Traffic sign detection method and system based on edge-guided hypergraph and detail preservation

Abstract

The present application belongs to the technical field of computer vision and deep learning, and discloses a traffic sign detection method and system based on edge-guided hypergraph and detail preservation. The present application constructs an edge-guided hypergraph module, optimizes the similarity measurement between nodes by using the edge structure information of the traffic sign, and combines the trainable temperature parameter and dynamic threshold control to enhance the modeling ability of the model for long-range spatial dependence and key structure information. A detail preservation interaction module is constructed, a multi-stage bidirectional interaction mechanism is used, and a detail preservation interaction module is designed to enhance the feature perception ability of the model for low-resolution traffic signs. A channel separation strategy is adopted to divide the feature channels into enhanced feature flow and preserved feature flow, so that the edge-guided hypergraph module and the detail preservation interaction module are compatible, and the model parameter amount is reduced. The present application can improve the spatial modeling and cross-scale perception ability of the target, thereby improving the overall detection accuracy of the model and improving the detection accuracy and robustness.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and deep learning technology, and in particular relates to a traffic sign detection method and system based on edge-guided hypergraph and detail preservation. Background Technology

[0002] Traffic sign detection is a crucial component of Intelligent Transportation Systems (ITS) and autonomous driving. Its goal is to accurately locate and identify traffic signs in complex road scenes, providing a reliable basis for path planning, behavioral decisions, and traffic rule understanding. Compared to general object detection tasks, traffic sign detection typically faces more significant challenges: First, traffic signs often occupy only a small area in road images, and after multiple downsampling operations, key details such as texture, edges, and internal symbols are easily degraded or even lost. Second, interfering factors in complex backgrounds, such as leaves, branches, power lines, and billboards, can easily confuse with the target area, affecting the detector's accurate localization and representation of traffic signs. Third, under conditions of small scale, long distance, or image degradation, some traffic signs exhibit weakened internal discriminative cues, such as reduced color contrast, blurred local symbols, or overlapping structural details, increasing the difficulty of fine-grained category differentiation. Furthermore, some samples are often affected by the combined effects of multiple adverse factors, further enhancing the challenge of robust traffic sign detection.

[0003] Existing traffic sign detection methods have undergone considerable research to address the aforementioned challenges. For example, in 2023, Xia et al. proposed DSRA-DETR, which enhances multi-scale feature representation by introducing dilated convolutions into low-level features and employing residual aggregation to mitigate the loss of small target information during downsampling of traffic signs. In 2025, Ren et al. proposed DHPE-DETR, which improves the detector's ability to identify details and spatially locate traffic signs by fusing high-frequency detail information with low-frequency positional coding information. In 2022, Chen et al. proposed TSR-SA, focusing on improving the real-time detection accuracy of small-sized traffic signs. In 2023, Wang et al. proposed CDFF, which improves the accuracy of traffic sign recognition through detail compensation and feature enhancement. However, most existing technologies most similar to this approach are still based on convolutional feature enhancement, general feature fusion, or local detail compensation, and still have significant shortcomings in traffic sign detection scenarios. Specifically, under complex background conditions, existing methods still fall short in modeling the key contours, edge responses, and relationships between structural regions of traffic signs. This leads to the introduction of background interference such as leaves, branches, wires, and billboards during feature aggregation, weakening the structural representation of the target region. For small-scale traffic signs, although existing methods attempt to supplement shallow details or enhance multi-scale representations, fine-grained discriminative cues such as internal color contrast, local symbols, arrow textures, and edge structures are easily weakened or even lost during cross-layer transmission and deep semantic fusion. The fundamental reason is that existing technologies lack a collaborative mechanism that can simultaneously consider structural relationship modeling in complex backgrounds and the continuous preservation of details in small targets, making it difficult to unify structural enhancement and detail retention. To address these shortcomings, this solution aims to resolve the problems of insufficient modeling of traffic sign structural information in complex backgrounds, easy degradation of fine-grained details of small-scale targets, and the difficulty in simultaneously addressing both.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0005] (1) Severe loss of texture and edge information in small-scale traffic signs: Existing detection methods based on convolutional neural networks (such as Faster R-CNN) and Transformers suffer from a gradual weakening or loss of high-frequency information (such as sign edges, internal arrows, and text outlines) when processing small-scale traffic signs in road images due to multiple downsampling operations in the network. This makes it difficult for the model to accurately distinguish traffic sign categories with similar structures or colors, and it is prone to missed detections or localization errors for small targets at a distance.

[0006] (2) Insufficient ability to model structural information in complex backgrounds and susceptible to interference from background noise: Traffic signs are often located in complex background environments such as leaves, utility poles, billboards, and buildings. Existing methods (especially the Transformer model based on self-attention) often treat background areas and target areas equally when calculating global feature associations, which leads to background noise being incorrectly included in the relationship aggregation process. This weakens the feature representation of the key structures of the traffic sign itself (such as circles, triangles, rectangular borders and internal graphics). Therefore, the model is prone to false detection in complex scenes, misidentifying non-target objects in the background as traffic signs.

[0007] (3) Existing RT-DETR series methods are not suitable for traffic sign detection: Real-time detection Transformer (RT-DETR) shows a good balance between accuracy and speed in general object detection tasks, but its application research in traffic sign detection scenarios is still limited. Due to the characteristics of traffic signs, such as extreme scale (especially the high proportion of small objects), high inter-class similarity, and strong background interference, when RT-DETR or its variants (such as RT-DETRv2) are used directly for detection, the multi-scale feature interaction and self-attention mechanism in its encoder fail to specifically enhance edge structure and retain detailed information, resulting in weak perception ability of fine-grained features of traffic signs, which is difficult to meet the needs of high-precision autonomous driving environment perception.

[0008] (4) There is a conflict between high-order relation modeling and local detail preservation: In existing methods, graph structures or hypergraph structures (such as ordinary hypergraph convolution) introduced to enhance long-range dependencies can improve global semantic consistency, but continuous graph propagation operations will gradually weaken the spatial heterogeneity of feature maps, leading to homogenization of texture and edge information in different regions. This causes the model to further lose key local details of small target traffic signs while enhancing structural modeling, that is, semantic enhancement and detail preservation are difficult to balance. Summary of the Invention

[0009] To overcome the problems existing in related technologies, the present invention discloses a traffic sign detection method and system based on edge-guided hypergraph and detail preservation, the technical solution of which is as follows:

[0010] This invention is implemented as follows: A traffic sign detection method based on edge-guided hypergraph and detail preservation includes the following steps:

[0011] S1. A channel separation strategy is applied to the feature map output by the backbone network. This channel separation strategy divides the feature channels into an enhanced feature stream and a retained feature stream. The enhanced feature stream is used to enhance structural relationships, and the retained feature stream is used to retain the original spatial distribution information. This alleviates the feature conflict between global structural modeling and local detail preservation, and enables the edge-guided hypergraph module and the detail preservation interaction module to be compatible.

[0012] We propose an Edge-Guided Hypergraph Module (EGHM), which uses the edge structure of traffic signs to optimize the similarity measure between nodes. It also combines trainable temperature parameters and dynamic threshold control to model long-range spatial dependencies and key structural information in the enhanced feature flow, thereby improving the model's ability to represent the contours, edges, and structural region relationships of traffic signs in complex backgrounds.

[0013] S2. Construct an edge-guided hypergraph module, which is applied to the feature map corresponding to the enhanced feature flow in the channel separation strategy. This edge-guided hypergraph module uses the edge structure information of traffic signs to optimize the similarity measurement between nodes. Combined with trainable temperature parameters and dynamic threshold control, it enhances the long-range spatial dependence in the input feature map and uses the edge structure information to strengthen the key contours and internal structural features of traffic signs.

[0014] A Detail-Preserving Interaction Module (DPIM) is constructed, and a Detail Preservation Module (DP-Module) is designed to collaboratively fuse shallow detail features and deep semantic features, thereby enhancing the model's ability to continuously perceive and recover low-resolution traffic sign textures, edges, and internal discrimination cues.

[0015] S3. Construct a detail-preserving interaction module and embed it into the encoder; the detail-preserving interaction module enhances the model's ability to perceive and restore the texture and edge details of traffic signs in deep feature maps whose resolution is reduced due to downsampling through a multi-stage bidirectional interaction mechanism.

[0016] The Channel Separation Strategy (CSS) is designed to divide the feature channels into an enhanced feature stream and a preserved feature stream, providing a collaborative feature processing foundation for EGHM and DPIM. The enhanced feature stream is used to guide the hypergraph module in structural relationship modeling at the edge, while the preserved feature stream is used to retain the original local spatial details, avoiding excessive weakening of detail information during the structural enhancement process. This provides a more complete feature foundation for detail interaction and recovery in the subsequent encoding stage, enabling the three to work together to solve the problems of complex background interference and easy degradation of small target details.

[0017] Preferably, step S1 specifically includes the following steps:

[0018] S11. Given an input feature map from a certain stage of the backbone network. EGHM first extracts two types of features in parallel: one for relationship modeling and the other for edge structure extraction. Specifically, it employs... Convolution projects input features into a low-dimensional semantic space to obtain features. Simultaneously, edge response maps are generated through convolution. and use The function normalizes it to highlight the outline and local geometry of the traffic sign. To control computational overhead, the features are... Input feature map and edge response map Adaptive average pooling is performed to obtain low-resolution features. , and Then Spatial location unfolded into a node feature matrix Based on this, the Euclidean distance matrix is ​​first calculated according to the node characteristics. And introduce trainable temperature parameters The basic similarity is constructed using an exponential kernel function, with the specific formula as follows:

[0019]

[0020] In the formula, For nodes and nodes The basic similarity between them For the first The node and the first Euclidean distance between nodes This is a trainable temperature parameter used to adjust the degree of influence of Euclidean distance on the basic similarity; when As the basic similarity increases, its sensitivity to changes in the distance between nodes decreases. When the basic similarity is reduced, its sensitivity to changes in the distance between nodes increases; It is a constant. The range of values ​​is More preferably This is used to avoid division by zero errors;

[0021] S12. To suppress spurious associations caused by complex backgrounds, edge structures are used to adjust similarity. Specifically, the edge map... Expand into vectors And construct the edge coordination matrix through outer product. Subsequently, this is fused element-wise with the basic similarity to obtain the edge-guided similarity matrix. :

[0022]

[0023] In the formula, For element-wise multiplication, This is the edge collaboration matrix.

[0024] S13. Direct propagation on the fully connected graph is not only computationally expensive but also prone to causing feature oversmoothing. Therefore, a dynamic thresholding strategy is used to sparsify the similarity matrix, resulting in a hypergraph association matrix. :

[0025]

[0026] In the formula, For nodes and nodes The connection between them The basic distance threshold is set in advance based on the distance distribution between nodes or determined based on the statistical results of training data. These are trainable parameters used to dynamically adjust the basic distance threshold so that the sparsity of the similarity matrix can adapt to the distribution of input features, thereby suppressing weak correlation noise connections while preserving the effective correlation of key structural regions. 1 indicates a connection, and 0 indicates no connection.

[0027] S14. Obtaining the sparse correlation matrix Then, in node features Hypergraph propagation is then used to integrate higher-order contextual information, resulting in enhanced node features. :

[0028]

[0029] In the formula, and These are the angle matrices of the nodes and hyperedges, respectively; It is a trainable projection matrix;

[0030] S15. Rearrange the enhanced node features and upsample them to the original resolution to obtain the enhanced features. To mitigate the distributional bias between enhanced and original features, and to avoid over-enhancement damaging local details, a residual fusion method with learnable weights is used to obtain the output. :

[0031]

[0032]

[0033] In the formula, for After convolution, and Features after activation function processing For learnable scalar parameters, To integrate weights, for function.

[0034] Preferably, step S2 specifically includes the following steps:

[0035] S21. First, the spatial details of the shallow feature map... Semantic information of deep feature maps The process involves enhancing both local detail perception and global context understanding capabilities, followed by fusion. The local feature extraction branch... use Capture local information, then use Expanding the receptive field to incorporate broader contextual information enhances the perception of traffic sign edges and internal textures. The global context extraction branch is... use and This process involves aggregating global spatial information to form a global feature representation, which is then mapped to global attention features using convolution, enabling the model to perceive the overall structure of the target and the scene context. The extracted local detail features are then summed element-wise with the global context features. Information exchange is then facilitated using a cross-scale interaction module, and finally, convolution is used to restore the dimensionality, yielding the fused features. The specific formula is as follows:

[0036]

[0037]

[0038]

[0039] In the formula, and These are depthwise separable convolution and dilated depthwise separable convolution, respectively. and These are global average pooling and global max pooling, respectively. This is for channel splicing operations. for convolution, Composed of multiple Cross-scale interaction module for convolutional compositing. To sum element by element;

[0040] S22. Introduce a spatial selection mechanism based on fusion characteristics. Two sets of spatial weight maps are adaptively generated to adjust the response strengths of shallow and deep features, respectively. Specifically, After convolutional mapping, spatial weights corresponding to the two types of input features are generated. Normalization constraints are then applied to ensure that these two weights compete at the same location, thus reflecting the relative contribution of shallow details to deep semantics. The specific formula is as follows:

[0041]

[0042]

[0043]

[0044] In the formula, for The calculated spatial selection weights, for function, This is for element-wise multiplication.

[0045] S23 and DPIM enhance the cross-layer interaction between shallow detail features and deep semantic features through a bidirectional gating mechanism. Unlike unidirectional feature enhancement, this stage modulates shallow detail features with deep semantic features and uses shallow detail features to inversely correct deep semantic features. The former suppresses noise responses in shallow features, making them more focused on semantically relevant regions; the latter injects edge and texture cues into higher-level representations to improve the boundary clarity and spatial localization capabilities of semantic features. To further enhance the coupling between the two types of features, multiplication fusion is used to promote information flow between the two types of features, resulting in interactive features. The two enhanced features and the interactive features are then aggregated in a three-way manner to obtain the final fused feature. The specific formula is as follows:

[0046]

[0047]

[0048]

[0049] In the formula, express Activation function and These are spatial gating weights generated based on semantic and detail features, respectively. These are shallow enhancement features modulated from deep semantic features. The deep enhancement feature is generated by reverse modulation of shallow detail features. The bidirectional gating mechanism refers to generating modulation weights for shallow detail features using deep semantic features and generating modulation weights for deep semantic features using shallow detail features, thereby achieving feature selection and enhancement in two directions. The cross-layer interaction refers to the information transmission, mutual modulation, and fusion process between shallow detail features and deep semantic features. The three-way aggregation refers to aggregating shallow enhancement features, deep enhancement features, and the interaction features obtained by multiplying the two to obtain the final fused feature that simultaneously contains local detail information and high-level semantic information.

[0050] S24. In the final design detail optimization stage of the module, the interaction features are further corrected by combining spatial attention and residual enhancement. Specifically, spatial attention is first used to enhance the response of regions sensitive to local structures; then, the enhanced features are... The data is fed into the Detail Preservation Module (DP-Module) to further extract and enhance high-frequency texture information. The DP-Module employs a design combining convolutional residual extraction and channel recalibration. On one hand, it mines fine-grained responses corresponding to internal structures such as numbers and arrows through multi-layer convolution; on the other hand, it adaptively emphasizes key textures using channel weights, thereby further restoring local details of small objects without compromising overall semantic consistency. Finally, a gating mechanism is used for weighted processing, and... Adding them together yields the output features. The specific formula is as follows:

[0051]

[0052]

[0053]

[0054] In the formula, for Activation function , They represent , convolution, , Indicates the weights of the fully connected layer. Indicates global average pooling. This indicates a gating mechanism.

[0055] Preferably, step S3 specifically includes the following steps:

[0056] S31. Divide the feature channels into raw data branches and enhanced data branches in a 4:6 ratio.

[0057] S32. The original data branch keeps 40% of the feature channels unchanged, preserving their spatial distribution; the enhanced data branch inputs 60% of the features into EGHM for feature enhancement.

[0058] S33. Restore the feature channels of the original data branch and the enhanced data branch.

[0059] Another objective of this invention is to provide a traffic sign detection system based on edge-guided hypergraph and detail preservation. This system is used to implement the aforementioned traffic sign detection method based on edge-guided hypergraph and detail preservation. The system includes:

[0060] An edge-guided hypergraph module, embedded in the encoder of an RT-DETR or variant framework, is used to optimize the similarity measurement between nodes by utilizing the edge structure information of traffic signs, and to enhance the modeling of long-range spatial dependencies in the input feature map by combining trainable temperature parameters and dynamic threshold control, while using the edge structure information to strengthen the key contours and internal structural features of traffic signs.

[0061] The detail-preserving interaction module, embedded in the encoder of the RT-DETR or variant framework, is used to enhance the model's ability to perceive and recover the texture and edge details of traffic signs in deep feature maps that have been downsampled due to downsampling, through a multi-stage bidirectional interaction mechanism and by utilizing the detail-preserving interaction module.

[0062] The channel separation strategy module is used to divide the feature channels into enhanced feature streams and preserved feature streams, so that the edge-guided hypergraph module and the detail-preserving interaction module can be compatible.

[0063] Another object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the traffic sign detection method based on edge-guided hypergraph and detail preservation.

[0064] Combining all the above technical solutions, the beneficial effects of this invention are as follows:

[0065] First, this invention proposes a traffic sign detection method based on edge-guided hypergraph and detail-preserving architecture. An edge-guided hypergraph module (EGHM) is proposed, which optimizes the similarity measurement between nodes using the edge structure of traffic signs. Combined with trainable temperature parameters and dynamic threshold control, it models long-range spatial dependencies and key structural information in the enhanced feature flow, improving the model's ability to represent the contours, edges, and structural region relationships of traffic signs in complex backgrounds. A detail-preserving interaction module (DPIM) is constructed, using a multi-stage bidirectional interaction mechanism and designing a detail-preserving interaction module to collaboratively fuse shallow detail features and deep semantic features, enhancing the model's continuous perception and recovery of low-resolution traffic sign textures, edges, and internal discrimination cues. A channel separation strategy (CSS) is designed, dividing feature channels into enhanced feature flows and preserved feature flows, providing a collaborative feature processing foundation for EGHM and DPIM. The enhanced feature flow is used by the edge-guided hypergraph module for structural relationship modeling, while the preserved feature flow is used to preserve original local spatial details, avoiding excessive weakening of detail information during structural enhancement. This provides a more complete feature foundation for detail interaction and recovery in the subsequent encoding stage, enabling the three to collaboratively solve the problems of complex background interference and the easy degradation of small target details. This invention addresses the problem that there is currently no high-precision detection algorithm for traffic sign detection tasks based on deep learning.

[0066] Secondly, this invention proposes a traffic sign detection method based on an edge-guided hypergraph and detail-preserving interaction. In traffic sign detection, this method improves spatial modeling and cross-scale perception capabilities, thereby enhancing overall detection accuracy and robustness. The method demonstrates high accuracy on various traffic sign datasets, providing environmental perception technology support for autonomous driving systems. It also exhibits excellent performance on general datasets, demonstrating its good generalization ability.

[0067] This invention addresses the problem in traffic sign detection scenarios where RT-DETRv2 suffers from interference from complex backgrounds such as leaves, power lines, and billboards, which can easily aggregate with the target region, thus weakening its feature discrimination capabilities. An EGHM module is constructed to optimize the similarity measurement between nodes using the edge structure of traffic signs. Combined with trainable temperature parameters and dynamic threshold control, this enhances the model's ability to model long-range spatial dependencies and key structural information, enabling the model to focus on critical regions and reduce interference from irrelevant background noise.

[0068] To address the issues of blurred boundaries and unstable positioning of small-scale traffic signs in deep features, this invention designs a DPIM module. It employs a multi-stage bidirectional interaction mechanism and incorporates a detail-preserving interaction module to restore downsampled and blurred high-frequency information, enhance the texture representation of small-scale traffic signs, and improve the model's ability to perceive the features of low-resolution traffic signs.

[0069] This invention addresses the issue of EGHM's continuous graph convolution operations progressively reducing the spatial heterogeneity of feature maps, leading to homogenization of feature representations. A CSS strategy is designed to divide feature channels into enhanced and preserved feature flows. Structural relationship modeling is performed on the enhanced channels to achieve effective compatibility between EGHM and DPIM, improving model accuracy while reducing the number of model parameters.

[0070] Third, the traffic sign detection method based on edge-guided hypergraph and detail preservation proposed in this invention can improve the accuracy and robustness of traffic sign detection in complex road scenarios, and has good engineering application prospects. This technology can be applied to intelligent transportation systems, driver assistance systems, autonomous driving perception systems, and vehicle-road cooperative scenarios, providing more reliable detection results for vehicle environmental perception, traffic rule understanding, and driving decision-making, thereby improving the safety, stability, and practical value of related products.

[0071] Existing traffic sign detection technologies mostly focus on multi-scale feature fusion, local detail enhancement, or general feature compensation. They lack a collaborative technical solution that can simultaneously address the problems of insufficient modeling of traffic sign structural relationships in complex backgrounds and the easy degradation of fine-grained discrimination information for small-scale targets. This invention provides a new technical path for the collaborative optimization of structural information modeling and continuous detail preservation in traffic sign detection through the combined design of an edge-guided hypergraph module, a detail-preserving interactive module, and a channel separation strategy.

[0072] The field of traffic sign detection has long faced challenges such as strong interference from complex backgrounds, easy loss of details from small-scale targets, and the difficulty in balancing structural enhancement and detail preservation. While existing methods have made improvements in some areas, they still struggle to simultaneously achieve both the ability to model structural relationships in complex backgrounds and the ability to represent fine-grained details of small targets. This invention proposes a synergistic solution to address these problems, alleviating to some extent the long-standing technical difficulties in existing technologies.

[0073] In existing technologies, it is generally believed that enhancing global structural relationship modeling is beneficial for improving semantic consistency, while fine-grained detail preservation relies more on shallow local features. These two approaches are often difficult to reconcile and may even conflict. This invention addresses this by implementing a channel separation strategy to separate the enhanced feature flow from the preserved feature flow, enabling effective compatibility between the edge-guided hypergraph module and the detail-preserving interaction module. This approach improves structural modeling capabilities while maintaining local detail information, overcoming the conventional approach that struggles to unify structural enhancement and detail preservation. Attached Figure Description

[0074] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of the disclosure of this invention.

[0075] Figure 1 This is a flowchart of the traffic sign detection method based on edge-guided hypergraph and detail preservation provided in the embodiments of the present invention;

[0076] Figure 2 This is an overall framework diagram of the traffic sign detection method based on edge-guided hypergraph and detail preservation provided in the embodiments of the present invention;

[0077] Figure 3 This is an overall flowchart of EGHM provided in the embodiments of the present invention;

[0078] Figure 4 This is a structural diagram of the DPIM module provided in an embodiment of the present invention;

[0079] Figure 5 This is a visualization of traffic sign detection results for images 143 and 4107 in the TT100K dataset provided in this embodiment of the invention, wherein: Figure 5 (a) is YOLOv8-L, Figure 5 (b) is YOLOv10-L. Figure 5 (c) is YOLO-TS. Figure 5 (d) represents DEIM-DFINE-N. Figure 5 (e) is RT-DETRv2-R18. Figure 5 (f) is a visualization of the traffic sign detection results of the present invention;

[0080] Figure 6 This is a visualization result of traffic sign detection from the CCTSDB2021 dataset provided in this embodiment of the invention, wherein: Figure 6 (a) is YOLOv8-L, Figure 6 (b) is YOLOv10-L. Figure 6 (c) is DEIM-DFINE-N. Figure 6 (d) is a visualization result of traffic sign detection according to the present invention.

[0081] Figure 7 This is a heatmap visualization comparison diagram provided in the embodiment of the present invention; wherein, the S5 feature map enhanced by EGHM is selected as the target layer. Detailed Implementation

[0082] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0083] The innovation of this invention lies in addressing the shortcomings of existing traffic sign detection methods, such as insufficient modeling of structural relationships in complex backgrounds, easy degradation of fine-grained details of small-scale traffic signs during downsampling and cross-layer transfer, and the difficulty in simultaneously achieving structural enhancement and detail preservation. This invention proposes a collaborative detection method consisting of an edge-guided hypergraph module, a detail-preserving interaction module, and a channel separation strategy. Specifically, the edge-guided hypergraph module utilizes the edge structural information of traffic signs to constrain the similarity measurement between nodes, and combines trainable temperature parameters and dynamic threshold control to construct sparse high-order association relationships, thereby enhancing the ability to model long-range dependencies between traffic sign contours, edges, and structural regions in complex backgrounds. The detail-preserving interaction module enhances the model's continuous perception and recovery of low-resolution traffic sign textures, edges, and internal discrimination cues through multi-stage bidirectional interaction between shallow detail features and deep semantic features, combined with the detail-preserving interaction module. The channel separation strategy alleviates the feature conflict between structural enhancement and detail preservation by dividing features into enhanced feature streams and preserved feature streams, effectively enabling compatibility between the edge-guided hypergraph module and the detail-preserving interaction module. Therefore, this invention can simultaneously improve the ability to characterize structures in complex backgrounds and preserve fine-grained details of small targets, thereby improving the accuracy and robustness of traffic sign detection.

[0084] Example 1, such as Figure 1 As shown, the traffic sign detection method based on edge-guided hypergraph and detail preservation provided in this embodiment of the invention includes the following steps:

[0085] S1. A channel separation strategy is applied to the feature map output by the backbone network. This channel separation strategy divides the feature channels into an enhanced feature stream and a retained feature stream. The enhanced feature stream is used to enhance structural relationships, and the retained feature stream is used to retain the original spatial distribution information. This alleviates the feature conflict between global structural modeling and local detail preservation, and enables the edge-guided hypergraph module and the detail preservation interaction module to be compatible.

[0086] S2. Construct an edge-guided hypergraph module, which is applied to the feature map corresponding to the enhanced feature flow in the channel separation strategy. This edge-guided hypergraph module uses the edge structure information of traffic signs to optimize the similarity measurement between nodes. Combined with trainable temperature parameters and dynamic threshold control, it enhances the long-range spatial dependence in the input feature map and uses the edge structure information to strengthen the key contours and internal structural features of traffic signs.

[0087] S3. Construct a detail-preserving interaction module and embed it into the encoder; the detail-preserving interaction module enhances the model's ability to perceive and restore the texture and edge details of traffic signs in deep feature maps whose resolution is reduced due to downsampling through a multi-stage bidirectional interaction mechanism.

[0088] like Figure 2 As shown in the figure, the overall framework of the traffic sign detection method based on edge-guided hypergraph and detail preservation provided in this embodiment of the invention consists of three parts: a backbone network, an encoder, and a decoder. First, the input image is processed by a ResNet-18 backbone network to extract multi-scale feature maps S3, S4, and S5. During this process, a channel separation strategy is applied to the S4 and S5 feature maps output by the backbone network to split the feature channels, thereby alleviating feature conflicts between the edge-guided hypergraph module and the detail-preserving interaction module. The edge-guided hypergraph module is nested within the enhanced feature flow branch of the CSS (Structured Feature Flow) module, used for structural relationship modeling of the enhanced feature flow. Subsequently, the S3, S4, and S5 feature maps are input into the encoder. After DPIM processing, S3, S4, and S5 output sequence features in a unified semantic space. Finally, the sequence features are input into the IoU-aware Transformer decoder, which combines positional encoding to model spatial relationships, generate predicted representations of candidate regions, and output the detection results.

[0089] like Figure 3 As shown, the Edge Guided Hypergraph Module (EGHM) provided in this embodiment of the invention specifically includes:

[0090] RT-DETRv2 relies on self-attention and multi-scale feature interactions to model global relationships. However, in traffic sign detection scenarios, interfering regions such as leaves, power lines, and billboards in complex backgrounds easily participate in relationship aggregation with the target region, thus weakening the discriminative ability of features. To address this, this invention proposes an edge-guided hypergraph module. This module utilizes the stable contour structure information of traffic signs to constrain high-order relationship modeling, thereby constructing a sparse and adaptive hypergraph. The hypergraph is then used for efficient information propagation in the feature space. Unlike relationship modeling methods based solely on feature similarity, EGHM introduces edge structure into the node association construction process, making information propagation more focused on the target structure itself, rather than background noise.

[0091] Given an input feature map from a certain stage of the backbone network EGHM first extracts two types of features in parallel: one for relationship modeling and the other for edge structure extraction. Specifically, it employs... Convolution projects input features into a low-dimensional semantic space to obtain features. Simultaneously, edge response maps are generated through convolution. and use The function normalizes it to highlight the outline and local geometry of the traffic sign. To control computational overhead, [the function is used to] normalize it. , and Adaptive average pooling is performed to obtain low-resolution features. , and Then Spatial location unfolded into a node feature matrix Based on this, the Euclidean distance matrix is ​​first calculated according to the node characteristics. And introduce trainable temperature parameters The basic similarity is constructed using an exponential kernel function, with the specific formula as follows:

[0092] (1)

[0093] In the formula, For nodes and nodes The basic similarity between them For the first The node and the first Euclidean distance between nodes This is a trainable temperature parameter used to adjust the degree of influence of Euclidean distance on the basic similarity; when As the basic similarity increases, its sensitivity to changes in the distance between nodes decreases. When the basic similarity is reduced, its sensitivity to changes in the distance between nodes increases; It is a constant. The range of values ​​is To avoid division by zero errors; and to suppress spurious associations caused by complex backgrounds, edge structures are used to adjust similarity. Specifically, the edge map... Expand into vectors And construct the edge coordination matrix through outer product. This is then fused element-wise with the basic similarity to obtain the edge-guided similarity matrix. :

[0094] (2)

[0095] In the formula, For element-wise multiplication, This is the edge collaboration matrix. This design enhances the connection strength of nodes located within the same contour or local structure, allowing higher-order relationship modeling to focus more on the structural regions of traffic signs. Direct propagation on the fully connected graph is not only computationally expensive but also prone to feature oversmoothing. Therefore, a dynamic thresholding strategy is used to sparsify the similarity matrix, resulting in the hypergraph association matrix. :

[0096] (3)

[0097] In the formula, For nodes and nodes The connection between them The basic distance threshold is set in advance based on the distance distribution between nodes or determined based on the statistical results of training data. These are trainable parameters used to dynamically adjust the basic distance threshold, allowing the sparsity of the similarity matrix to adaptively change with the input feature distribution. This suppresses weak, noisy connections while preserving effective associations in key structural regions; 1 indicates a connection, and 0 indicates no connection. Through this process, each node only establishes associations with nodes that are semantically and structurally similar, thus achieving adaptive neighborhood selection. This process is repeated after obtaining the sparse association matrix. Then, in node features Hypergraph propagation is then used to integrate higher-order contextual information, resulting in enhanced node features. :

[0098] (4)

[0099] In the formula, and These are the angle matrices of the nodes and hyperedges, respectively; It is a trainable projection matrix;

[0100] This process enables information to flow along edge-guided higher-order association paths within the graph, thereby enhancing the structured representation of the target region. Finally, the enhanced node features are rearranged and upsampled to the original resolution to obtain the enhanced features. To mitigate the distributional bias between enhanced and original features, and to avoid over-enhancement damaging local details, a residual fusion method with learnable weights is used to obtain the output. The expression is:

[0101]

[0102] (5)

[0103] In the formula, for After convolution, and Features after activation function processing For learnable scalar parameters, To integrate weights, for This design can adaptively introduce higher-order contextual information guided by edges while preserving the original local details, thereby improving the structure perception capability in complex traffic scenarios.

[0104] like Figure 4 As shown, the Detail Preservation Interaction Module (DPIM) provided in this embodiment of the invention specifically includes:

[0105] In traffic sign detection tasks, small targets constitute a significant proportion of the dataset. Because continuous downsampling in deep networks weakens high-frequency information such as texture and edges, small-scale traffic signs often exhibit semantically sufficient but blurry boundaries and unstable localization in deep features. To address this, a detail-preserving interaction module is constructed. This module uses deep semantics to filter shallow noise while simultaneously sharpening deep semantic boundaries with high-frequency information from the shallow layers, thus preserving information about small targets. Furthermore, the detail-preserving interaction module at the end uses residual learning to further enhance weak texture responses, ensuring the integrity of traffic sign details. This module consists of four parts: multi-scale feature fusion, spatial selection, spatial interaction and cross-modulation, and detail-preserving optimization.

[0106] Multi-scale feature fusion: This stage aims to efficiently integrate feature information from different network layers. First, it integrates the spatial details of shallow feature maps. Semantic information of deep feature maps The process involves enhancing both local detail perception and global context understanding capabilities, followed by fusion. The local feature extraction branch... use Capture local information, then use Expanding the receptive field to incorporate broader contextual information enhances the perception of traffic sign edges and internal textures. The global context extraction branch is... use and This process involves aggregating global spatial information to form a global feature representation, which is then mapped to global attention features using convolution, enabling the model to perceive the overall structure of the target and the scene context. The extracted local detail features are then summed element-wise with the global context features. Information exchange is then facilitated using a cross-scale interaction module, and finally, convolution is used to restore the dimensionality, yielding the fused features. The specific formula is as follows:

[0107] (6)

[0108] (7)

[0109] (8)

[0110] In the formula, and These are depthwise separable convolution and dilated depthwise separable convolution, respectively. and These are global average pooling and global max pooling, respectively. This is for channel splicing operations. for convolution, Composed of multiple Cross-scale interaction module for convolutional compositing. To sum element by element;

[0111] Spatial selection: Directly fusing shallow and deep features easily overlooks the differences in importance between the two types of information at different spatial locations. Therefore, a spatial selection mechanism is introduced, based on the fused features... Two sets of spatial weight maps are adaptively generated to adjust the response strengths of shallow and deep features, respectively. Specifically, After convolutional mapping, spatial weights corresponding to the two types of input features are generated. Normalization constraints are then applied to ensure that these two weights compete at the same location, thus reflecting the relative contribution of shallow details to deep semantics. The specific formula is as follows:

[0112] (9)

[0113]

[0114] (10)

[0115] In the formula, for The calculated spatial selection weights, for function, This is for element-wise multiplication.

[0116] Spatial Interaction and Crossover: After spatial selection, DPIM enhances cross-layer interaction between shallow detail features and deep semantic features through a bidirectional gating mechanism. Unlike unidirectional feature enhancement, this stage modulates shallow detail features with deep semantic features and uses shallow detail features to inversely correct deep semantic features. The former suppresses noise responses in shallow features, making them more focused on semantically relevant regions; the latter injects edge and texture cues into higher-level representations to improve the boundary clarity and spatial localization capabilities of semantic features. To further enhance the coupling between the two types of features, multiplication fusion is used to promote information flow between the two types of features, resulting in interactive features. The two enhanced features and the interactive features are then aggregated in a three-way manner to obtain the final fused feature. The specific formula is as follows:

[0117]

[0118] (11)

[0119] (12)

[0120] In the formula, express Activation function and These are spatial gating weights generated based on semantic and detail features, respectively. These are shallow enhancement features modulated from deep semantic features. The deep enhancement feature is generated by reverse modulation of shallow detail features. The bidirectional gating mechanism refers to generating modulation weights for shallow detail features using deep semantic features and generating modulation weights for deep semantic features using shallow detail features, thereby achieving feature selection and enhancement in two directions. The cross-layer interaction refers to the information transmission, mutual modulation, and fusion process between shallow detail features and deep semantic features. The three-way aggregation refers to aggregating shallow enhancement features, deep enhancement features, and the interaction features obtained by multiplying the two to obtain the final fused feature that simultaneously contains local detail information and high-level semantic information.

[0121] Detail Preservation and Optimization: Although the features have incorporated both shallow and deep information after two-way interaction, further detail recovery and enhancement are still needed for weakly responsive structures such as numbers, arrows, and fine outlines in traffic signs. Therefore, a detail preservation and optimization stage is designed at the end of the module, using a combination of spatial attention and residual enhancement to further correct the interactive features. Specifically, spatial attention is first used to enhance the response of regions sensitive to local structures; subsequently, the enhanced features... The data is fed into a detail-preserving interaction module to further extract and enhance high-frequency texture information. This module employs a design combining convolutional residual extraction and channel recalibration. On one hand, it mines fine-grained responses corresponding to internal structures such as numbers and arrows through multi-layer convolution; on the other hand, it uses adaptive channel weights to emphasize key textures, thereby further restoring local details of small objects without compromising overall semantic consistency. Finally, a gating mechanism is used for weighted processing, and... Adding them together yields the output features. The specific formula is as follows:

[0122] (13)

[0123] (14)

[0124] (15)

[0125] In the formula, for Activation function , They represent , convolution, , Indicates the weights of the fully connected layer. Indicates global average pooling. This indicates a gating mechanism.

[0126] The Channel Separation Strategy (CSS) provided in this embodiment of the invention specifically includes:

[0127] EGHM enhances global semantic consistency through high-order relation propagation, but continuous graph augmentation can weaken the spatial heterogeneity of feature maps, thus hindering the preservation of texture and boundary information of small objects. To address this, this invention proposes a Channel Separation Strategy (CSS), dividing feature channels into original data branches and augmented data branches in a 4:6 ratio. The former directly preserves the original local representation, while the latter is fed into EGHM for structural relation modeling. Finally, the feature channels of the two paths are restored. The 4:6 ratio for dividing feature channels into original data branches and augmented data branches has been verified through ablation experiments, achieving an optimal balance between accuracy and parameter count (see Table X). This strategy not only alleviates the feature conflict between semantic augmentation and detail preservation, improving model accuracy, but also reduces computational complexity and parameter count, thus improving model efficiency.

[0128] Example 2: The traffic sign detection system based on edge-guided hypergraph and detail preservation provided in this embodiment of the invention includes:

[0129] An edge-guided hypergraph module, embedded in the encoder of an RT-DETR or variant framework, is used to optimize the similarity measurement between nodes by utilizing the edge structure information of traffic signs, and to enhance the modeling of long-range spatial dependencies in the input feature map by combining trainable temperature parameters and dynamic threshold control, while using the edge structure information to strengthen the key contours and internal structural features of traffic signs.

[0130] The detail-preserving interaction module, embedded in the encoder of the RT-DETR or variant framework, is used to enhance the model's ability to perceive and recover the texture and edge details of traffic signs in deep feature maps that have been downsampled due to downsampling, through a multi-stage bidirectional interaction mechanism and by utilizing the detail-preserving interaction module.

[0131] The channel separation strategy module is used to divide the feature channels into enhanced feature streams and preserved feature streams, so that the edge-guided hypergraph module and the detail-preserving interaction module can be compatible.

[0132] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solutions.

[0133] This invention was validated on the TT100K, CCTSDB2021, and MS-COCO datasets. The TT100K dataset, composed of 100,000 images from Tencent Street View Maps with a resolution of 2048×2048, includes 10,000 labeled images with approximately 30,000 traffic signs. TT100K contains 150 traffic sign categories. Following the setup of Zhu et al., categories with fewer than 100 samples were removed, adjusting the number of categories to 45. The training set contains 6,105 images with approximately 15,000 traffic sign instances. The test set contains 3,071 images with 7,070 traffic sign instances. The CCTSDB2021 dataset, created by Changsha University of Science and Technology, is a traffic sign detection dataset containing 17,856 images for both training and testing. Traffic signs are categorized into three types: mandatory, warning, and cautionary. The training set contains 16,356 images, and the testing set contains 1,500 images. MS-COCO is a commonly used large-scale general-purpose dataset in object detection, containing 80 object categories. The training set contains 118,287 images, and the validation set contains 5,000 images. TT100K was used for ablation experiments, and TT100K, CCTSDB2021, and MS-COCO were used for comparative experiments. The experiments used Precision, Recall, F1 score, mAP50, mAP50–95, and speed (FPS) to evaluate detection performance. Specific definitions of these evaluation metrics are provided in the supplementary materials. To make the analysis more comprehensive, this paper also presents mAP50–95, mAPm50–95, and mAPl50–95.

[0134] The experimental environment for this invention was an Ubuntu 20.04 operating system, an Intel® Xeon® Silver4310 CPU, and an NVIDIA GeForce RTX 3090 GPU. The model training batch size was set to 8, with 200 training epochs on the TT100K dataset and 50 training epochs on the CCTSDB2021 dataset. The AdamW optimizer was selected for parameter optimization.

[0135] In specific implementation examples, the datasets used were the traffic sign dataset TT100K, CCTSDB2021, and the general dataset MS-COCO. Our method was compared with representative object detection algorithms on these datasets to verify the effectiveness of the proposed model. To ensure fairness in the comparison experiments, all comparison models adopted their officially recommended training strategies and hyperparameter configurations, and were retrained or tested under the same dataset partitioning conditions; the partitioning of the training set, validation set, and test set remained consistent. All experiments were conducted in the same hardware and software environment to ensure the comparability and objectivity of the experimental results.

[0136] Table 1. Comparison results of the present invention and a general object detection model on the TT100K dataset.

[0137]

[0138] Table 1 shows the comparison results with general object detection models on the TT100K dataset. As can be seen from Table 1, the proposed method demonstrates strong competitiveness across all metrics. Specifically, it achieves 85.4% Precision and 87.0% Recall, reaching 92.2% on the key metric mAP50. This result indicates that the proposed method can maintain high classification and discrimination capabilities and localization accuracy in complex traffic scenarios. In terms of efficiency, the proposed method achieves 41 FPS. Although pure YOLO route analysis has an advantage in inference speed, this method maintains deployable real-time inference speed while achieving higher detection accuracy, demonstrating that the introduction of structural relationship modeling and detailed interaction mechanisms does not incur unacceptable additional overhead. Figure 5 The results of multiple models and Ours on the TT100K dataset are shown. Figure 5 (a) is YOLOv8-L, Figure 5 (b) is YOLOv10-L. Figure 5 (c) is YOLO-TS. Figure 5 (d) represents DEIM-DFINE-N. Figure 5 (e) is RT-DETRv2-R18. Figure 5(f) is a visualization of the traffic sign detection results of this invention. Figure 5 It can be observed that in scenarios involving small targets at long distances and complex background interference, this method can detect traffic signs more accurately and reduce missed detections and false detections to some extent.

[0139] Table 2. Comparison of experimental results between the present invention and a general object detection model on the CCTSDB2021 dataset.

[0140]

[0141] Table 2 presents the comparative experimental results on the CCTSDB2021 dataset. On CCTSDB2021, the method also demonstrates stable performance, achieving an mAP50 of 88.3%, a Precision of 91.6%, a Recall of 79.0%, and an F1 score of 0.849. Figure 6 This section showcases the detection results of other models and Ours on the CCTSDB2021 dataset, including... Figure 6 (a) is YOLOv8-L, Figure 6 (b) is YOLOv10-L. Figure 6 (c) is DEIM-DFINE-N. Figure 6 (d) is a visualization result of traffic sign detection according to the present invention. Figure 6 This demonstrates that the proposed method not only maintains high detection accuracy but also possesses excellent cross-dataset generalization ability. Furthermore, achieving this performance with only 25.87M parameters and a computational scale of 88.0 GFLOPs reflects a balance between accuracy and parameter count. Even in adverse weather conditions and with small targets, the model can deliver reliable outputs with manageable complexity, thus ensuring the safety of autonomous driving.

[0142] Table 3. Comparison results of the present invention and the general object detection model on the MS-COCO dataset.

[0143]

[0144] To further verify the versatility and cross-scene generalization ability of the proposed method, comparative experiments were conducted on the MS-COCO dataset, and the results are shown in Table 3. The proposed method achieved 65.8% mAP50, 48.3% mAP50-95, and 31.1% mAPs50-95 on MS-COCO. Compared to the traffic sign dataset, MS-COCO contains more categories and more complex scenes, placing higher demands on the detector's general representation capabilities. Therefore, this result demonstrates that the proposed method is not only applicable to traffic sign scenes but also exhibits good detection performance in general object detection tasks. In particular, the performance on the small object metric mAPs50-95 indicates that the proposed method maintains strong localization capabilities for small-scale objects under complex category distributions and diverse scenes, verifying that the edge-guided structure modeling and detail-preserving interaction mechanism have a certain degree of out-of-scene generalization ability.

[0145] Table 4 Comparison of ablation experimental results in dataset TT100K

[0146]

[0147] (1) Effectiveness of the edge-guided hypergraph module: As shown in Table 4, when using only EGHM, mAP50 increased from 88.8% to 90.8% of the baseline, mAP50–95 increased from 69.0% to 70.4%, the number of parameters increased from 20.14M to 21.30M, and GFLOPs increased from 61.35 to 66.12. This result shows that EGHM maintains high computational efficiency while improving accuracy. To verify the ability of EGHM to focus on key regions, the S5 feature map enhanced by EGHM was selected as the target layer for heatmap visualization, as shown in the figure. Figure 7 As shown, the top image is the heatmap of the baseline model, and the bottom image is the heatmap after adding EGHM to S4 and S5 on the baseline model. This indicates that EGHM makes the model focus more on the traffic sign area and suppresses noise interference from complex backgrounds.

[0148] (2) Effectiveness of the detail-preserving interaction module: As shown in Table 4, after adding DPIM, mAP50 reached 91.5%, mAP50–95 reached 70.8%, and the small target index mAPs50–95 increased from 55.4% to 56.2%. This result shows that DPIM can improve the model's ability to preserve small target information and more accurately locate the position of small targets. To further verify the effectiveness of the DP-Module at the end of DPIM, ablation experiments were conducted on only the DP-Module under the same settings, as shown in Table 5. After adding DP-Module, mAP50 increased from 91.2% to 91.5%, mAP50–95 increased from 70.6% to 70.8%, and brought a slight improvement in the performance of small targets, mAPs50–95, from 56.1% to 56.2%. This indicates that the module can amplify weak texture responses and enhance the integrity of the representation of internal structures such as numbers and arrows through high-frequency detail convolution and gating mechanism.

[0149] Table 5. Results of DMSIA ablation experiments on dataset TT100K

[0150]

[0151] 3) Effectiveness of the channel separation strategy: As shown in Table 4, when EGHM and DPIM are introduced simultaneously, mAP50 is 91.5% and Map50-95 is 71.0%. After adding EGHM on top of DPIM, mAP50-95 drops from 56.2% to 55.9%, indicating that the global semantic enhancement of EGHM destroys the local detail preservation of DPIM, and the two conflict at the deep feature level. When CSS is further added on top of EGHM+DPIM, the model achieves the best performance across the entire table, as shown in Table 4, mAP50 reaches 92.2%, mAP50-95 reaches 71.7%, small target mAP50-95 reaches 57.5%, medium target mAP50-95 reaches 77.5%, and large target mAP50-95 reaches 87.0%. Meanwhile, the model parameters were reduced from 26.57M to 25.87M, and Gflops were reduced from 90.23 to 88.14, indicating that CSS improves accuracy while mitigating the increased computational overhead of the two modules. Experimental results show that CSS achieves effective compatibility between EGHM and DPIM by retaining some channels as the original detail stream and others as the semantic enhancement stream.

[0152] 4) Analysis of Channel Splitting Ratio: To verify the most effective channel splitting ratio, experiments were designed, changing only the channel splitting ratio under the same settings. As shown in Table 6, when the enhancement branch ratio was adjusted from 0.5 to 0.6, the model performance further improved, reaching its optimum at 0.6, with mAP50 at 92.2% and mAP50–95 at 71.7%. Further increasing to 0.7 caused mAP50 to drop back to 91.8% and mAP50–95 to drop back to 71.2%. This indicates that a higher enhancement branch ratio is not always better. A ratio that is too low limits the global modeling of EGHM, while a ratio that is too high leads to homogenization, compressing the expression space of detail flow and weakening the preservation of small object texture and boundary information. Experimental results show that a 4:6 split achieves the best balance between EGHM and DPIM, fully leveraging the long-range dependency modeling advantages of EGHM while ensuring the preservation of small object texture information by DPIM.

[0153] Table 6. Results of channel partitioning ablation experiments on dataset TT100K.

[0154]

[0155] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A traffic sign detection method based on edge-guided hypergraph and detail preservation, characterized in that, This method is implemented based on the RT-DETR or variant framework and includes the following steps: S1. A channel separation strategy is applied to the feature map output by the backbone network. This channel separation strategy divides the feature channels into an enhanced feature stream and a retained feature stream. The enhanced feature stream is used to enhance structural relationships, and the retained feature stream is used to retain the original spatial distribution information. This alleviates the feature conflict between global structural modeling and local detail preservation, and enables the edge-guided hypergraph module and the detail preservation interaction module to be compatible. S2. Construct an edge-guided hypergraph module, which is applied to the feature map corresponding to the enhanced feature flow in the channel separation strategy. This edge-guided hypergraph module uses the edge structure information of traffic signs to optimize the similarity measurement between nodes. Combined with trainable temperature parameters and dynamic threshold control, it enhances the long-range spatial dependence in the input feature map and uses the edge structure information to strengthen the key contours and internal structural features of traffic signs. S3. Construct a detail-preserving interaction module and embed it into the encoder; the detail-preserving interaction module enhances the model's ability to perceive and restore the texture and edge details of traffic signs in deep feature maps whose resolution is reduced due to downsampling through a multi-stage bidirectional interaction mechanism.

2. The traffic sign detection method based on edge-guided hypergraph and detail preservation according to claim 1, characterized in that, In step S1, the channel separation strategy includes: The feature channels are divided into an original data branch and an enhanced data branch in a 4:6 ratio. The original data branch keeps 40% of the feature channels unchanged, preserving the spatial distribution of the channels. The enhanced data branch inputs 60% of the features to the edge-guided hypergraph module for feature enhancement. The feature channels processed by the original data branch and the enhanced data branch are then concatenated.

3. The traffic sign detection method based on edge-guided hypergraph and detail preservation according to claim 1, characterized in that, In step S2, an edge-guided hypergraph module is constructed, and the feature map corresponding to the enhanced feature flow in the channel separation strategy includes: For the input feature map The edge-guided hypergraph module extracts relationship modeling features and edge structure extraction features in parallel; it employs... Convolution projects input features into a low-dimensional semantic space to obtain features. Edge response maps are generated through convolution. and use Normalize the function; Features Input feature map and edge response map Adaptive average pooling is performed to obtain low-resolution features. , and Then Spatial location unfolded into a node feature matrix Calculate the Euclidean distance matrix based on node characteristics. And introduce trainable temperature parameters The basic similarity is constructed using an exponential kernel function, and the calculation formula is as follows: In the formula, For nodes and nodes The basic similarity between them For the first The node and the first Euclidean distance between nodes This is a trainable temperature parameter used to adjust the degree of influence of Euclidean distance on the basic similarity; when As the basic similarity increases, its sensitivity to changes in the distance between nodes decreases. When the basic similarity is reduced, its sensitivity to changes in the distance between nodes increases; It is a constant. The range of values ​​is This is used to avoid division by zero errors; Adjusting similarity using edge structure: [Edge map] Expand into vectors Constructing the marginal cooperative matrix through outer product This similarity matrix is ​​then fused element-wise with the basic similarity to obtain an edge-guided similarity matrix. : In the formula, For element-wise multiplication, This is the edge collaboration matrix.

4. The traffic sign detection method based on edge-guided hypergraph and detail preservation according to claim 3, characterized in that, In step S2, the dynamic threshold control includes: A dynamic thresholding strategy is used to sparsify the similarity matrix to obtain the hypergraph association matrix. : In the formula, For nodes and nodes The connection between them The basic distance threshold is set in advance based on the distance distribution between nodes or determined based on the statistical results of training data. These are trainable parameters used to dynamically adjust the basic distance threshold so that the sparsity of the similarity matrix can adapt to the distribution of input features, thereby suppressing weak correlation noise connections while preserving the effective correlation of key structural regions. 1 indicates a connection, and 0 indicates no connection.

5. The traffic sign detection method based on edge-guided hypergraph and detail preservation according to claim 4, characterized in that, Combining trainable temperature parameters with dynamic threshold control, enhanced modeling of long-range spatial dependencies in the input feature map includes: In obtaining the sparse correlation matrix Then, in node features Hypergraph propagation is then used to integrate higher-order contextual information, resulting in enhanced node features. The expression is: In the formula, and These are the angle matrices of the nodes and hyperedges, respectively; It is a trainable projection matrix; The enhanced node features are rearranged and upsampled to the original resolution to obtain the enhanced features. The output is obtained by using a residual fusion method with learnable weights. The expression is: In the formula, for After convolution, and Features after activation function processing For learnable scalar parameters, To integrate weights, for function.

6. The traffic sign detection method based on edge-guided hypergraph and detail preservation according to claim 1, characterized in that, In step S3, constructing the detail-preserving interaction module includes: Multi-scale feature fusion: integrating spatial details from shallow feature maps Semantic information of deep feature maps The processing enhances both local detail perception and global context understanding capabilities, followed by fusion; the local feature extraction branch... use Capture local information and reuse it Expanding the receptive field and fusing contextual information enhances the perception of traffic sign edges and internal textures; the global context extraction branch, through... use and Global spatial information is aggregated to form a global feature representation, which is then mapped to global attention features using convolution. The extracted local detail features are then summed element-wise with the global context features, and finally, a cross-scale interaction module is used. Information is exchanged, and finally convolution is used to restore the dimensions to obtain the fused features. ; In the formula, and These are depthwise separable convolution and dilated depthwise separable convolution, respectively. and These are global average pooling and global max pooling, respectively. This is for channel splicing operations. for convolution, Composed of multiple Cross-scale interaction module for convolutional compositing. To sum element by element; Spatial selection: Introducing a spatial selection mechanism based on fusion characteristics. Two sets of spatial weight maps are adaptively generated to adjust the response intensity of shallow and deep features respectively. After convolutional mapping, spatial weights corresponding to the two types of input features are generated, and normalization constraints are used to make the two compete in the same position, thereby reflecting the relative contribution of shallow details and deep semantics. In the formula, for The calculated spatial selection weights, for function, This is for element-wise multiplication.

7. The traffic sign detection method based on edge-guided hypergraph and detail preservation according to claim 1, characterized in that, In step S3, the detail-preserving interaction module enhances the cross-layer interaction between shallow detail features and deep semantic features through a bidirectional gating mechanism; This stage modulates shallow detail features using deep semantic features, and then uses the shallow detail features to inversely correct the deep semantic features. Multiplication and fusion are used to promote information flow between the two types of features, resulting in interactive features. Finally, the two enhanced features and the interactive features are aggregated in a three-way process to obtain the final fused feature. ; In the formula, express Activation function and These are spatial gating weights generated based on semantic and detail features, respectively. These are shallow enhancement features modulated from deep semantic features. The deep enhancement feature is generated by reverse modulation of shallow detail features. The bidirectional gating mechanism refers to generating modulation weights for shallow detail features using deep semantic features and generating modulation weights for deep semantic features using shallow detail features, thereby achieving feature selection and enhancement in two directions. The cross-layer interaction refers to the information transmission, mutual modulation, and fusion process between shallow detail features and deep semantic features. The three-way aggregation refers to aggregating shallow enhancement features, deep enhancement features, and the interaction features obtained by multiplying the two to obtain the final fused feature that simultaneously contains local detail information and high-level semantic information.

8. The traffic sign detection method based on edge-guided hypergraph and detail preservation according to claim 1, characterized in that, In step S3, the detail-preserving interaction module's end design detail-preserving optimization stage corrects the interaction features by combining spatial attention and residual enhancement. Enhance the response of regions sensitive to local structure using spatial attention; and apply the enhanced features The enhanced features are then fed into the detail-preserving interaction module; the detail-preserving interaction module first processes the enhanced features... Multiple convolutions are performed to extract high-frequency detail features, enhancing the model's ability to perceive internal structures, including numbers and arrows; simultaneously, the features are... Weights are generated using global average pooling, fully connected layers, and activation functions, then weighted based on detailed features, and finally combined with... Adding them together yields the output features. ;Then, go through Convolution and Convolutional processing enhances the weights, which are then processed by global average pooling, fully connected layers, and activation functions to generate importance weights. Finally, a gating mechanism is used for weighting, and... Adding them together yields the output features. ; In the formula, for Activation function , They represent , convolution, , Indicates the weights of the fully connected layer. Indicates global average pooling. This indicates a gating mechanism.

9. A traffic sign detection system based on edge-guided hypergraph and detail preservation, characterized in that, This system is used to implement the traffic sign detection method based on edge-guided hypergraph and detail preservation as described in any one of claims 1 to 8, the system comprising: The channel separation strategy module, embedded in the backbone network of the RT-DETR or variant framework, is used to divide the feature channels into enhanced feature streams and preserved feature streams, so that the edge-guided hypergraph module and the detail-preserving interaction module can be compatible. The edge-guided hypergraph module is embedded in the enhanced feature flow of the channel separation strategy. It is used to optimize the similarity measurement between nodes by utilizing the edge structure information of traffic signs, and combined with trainable temperature parameters and dynamic threshold control, to enhance the long-range spatial dependence in the input feature map. At the same time, it uses the edge structure information to strengthen the key contours and internal structural features of traffic signs. The detail-preserving interaction module, embedded in the encoder of the RT-DETR or variant framework, is used to enhance the model's ability to perceive and recover the texture and edge details of traffic signs in deep feature maps that have been downsampled due to downsampling, through a multi-stage bidirectional interaction mechanism and by utilizing the detail-preserving interaction module.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the traffic sign detection method based on edge-guided hypergraph and detail preservation as described in any one of claims 1 to 8.

Images