A traffic sign detection method based on ADLF-Net

Abstract

The application discloses a traffic sign detection method based on an adaptive double-path local-global fusion network and belongs to the technical field of computer vision and intelligent traffic. Three innovative modules work together: the nonlinear adaptive feature extraction module adopts a wide-narrow parallel structure, combines an adaptive deformation activation function and a partial convolution mechanism, and enhances the capture ability of weak features; the dynamic channel calibration attention module realizes dynamic calibration of key channels through a local-global double-path interaction strategy and double-view attention fusion; and the local-global feature fusion module efficiently fuses multi-scale features based on multi-granularity parallel attention and reparameterization convolution technology. The application constructs an end-to-end detection network on the basis of a YOLO framework, significantly improves the precision and environmental adaptability of traffic sign detection, balances the calculation efficiency and resource consumption through light-weight design, and is suitable for high real-time requirement scenes such as automatic driving and intelligent traffic systems.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly to a traffic sign detection method based on ADLF-Net. BACKGROUND

[0002] With the rapid development of autonomous driving and advanced driver assistance systems, accurate and real-time traffic sign detection technology has become a core requirement in intelligent transportation systems. However, the complexity and variability of real-world traffic scenarios pose great challenges to detection systems. Traffic signs often only occupy a very small area of the image when photographed at a long distance; complex weather and lighting conditions result in reduced contrast and distorted colors of the signs; occlusion and contamination problems are widespread; and there are many visual interference elements in urban environments

[0003] In recent years, deep learning-based target detection algorithms such as the YOLO series have made significant progress in general target detection tasks, but in the face of traffic signs, which are small targets with high precision requirements, existing detection networks still have some technical bottlenecks. Traditional convolution modules generally use fixed form activation functions and standard convolution structures, lack the ability to adaptively capture weak visual features in complex scenes, and result in insufficient detection accuracy for small traffic signs at a long distance or signs in low light conditions. Existing attention mechanisms usually use single-path and fixed-scale channel attention calculations, lack the ability to model local channel interactions and global channel contexts, and are difficult to accurately identify and enhance key feature channels in complex background interference. The simple concatenation or weighted average operation of traditional neck networks ignores the differences in semantic intensity and spatial resolution of different levels of features, fails to effectively model the complex interaction between features, and results in an imbalance between detailed information and contextual understanding, affecting detection accuracy and environmental adaptability.

[0004] In view of the above problems, the academia and industry have proposed various improvement schemes, but each has different degrees of limitations. Lightweight networks such as MobileNet reduce computational complexity through depth separable convolution, but the feature expression capability is also correspondingly reduced, resulting in insufficient capture of subtle features of traffic signs under complex lighting conditions; the Kolmogorov-Arnold network introduces high-order nonlinear expression, but the application form in visual tasks is still imperfect, lacking organic combination of spatial feature mixing. The ECA-Net and CBAM methods attempt to optimize attention calculation, but are still limited to single-granularity feature relationship modeling, lacking collaborative optimization of local channel interaction and global channel dependency, making it difficult to adapt to dynamic characteristic changes of traffic signs in different environments. The FPN, PANet and ASFF methods propose top-down and bottom-up feature fusion strategies, and the fusion operation is still too simple, without considering the semantic differences of features at different abstraction levels, lacking adaptive evaluation mechanism for feature quality. These methods have their own advantages, but in the detection of traffic signs and other special scenes, it is often difficult to balance the detection accuracy, environmental adaptability and computational efficiency. SUMMARY

[0005] (I) Technical problems solved

[0006] In view of the deficiencies of the prior art, the present application proposes an adaptive dual-path local-global fusion network (ADLF-Net) which, through the organic integration of three innovative modules, constructs a high-performance and environmentally adaptable detection framework, effectively solving the key technical problems of sign detection in complex traffic scenes.

[0007] (II) Technical solutions

[0008] To achieve the above object, the present application is implemented by the following technical solutions: a traffic sign detection method based on ADLF-Net, comprising the following steps:

[0009] Step 1: Construct an ADLF-Net network, which successively contains a nonlinear adaptive feature extraction module (NSFE), a dynamic channel calibration attention module (DCCA) and a local-global feature fusion module (LGFF);

[0010] Step 2: input the image to be detected into the ADLF-Net network, and extract enhanced multi-level features through the NSFE module;

[0011] Step 3: use the DCCA module to calibrate the channel of the multi-level features to generate dynamic enhanced channel attention weights;

[0012] Step 4: Multi-scale fusion of calibrated features by LGFF module to generate integrated features that fuse local details and global context;

[0013] Step 5: Output the class and location information of the traffic sign based on the integrated features.

[0014] Preferably, the NSFE module adopts a wide-narrow parallel structure, including:

[0015] Standard convolution path, which preserves the basic features through short circuit connection;

[0016] Deformation activation path, composed of deformation activation unit, partial convolution layer and random path deactivation layer, used to mine the deep nonlinear relationship of input data;

[0017] The deformation activation unit enhances the nonlinear expression ability of the model through adaptive deformation activation function and partial convolution mechanism.

[0018] Preferably, the DCCA module realizes channel calibration through a dual-view attention fusion mechanism, specifically including:

[0019] Local path: adopts one-dimensional convolution branch, whose convolution kernel size is dynamically adjusted as a function of the number of input channels;

[0020] Global path: models the full-channel dependency relationship through a fully connected layer;

[0021] Fusion strategy: fuse local and global attention weights through matrix multiplication and weight sharing strategy to generate dynamic channel attention weights.

[0022] Preferably, the LGFF module realizes multi-scale fusion through a multi-granularity parallel attention mechanism, including:

[0023] Apply attention units of different receptive field sizes to input features to extract local details and global semantic information respectively;

[0024] Adjust the information flow weight of local and global features through a similarity adaptive gating mechanism;

[0025] Use reparameterization convolution technology to enhance the expression ability through multi-branch structure in the training stage, and equivalent conversion to single convolution layer in the inference stage to maintain computational efficiency.

[0026] Preferably, the ADLF-Net network is built based on the YOLO framework, and the network parameters are optimized through end-to-end training to output detection results including: traffic sign bounding box coordinates, class confidence, classification label.

[0027] Preferably, in the shape activation path, the convolution kernel size of the partial convolution layer is dynamically adapted to the spatial dimension of the input feature, and part of the neuron connection is randomly shielded through the random path deactivation layer to enhance the feature robustness.

[0028] Preferably, the similarity adaptive gating mechanism generates a gating coefficient to weight and fuse the local and global features by calculating the correlation between the local and global features.

[0029] Preferably, the multi-level feature is composed of the convolution layer outputs at different depths of the network, and the NSFE module enhances the nonlinear expression ability of the feature layer by layer through the wide-narrow parallel structure.

[0030] (Three) beneficial effects

[0031] The application provides a traffic sign detection method based on ADLF-Net. The application has the following beneficial effects:

[0032] (1) The nonlinear adaptive feature extraction and dynamic channel calibration attention mechanism significantly improve the detection accuracy of traffic signs and outperform existing mainstream methods in complex scenes.

[0033] (2) For low light, occlusion and complex background interference, ADLF-Net uses adaptive shape activation function and local-global feature fusion mechanism to show stronger environmental adaptability, effectively alleviating the performance degradation problem of traditional methods in extreme conditions.

[0034] (3) The network uses lightweight design and reparameterization convolution technology to maintain high inference speed while significantly reducing the number of model parameters, achieving a balance between performance and computing resources to meet real-time detection requirements.

[0035] (4) Through the multi-granularity parallel attention mechanism and adaptive gating strategy, ADLF-Net can efficiently fuse features at different levels, enhance the ability to capture long-distance small targets and detailed features, and improve the overall detection. BRIEF DESCRIPTION OF DRAWINGS

[0036] Fig. 1 is the overall structure diagram of the ADLF-Net traffic sign detection network proposed by the application;

[0037] Fig. 2 is the structure diagram of the nonlinear adaptive feature extraction (NSFE) module of the application;

[0038] Fig. 3 is the structure diagram of the dynamic channel calibration attention (DCCA) module in the application;

[0039] Fig. 4Fig. 1 is a structural schematic diagram of a local-global feature fusion (LGFF) module in the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0041] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application. Fig. 1 to Fig. 4 The present application will be further described in detail.

[0042] Referring to Fig. 1 The present application proposes a traffic sign detection method based on an adaptive dual-path local-global fusion network (ADLF-Net). The method innovates from the perspective of network architecture design to address the deficiencies of existing traffic sign detection algorithms in small target recognition, environmental adaptability and computational efficiency. Three original modules are the core of the end-to-end detection framework: a nonlinear adaptive feature extraction (NSFE) module, a dynamic channel calibration attention (DCCA) module and a local-global feature fusion (LGFF) module. Among them, Adaptive corresponds to the adaptive deformation activation feature of the NSFE module, Dual-path corresponds to the local-global dual-path interaction strategy of the DCCA module, and Local-global Fusion corresponds to the local-global feature fusion mechanism of the LGFF module. The NSFE module improves the feature expression ability of the network by introducing an adaptive deformation activation function and a partial convolution mechanism. The DCCA module adopts an innovative local-global dual-path interaction strategy to achieve efficient dynamic calibration of channel information. The LGFF module solves the key problem of multi-scale feature fusion through multi-view parallel attention. The ADLF-Net detection method can improve the precision and robustness of traffic sign detection while maintaining low computational complexity.

[0043] Referring to Fig. 2, the nonlinear self-adaptive feature extraction (NSFE) module adopts a "wide-narrow parallel + adaptive activation" architecture design, which shows a new idea of feature processing. After the initial processing of the input feature map, the system divides it into two complementary paths: one path maintains the continuity and stability of the basic information flow through standard convolution; the other path contains multiple serial deformation activation units, which aims to mine the deep nonlinear relationship in the input data. This split-integrated architecture enables the module to obtain feature representations at different levels, preserving the integrity of the original visual features while enhancing the model's ability to recognize complex patterns.

[0044] Specifically, the input features are first processed by two standard convolution layers cv1 and cv2. Subsequently, the feature stream is divided into two parallel paths: one path forms a short circuit connection shortcut = x to preserve the basic feature information by adjusting the dimensions through an additional Conv layer according to whether the input and output dimensions are consistent; the other path focuses on the mining of deep nonlinear features, sequentially passing through a Partial_conv3 partial convolution layer to reduce computational complexity, an mlp layer to introduce an adaptive deformation activation function to enhance nonlinear expression ability, and a drop_path layer for random path deactivation regularization. The outputs of the two paths are preliminarily fused by element-wise addition operation. The fused features and the output features of the cv1 layer are concatenated in the channel dimension, integrating feature information at different levels. Finally, the concatenated features are processed by a cv3 convolution layer for final feature extraction and dimension adjustment, generating the output features of the module. The "wide-narrow parallel + adaptive activation" architecture, combined with partial convolution, parameterized activation, and residual connection ideas, effectively improves the model's ability to capture complex visual patterns, especially weak features in traffic signs, while considering computational efficiency.

[0045] Referring to Fig. 3 The DCCA module considers the multi-level dependency relationship of channel information in its design and adopts a "local-global dual-path interaction" processing architecture. Unlike traditional channel attention mechanisms that only focus on a single information granularity, the module constructs two complementary feature processing paths in parallel, focusing on channel correlation at different scales.

[0046] The module first performs global average pooling on the input feature map, compressing the spatial dimension to a one-dimensional channel descriptor x. Subsequently, x is sent to two parallel processing branches: the first path uses a one-dimensional convolution layer, with the key innovation being that the convolution kernel size dynamically changes with the input channel number, allowing adaptive adjustment of the local receptive field to accurately capture local correlation patterns between adjacent channels, generating local features x1; the second path processes x through a fully connected layer, aiming to establish a dependency relationship model across all channels to obtain more macro global semantic context information, generating global features x2.

[0047] In the information fusion process, the DCCA module applies an innovative "dual-view attention fusion" mechanism. Through matrix multiplication and subsequent processing, two complementary attention weight maps are generated according to x1 and x2, respectively, emphasizing features from different perspectives: one perspective focuses on local patterns with global context support, and the other highlights local details in global patterns, together building a comprehensive channel importance evaluation system. After the weight map is fused through the Mix operation, in order to balance the effect and computational efficiency, the weight sharing strategy is adopted to fine-tune the fused attention, and the final channel attention weight out is obtained. This refining step not only reduces the parameter quantity, but also makes the attention calculation more coherent and efficient.

[0048] Finally, the fine-tuned channel attention weight out is applied to the original input feature Input through element-wise multiplication operation, realizing adaptive enhancement and suppression of each channel feature. The output Output feature map is the result of the DCCA module calibration, and the key information channel is effectively focused. The DCCA module combines local-global information interaction and dual-view fusion, enabling the network to more flexibly, accurately and efficiently adapt to different levels of visual information, improving overall performance.

[0049] Referring to Fig. 4 In the target detection system, effective fusion of multi-scale features is crucial for accurate positioning and recognition of different size targets. Traditional feature fusion methods mostly use simple concatenation or weighted average strategy, which often shows limitations when dealing with traffic signs with large scale changes and obvious semantic differences. The Local-Global Feature Fusion (LGFF) module uses a solution that enhances the interaction of features at different levels through multi-granularity attention mechanism and adaptive feature integration technology.

[0050] The LGFF module uses a "three-way parallel + fine integration" approach to process feature inputs from different depth levels of the network. To ensure the semantic compatibility of the features, the module first performs channel dimension remapping on the input features through point convolution operation to create a unified feature representation space. On this basis, the module expands three complementary processing paths in parallel: the basic path extracts and preserves the essential information of the input features through simple addition and convolution operations; the two enhancement paths apply specially designed attention mechanisms to the features, enhancing the expression ability of the features from different perspectives.

[0051] The LGFF uses a "multi-receptive field parallel processing" implementation form. Unlike the conventional attention calculation which only considers a single granularity, the module simultaneously applies attention units of different receptive field sizes to each feature map, forming complementary feature enhancement effects: the attention unit of a smaller receptive field (p = 2) focuses on capturing local details and texture information, which is of great significance for identifying detail features such as pictographic patterns and text on traffic signs; the attention unit of a larger receptive field (p = 4) focuses on extracting broader environmental context and semantic associations, prompting the model to understand the scene background in which the target is located. This multi-granularity parallel processing strategy allows the module to simultaneously consider microscopic details and macroscopic structures, forming a more comprehensive feature representation.

[0052] To improve the discriminative ability of the fused features, the module also studies an adaptive gating mechanism based on similarity, by introducing a trainable parameterized vector to establish a feature correlation evaluation standard, enabling the system to automatically adjust the information flow weight according to the similarity between the features and the preset patterns. This effectively enhances the feature response related to the target while reducing the impact of background interference. In the final feature integration stage, the outputs of the three processing paths are fused and purified through channel dimension splicing and specially designed reparameterization convolution, forming a comprehensive feature representation rich in multi-level information.

[0053] Reparameterization convolution, as the last link of the module, significantly enhances the model's expressive ability in the training phase while maintaining the computational efficiency in the inference phase. It uses a multi-branch structure to enhance learning ability during training, and is equivalent to a single efficient convolution during inference, skillfully balancing performance and efficiency.

[0054] The effectiveness of the present application will be further illustrated in combination with experiments. To comprehensively verify the effectiveness of the method proposed in the present application, we designed and carried out systematic experimental evaluation, including determination of experimental evaluation indicators, implementation of ablation experiments and comparative experiments, and quantitative and qualitative analysis of results. The following is a detailed description of experimental evaluation and results.

[0055] Experimental evaluation indicators

[0056] The following indicators are used by the present application to comprehensively evaluate the performance of traffic sign detection:

[0057] Precision P (%): the ratio of the number of correctly detected samples to all detection results, reflecting the accuracy of the model's prediction.

[0058]

[0059] Among them, TP (True Positive) represents the number of correctly detected samples, and FP (False Positive) represents the number of incorrectly detected samples.

[0060] Recall R(%) : the ratio of correctly detected samples to all samples that should be detected, reflecting the comprehensiveness of the model in capturing targets.

[0061]

[0062] where FN(False Negative) represents the number of missed samples.

[0063] Mean Average Precision mAP50(%) : the average precision at an IoU threshold of 0.5, which is the most commonly used comprehensive performance indicator in target detection tasks, used to evaluate the detection performance of the model on all classes.

[0064]

[0065] where C represents the number of classes, represents the average precision of the i-th class at an IoU threshold of 0.5, calculated as the area under the precision-recall curve:

[0066]

[0067] where P i (r) represents the precision when the recall is r.

[0068] Inference Speed FPS: the number of image frames processed per second, used to evaluate the real-time performance of the algorithm in actual application.

[0069]

[0070] where N is the total number of processed images, T i is the time required to process the i-th image.

[0071] Parameter Amount (M) : the total amount of parameters of the model, in millions (M).

[0072] Experimental content and result analysis

[0073] The experiments of the present application were carried out on public traffic sign datasets, including TT100K and GTSDB. To ensure the reliability of the experiments, we adopted standard training-validation-testing set division, and used data augmentation techniques such as random cropping, rotation, noise, etc. to simulate various real scenarios.

[0074] All experiments were conducted under the same hardware configuration (NVIDIA RTX 3060 GPU, Intel i3 8100), ensuring the fairness of performance comparison. The basic detection framework used YOLOv8, and the training hyperparameters were set as follows: batch size 32, total training rounds 300.

[0075] Table 1. Ablation experiment results of each module of the present application

[0076] Model configuration P(％) R(％) mAP50(%) FPS Parameter(M) Baseline model 87.5 88.9 88.2 65.3 11.2 +NSFE 89.7 91.1 90.4 62.8 11.8 +DCCA 91.0 92.5 91.7 60.6 12.3 +LGFF 92.5 93.8 93.2 58.2 13.0 ADLF-Net 94.1 95.2 94.6 56.5 13.5

[0077] To verify the effectiveness of each core module of the present application, we conducted detailed ablation experiments, and the results are shown in Table 1. The ablation experiment results show that with the gradual introduction of each key module of the present application, the model performance shows a steady upward trend. The introduction of the NSFE module improves the model mAP50 by 2.2 percentage points (88.2% to 90.4%), and the precision and recall are increased by 2.2% and 2.2%, respectively, proving that the adaptive deformation activation function indeed enhances the network's ability to capture complex features. On this basis, the addition of the DCCA module makes the mAP50 reach 91.7%, further verifying the effectiveness of the local-global dual-path interaction strategy in enhancing the response of key channels, especially for the detection of small targets at a distance. The integration of the LGFF module improves the mAP50 to 93.2%, fully embodying the effective fusion ability of the multi-granularity attention mechanism and adaptive feature integration technology for multi-scale features. The final complete method reaches an mAP50 of 94.6%, an increase of 6.4 percentage points compared to the baseline model, and the precision and recall reach 94.1% and 95.2%, respectively. It is worth noting that although the parameter amount increases by 20.5% (11.2M to 13.5M), the detection performance improvement is far greater than the parameter growth rate, and the inference speed still maintains at 56.5FPS, meeting the real-time application requirements, indicating that the present application has achieved an excellent balance between performance and computational efficiency.

[0078] Table 2. Performance comparison of the present application and existing methods

[0079] Method P(％) R(％) mAP50(%) FPS Parameter(M) FasterR-CNN 88.6 90.3 89.5 28.6 41.5 RetinaNet 89.2 91.0 90.1 32.7 36.4 YOLOv5-M 90.5 92.2 91.4 62.3 21.2 YOLOv7 91.8 93.5 92.7 60.1 36.9 YOLOv8-M 92.3 93.9 93.1 63.5 25.9 ADLF-Net 94.1 95.2 94.6 56.5 13.5

[0080] Meanwhile, we compared the performance of the present application with the current mainstream traffic sign detection method, and the results are shown in Table 2. The comparison results show that the present application has obvious advantages in both detection performance and resource utilization. In terms of detection indicators, the present application achieves an mAP50 of 94.6%, while the precision and recall rates reach 94.1% and 95.2%, respectively, all of which are higher than the current most advanced YOLOv8-M method (mAP50 of 93.1%). What is particularly prominent is the significant optimization of the parameter quantity of the present application, which is only 13.5M, about 52.1% of YOLOv8-M (25.9M), which greatly reduces the computational burden while maintaining the leading detection accuracy. From the perspective of inference efficiency, the 56.5FPS of the present application is slightly lower than the 63.5FPS of YOLOv8-M, but it is still far beyond the real-time application requirement (25FPS), and considering the improvement of detection accuracy and the significant reduction of resource occupation, this slight speed difference is reasonable and acceptable. In addition, in the complex scene adaptability test, the performance degradation of the present application under low light and partial occlusion conditions is 8.3% and 9.1%, respectively, which is significantly better than the 12.5% and 13.2% of YOLOv8-M, proving the enhanced environmental robustness brought by the innovative module design of the present application. These results collectively verify the comprehensive technical advantages of the present application in the traffic sign detection task, and it is particularly suitable for deployment in practical application scenarios with limited resources.

[0081] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A traffic sign detection method based on ADLF-Net, characterized in that, The method comprises the following steps: Step 1: constructing an ADLF-Net network, which comprises a nonlinear adaptive feature extraction module NSFE, a dynamic channel calibration attention module DCCA and a local-global feature fusion module LGFF in sequence; Step 2: inputting an image to be detected into the ADLF-Net network, and extracting enhanced multi-level features through the NSFE module; Step 3: calibrating the multi-level features through the DCCA module to generate dynamic enhanced channel attention weights; Step 4: performing multi-scale fusion on the calibrated features through the LGFF module to generate comprehensive features fusing local details and global context; Step 5: outputting class and position information of the traffic sign based on the comprehensive features. The NSFE module adopts a wide-narrow parallel structure and comprises: a standard convolution path for retaining basic features through a short circuit connection; a deformation activation path composed of a deformation activation unit, a partial convolution layer and a random path inactivation layer, for mining deep nonlinear relationships of input data; wherein the deformation activation unit enhances the nonlinear expression ability of the model through an adaptive deformation activation function and a partial convolution mechanism; The DCCA module calibrates channels through a dual-view attention fusion mechanism, specifically comprising: a local path adopting a one-dimensional convolution branch, whose convolution kernel size is dynamically adjusted as a function of the number of input channels; a global path modeling full-channel dependency through a fully connected layer; a fusion strategy for fusing local and global attention weights through matrix multiplication and weight sharing strategy to generate dynamic channel attention weights; The LGFF module realizes multi-scale fusion through a multi-granularity parallel attention mechanism, comprising: applying attention units of different receptive field sizes to input features to extract local details and global semantic information respectively; adjusting the information flow weight of local and global features through a similarity adaptive gating mechanism; using a reparameterization convolution technique to enhance the expression ability through a multi-branch structure in the training stage, and equivalent conversion to a single convolution layer in the inference stage to maintain computational efficiency.

2. The ADLF-Net-based traffic sign detection method according to claim 1, characterized in that: The ADLF-Net network is constructed based on the YOLO framework, and the network parameters are optimized through end-to-end training to output detection results including: the bounding box coordinates of the traffic sign, the class confidence, and the classification label.

3. The ADLF-Net based traffic sign detection method of claim 1, wherein: In the deformation activation path, the convolution kernel size of the partial convolution layer dynamically adapts to the spatial dimension of the input feature, and the random path inactivation layer randomly masks part of the neuron connections to enhance the robustness of the feature.

4. The ADLF-Net-based traffic sign detection method of claim 1, wherein: The similarity adaptive gating mechanism calculates the correlation between local and global features to generate a gating coefficient for weighted fusion of local and global features.

5. The ADLF-Net based traffic sign detection method of claim 1, wherein: The multi-level features are composed of convolution layer outputs at different depths of the network, and the NSFE module enhances the nonlinear expression ability of the features layer by layer through the wide-narrow parallel structure.

Images