the accuracy of downstream traffic visual processing tasks.

[0099] Example 2: This embodiment further verifies the advantages and beneficial effects of the image enhancement method for foggy scenes in intelligent transportation scenarios proposed in Embodiment 1 through specific experiments.

[0100] The environment setup includes: This embodiment uses a Windows 11 64-bit operating system and Python (version 3.8.12) as the software language, based on the deep learning framework PyTorch (version 1.10.2). The hardware environment consists of an Intel Core i5-12450H CPU, an NVIDIA GeForce RTX 3080 Ti (12GB) graphics card, 32GB of RAM, and a 2TB hard drive. Specific experimental conditions are as follows: Figure 11 As shown: Dataset selection includes: This embodiment evaluates the proposed network on both synthetic and real-world datasets. RESIDE (Realistic Single Image Dehazing) is a widely used benchmark dataset containing five subsets. This embodiment selects the Outdoor Training Set (OTS) from these five subsets as the training and validation dataset. From the OTS, this embodiment selects images related to traffic scenes as the dataset, named OTS-Traffic, which contains 54,930 foggy traffic scene images for training. 500 images are used for validation and 500 for testing. For real-world scenes, this embodiment uses the datasets Dense-Haze (Dense Homogeneous Haze Dataset) and NH-Haze (Non-Homogeneous Haze Dataset). Dense-Haze and NH-Haze are the single-image dehazing task datasets from the NTIRE 2019 and NTIRE 2020 challenges, respectively, containing dense and non-uniform fog scenes. These smog scenes were created using real smog records generated by a professional smog camera. The two datasets include 45 training images, 5 validation images, and 5 test images. Both Dense-Haze and NH-Haze have a size of 1600×1200.

[0101] Evaluation indicators and comparisons include: This embodiment uses Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) to evaluate the dehazing performance of the proposed network. This embodiment compares the performance of the proposed lightweight model with other related deep learning-based dehazing algorithms, including DCP (Dark Channel Prior), DehazeNet, GridDehazeNet, AOD-Net (All-in-One Dehazing Network), MSBDN (Multi-Scale Boosted Dehazing Network), GCANet (Gated Context Aggregation Network), FFA-Net (Feature Fusion Attention Network), USCFormer (Unified Semantic Contrastive Transformer), and DehazeFormer. DehazeFormer-B is the baseline model of DehazeFormer, and DehazeFormer-M is an intermediate variant.

[0102] Experimental parameter settings include: For the OTS-Traffic dataset, this embodiment randomly crops the images used for model training to a size of 256×256. The batch size is set to 8. Following the linear scaling rule, this embodiment sets the initial learning rate to 4×10⁻⁶. 4. This embodiment uses the AdamW optimizer and cosine annealing strategy to train the model, where the learning rate gradually decreases from the initial learning rate to 1×10⁻⁶. 6. When the training period reaches 40, the overall loss function tends to stabilize. For the real-world datasets Dense-Haze and NH-Haze, this embodiment uses a model trained on the OTS-Traffic dataset as initialization, sets the batch size to 2, randomly crops the training images to 900×900, and trains the model for 5000 epochs.

[0103] Quantitative comparisons include: Figure 12 This table compares the proposed lightweight model with other state-of-the-art image dehazing methods in terms of performance, number of parameters, and computation on the synthetic dataset OTS-Traffic. The hyphens (-) in the table indicate that no relevant evaluation metrics were tested on the current dataset. In this embodiment, DehazeFormer-S was used as the teacher network in the experiments. Figure 12As can be seen, the proposed network achieves excellent dehazing performance using knowledge distillation. The average PSNR and SSIM values ​​reach 34.68 dB and 0.987, respectively, very close to the teacher network. Because the module proposed in this embodiment uses depthwise convolutions to extract multi-scale information through different convolutions, the network parameters are reduced by 66% compared to DehazeFormer-S, but the performance only decreases by 0.17%. These results demonstrate that this lightweight network has better overall performance than the teacher network. Furthermore, compared to other algorithms such as DehazeFormer-T, MSBDN, FFA-Net, and GCANet, the model in this embodiment uses fewer parameters and has lower computational complexity, yet surpasses their dehazing performance, demonstrating superior overall performance. Although USCFormer has the highest evaluation metric among the compared methods, its parameter count is as high as 61.92M, which is unsuitable for use on resource-constrained devices.

[0104] Figure 13 The comparison shows the proposed lightweight network's performance on the Dense-Haze and NH-Haze datasets, compared to other state-of-the-art models in terms of PSNR and SSIM. Figure 12 It can be observed that all dehazing models perform poorly on both real and synthetic datasets. This is partly because the degradation process of real haze is more complex, and the haze in these two datasets is denser. More importantly, due to the limited number of image pairs in the datasets, deep learning-based methods cannot generalize knowledge about the degradation process through sufficient training, which limits the model's performance. Nevertheless, thanks to the knowledge distillation strategy, the lightweight model learns many useful features from the high-performing teacher network DehazeFormer-S, achieving superior performance in PSNR and SSIM, approaching the performance of DehazeFormer-S.

[0105] Qualitative comparisons include: This section provides a qualitative analysis of the model's performance. Figure 14 and Figure 15 This section showcases the dehazing results of various algorithms on the synthetic dataset OTS-Traffic. To make the differences between images produced by different models clearer, this section magnifies local areas of the images and... Figure 15 It is displayed in the middle. For example, Figure 14As shown, the images restored by AOD-Net and GCANet suffer from severe color and brightness distortion, with significant loss of image detail. In heavy fog, the image generated by MSBDN exhibits noticeable distortion and appears very blurry. In the faded image from FFA-Net, the color and features of most areas can be recovered, but the edges of some objects show color and texture distortion. Since this embodiment employs the proposed MFEM at each layer on one side of the network, more multi-scale feature information can be learned through MFEM. Therefore, the edge features and texture details of the image generated by the network proposed in this embodiment are clearer, such as... Figure 15 As shown, the dehazed images generated by DehazeFormer-T are visually similar to those generated by the proposed network, but the average PSNR and average SSIM of the dehazed images generated by the proposed lightweight dehazing network are slightly higher than those generated by DehazeFormer-T. Furthermore, the proposed lightweight dehazing network has fewer parameters and fewer computations than DehazeFormer-T. These results demonstrate that the proposed lightweight dehazing network outperforms the comparison models.

[0106] Figure 16 and Figure 17 The dehazing results of various dehazing networks on the real-world datasets Dense-Haze and NH-Haze are presented. It can be observed that the haze in both datasets is denser than in the actual environment and exhibits significant non-uniformity. Overall, all methods perform poorly on the Dense-Haze and NH-Haze datasets, with dehazing results typically containing residual haze and artifacts. Images processed by DehazeNet and AOD-Net still contain some dense haze, MSBDN results contain color patch artifacts, and FFA-Net suffers from severe color distortion. DehazeFormer-S and the dehazing network proposed in this embodiment do not produce severe color distortion like FFA-Net, but some areas in the dehazed images still fail to recover their original colors. In summary, the dehazing model proposed in this embodiment and the teacher network DehazeFormer-S show fewer artifacts and less residual haze, resulting in a better overall appearance.

[0107] To verify the advantages of the modules and knowledge distillation strategy proposed in this embodiment, this section conducts an ablation study, performing the following experiments to train the proposed lightweight image dehazing model. Experiment 1: To investigate the impact of different terms on the total loss function, the lightweight model was trained based on each loss function. Experimental results are shown below. Figure 18 Experiment 2: Based on the complete loss function, ablation experiments were conducted on each important module. The results are as follows. Figure 19 As shown.

[0108] Experiment 1 shows that the perceptual loss, SSIM loss, and knowledge distillation loss terms significantly impact the performance of the proposed model. Specifically, without knowledge distillation, the model's average PSNR is 33.87 dB. With knowledge distillation, the average PSNR of the lightweight model in this embodiment reaches 34.68 dB, an improvement of 0.8 dB, representing a 2.4% performance improvement compared to the model without knowledge distillation. These results demonstrate that the constructed loss function is suitable for the lightweight model proposed in this embodiment. These two knowledge distillation terms can transfer knowledge from the teacher model to the student model.

[0109] In Experiment 2, this section conducted ablation experiments on the proposed network modules, including PMU, LMFAB, and the improved SOS enhancement module. From... Figure 19 As can be seen, the designed lightweight dehazing network achieves a PSNR of 33.1 dB without using any of the aforementioned modules. Adding the SOS enhancement module improves the PSNR by 0.26 dB. Applying the PMU module results in an average PSNR of 34.08 dB. This is mainly because the PMU enables the model to simultaneously capture global and local dependencies, thereby generating discriminative features and improving the algorithm's performance. Adding LMFAB results in an average PSNR of 34.68 dB. This is primarily attributed to the adoption of LMFAB, which allows the proposed model to fully utilize the spatial details of the layers, generating rich multi-scale features and thus improving the dehazing performance of the proposed model.

[0110] This embodiment proposes a novel lightweight image dehazing algorithm for intelligent transportation applications. To compensate for the performance degradation caused by the reduction in parameters, this embodiment designs two new modules for the lightweight network, enabling the network to have adaptive learning capabilities and retain fine spatial details using attention features, thereby improving network performance. Specifically, this embodiment integrates the PMU module into the backbone network to further extract the physical features of the image. Simultaneously, an MFEM module is proposed on one side of the backbone network to fully utilize the contextual features and spatial details of each layer. This embodiment also proposes a knowledge distillation training strategy suitable for the lightweight network, which significantly improves its performance. Experimental results show that, compared with other algorithms, the lightweight image dehazing network proposed in this embodiment has fewer model parameters and lower computational cost, yet its performance surpasses most state-of-the-art dehazing algorithms.

[0111] Example 3: In Embodiment 1, a method for enhancing foggy images in a smart transportation scenario was provided. In this embodiment, an apparatus for enhancing foggy images in a smart transportation scenario is proposed. The apparatus for enhancing foggy images in a smart transportation scenario includes: a processor and a memory for storing processor-executable instructions; wherein, the processor is configured to execute the method for enhancing foggy images in a smart transportation scenario described in Embodiment 1.

[0112] like Figure 20 As shown, the image enhancement processing device for foggy images in intelligent transportation scenarios includes a processor 21 and a memory 22, wherein the processor 21 and the memory 22 can be connected by a bus or other means.

[0113] Processor 21 can be a Central Processing Unit (CPU). Processor 21 can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0114] The memory 22, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the enhancement processing method for foggy images in the intelligent transportation scenario in Embodiment 1 of the present invention. The processor executes various functional applications and training processes by running the non-transitory software programs, instructions, and modules stored in the memory.

[0115] The memory 22 may include a program storage area and a training storage area. The program storage area may store the operating system and applications required for at least one function; the training storage area may store training data created by the processor. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory, or other non-transitory solid-state storage device. In some embodiments, the memory 22 may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof. The one or more modules stored in the memory 22, when executed by the processor 21, perform functions such as... Figure 1The method for enhancing foggy images in a smart transportation scenario shown in Example 1 is illustrated. For specific details regarding the above method for enhancing foggy images in a smart transportation scenario, please refer to the relevant documentation. Figure 1 , Figure 2 and Figure 3 The relevant descriptions and effects in the embodiments shown are for reference only and will not be repeated here.

[0116] This embodiment also provides a computer storage medium storing a computer program that can be executed by a processor to complete the method for enhancing foggy images in a smart transportation scenario as described in Embodiment 1.

[0117] The computer storage medium stores computer-executable instructions, which can execute the enhancement processing method for foggy images in intelligent transportation scenarios in any of the above method embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium may also include combinations of the above types of memory.

[0118] The specific steps of the method for enhancing foggy images in the intelligent transportation scenario are described in Example 1, and will not be repeated in this example.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for enhancing foggy images in intelligent transportation scenarios, characterized in that, include: Acquire foggy images collected by traffic data acquisition equipment, and extract multi-scale basic image features from the foggy images in layers; in the shallow feature extraction stage, introduce atmospheric scattering physical priors to obtain physical perception enhancement features, and fuse the basic image features of the same level with the physical perception enhancement features to obtain single-layer backbone features of each level; Multi-scale feature aggregation and adaptive attention weighting are performed on the single-layer backbone features of each layer, and the weighting results of all layers are fused to obtain multi-layer enhanced side features. The single-layer backbone features of each level are sampled and multi-scale fusion correction is performed to obtain the corrected overall backbone features; the overall backbone features are fused with multi-layer enhanced side features to obtain comprehensive features. Based on the comprehensive features, a fog-free traffic image is reconstructed and output. The fog-free traffic image is then input into a downstream traffic vision processing device to perform traffic vision processing tasks.

2. The method for enhancing foggy images in intelligent transportation scenarios according to claim 1, characterized in that, The process of introducing atmospheric scattering physical priors in the shallow feature extraction stage to obtain physical perception enhancement features, and fusing the basic image features at the same level with the physical perception enhancement features to obtain single-layer backbone features at each level, specifically includes: Construct at least three parallel feature extraction branches to generate approximate transmittance features, approximate atmospheric light features, and joint features of transmittance and atmospheric light, respectively. The physical perception enhancement features are obtained by performing pixel-by-pixel multiplication, pixel-by-pixel subtraction, and pixel-by-pixel addition on the transmittance approximation features, atmospheric light approximation features, and transmittance and atmospheric light joint features based on the atmospheric scattering model. The obtained physical perception enhancement features and the basic image features of the same level are input into the selective kernel fusion unit to complete the fusion of the two features to generate the single-layer backbone features of the corresponding level.

3. The method for enhancing foggy images in intelligent transportation scenarios according to claim 2, characterized in that, Approximate characteristics of transmittance Represented as: ； Atmospheric light approximation characteristics Represented as: ； Combined characteristics of transmittance and atmospheric light Represented as: ； Physical perception enhancement features Represented as: ； in, As basic image features at the same level, Indicates lightweight convolution. Indicates nonlinear activation. This represents compressing the eigenvalue range to [0,1] using global average pooling. This indicates sampling of the global mean of the entire feature map. It represents the Hadamardi (or Hadama) stack.

4. The method for enhancing foggy images in intelligent transportation scenarios according to claim 1, characterized in that, The process involves performing multi-scale feature aggregation and adaptive attention weighting on the single-layer backbone features of each layer, and fusing the weighting results of all layers to obtain multi-layer enhanced side features, specifically including: After batch regularization preprocessing of the single-layer backbone features of the current level, different convolution operations are used to extract texture sub-features of different granularities. The texture sub-features are added and fused pixel by pixel and then processed by convolution to generate aggregated features that aggregate multi-scale texture information, thus completing multi-scale feature aggregation. The single-layer backbone features of the original input are extracted into branches, and adaptive attention weights are generated through convolution and activation operations. The adaptive attention weights are multiplied pixel by pixel with the aggregated features to complete the feature filtering and weighting, and the weighted texture features are obtained. The weighted texture features are convolved to adjust the channels, and then added to the pixel-by-pixel residual of the original input single-layer backbone features to obtain the output features. The neighboring layer backbone features of the output of adjacent layers are obtained, and the neighboring layer backbone features are upsampled or downsampled to complete the scale alignment. The scale-aligned neighboring layer backbone features are added to the output features pixel by pixel and fused to generate the layer enhancement features corresponding to the current layer. After obtaining the corresponding layer enhancement features for each layer in sequence, all layer enhancement features are summed pixel by pixel according to preset weights to obtain the multi-layer enhancement side features.

5. The method for enhancing foggy images in intelligent transportation scenarios according to claim 4, characterized in that, The output features The formula for generating it is: ； ； ； ； in, For small-scale texture sub-features, It represents a large-scale texture sub-feature. As an aggregation feature, This represents the single-layer backbone features of the current level after batch regularization preprocessing. To add pixels one by one, To multiply pixel by pixel, To simplify convolution operations, This is the activation function.

6. The method for enhancing foggy images in intelligent transportation scenarios according to claim 4, characterized in that, The multi-layered enhanced side features Represented as: ； Where n is the total number of layers. For hyperparameters, This represents the layer-level enhancement feature of the i-th layer.

7. The method for enhancing foggy images in intelligent transportation scenarios according to claim 1, characterized in that, The sampling and multi-scale fusion correction of the single-layer backbone features at each level to obtain the corrected overall backbone features specifically includes: The process is iterated layer by layer from deep to shallow, and the corrected backbone features obtained after the previous iteration are sent into the upsampling unit to complete the scale magnification. The scaled-up corrected backbone features are added and fused pixel by pixel with the single-layer backbone features of the current level, and then passed through two levels of residual units to complete texture detail enhancement. The enhanced features are subtracted pixel by pixel from the corrected backbone features obtained after the previous iteration to complete difference correction. After correction, the features are sent to the selective kernel fusion unit to dynamically fuse multi-scale channel information to obtain the corrected backbone features after the current level iteration. Complete all iterative correction operations at each level, integrate the corrected backbone features after all levels of iteration, and summarize to form the corrected overall backbone features.

8. The method for enhancing foggy images in intelligent transportation scenarios according to claim 1, characterized in that, The method further includes: Construct teacher and student models; Collect real traffic scene foggy image samples, input them simultaneously into the teacher model and student model, and output the corresponding predicted feature map and fog-free reconstructed image respectively; The feature distribution output by the teacher model and the grayscale distribution of the reconstructed image are used as soft labels, and combined with hard labels composed of clear real-world images to construct a composite loss function. The weights of the convolutional and fusion units of the student model are iteratively updated by backpropagation using a composite loss function, continuously reducing the distribution difference between the output results of the student model and the teacher model, and completing the real distillation training. After training, the student model weights are fixed and used to process foggy images output by traffic data collection equipment in real time.

9. An enhancement processing device for foggy images in intelligent transportation scenarios, characterized in that, The enhanced processing device for foggy images in the intelligent transportation scenario includes: a processor and a memory for storing processor-executable instructions; The processor is configured to perform the enhancement processing method for foggy images in a smart transportation scenario as described in any one of claims 1-8.

10. A non-volatile computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which are executed by one or more processors to perform the enhancement processing method for foggy images in intelligent transportation scenarios as described in any one of claims 1-8.

Images