Road sign recognition method based on lightweight target recognition network

Abstract

The application discloses a road sign recognition method based on a lightweight target recognition network, constructs an MGA module, improves the model's extraction capability for low-resolution and small-shield sign features, and significantly reduces the false detection and missed detection; an improved feature fusion Neck module is designed, a branch is introduced from an improved feature extraction Backbone module, multi-scale interaction operation of a maximum pooling layer is used, and the features of small-scale targets are well reserved; an improved SPPF module is designed, a designed GLS module is added after each maximum pooling operation of each level of the original SPPF module, and the perception capability for long-distance imaging and blurred signs is effectively enhanced. The application has high robustness and perception accuracy, improves the target detection accuracy and recall rate, strictly controls the parameter quantity through a lightweight convolution structure, and maintains excellent detection real-time performance.

Description

Technical Field

[0001] This invention belongs to the field of image processing and target recognition, specifically relating to a road sign recognition method based on a lightweight target recognition network, which is used to improve the target recognition accuracy and speed of road signs in vehicle driving scenarios. Background Technology

[0002] Road traffic images, captured by in-vehicle cameras or road monitoring equipment, are widely used in autonomous driving, advanced driver assistance systems (ADAS), and intelligent traffic management. Particularly in the field of autonomous driving, road sign recognition is crucial for safe driving and compliant decision-making. Real-world road environments are complex, presenting challenges such as varying lighting conditions, obstruction by trees or vehicles, worn and faded signs, multi-scale characteristics of road targets in the image, and small objects.

[0003] Unlike the single background of remote sensing images, road environments are greatly affected by changes in lighting (such as strong light and nighttime), inclement weather, and the dynamic effects of trees and vehicles obstructing the view. In addition, traffic signs in images often appear as small targets with low pixel ratios and motion blur, which makes existing algorithms prone to problems such as insufficient feature extraction, high false negative rates for distant small targets, and frequent false positives in complex backgrounds.

[0004] In summary, existing target recognition algorithms still have significant limitations in practical applications, making it difficult to achieve high-precision and robust traffic sign recognition in dynamic driving scenarios with varying lighting, complex interference, varying target sizes, and small targets. Summary of the Invention

[0005] This invention provides a road sign recognition method based on a lightweight target recognition network. By designing a multi-path gated aggregation module (MGA), an improved spatial pyramid pooling-fast (SPPF) module, and an improved feature fusion neck module, the method significantly improves the recognition accuracy of the model for occluded, low-resolution, and small-sized road signs. It can also adapt well to multi-scale target recognition while maintaining real-time performance.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention relates to a road sign recognition method based on a lightweight target recognition network, as detailed below:

[0008] Step 1: Collect road sign images, adjust pixel size, label road signs, construct a road sign dataset and divide it into training and validation sets, and preprocess the training set;

[0009] Step 2: Construct the MGA module, improve the SPPF module, and improve the feature fusion Neck module;

[0010] The MGA module includes a receptive field enhancement branch, a color feature extraction branch, a detail feature extraction branch, a color-edge enhancement feature layer, and a feature integration layer. The color-edge enhancement feature layer concatenates the output of the color feature extraction branch with the output of the detail feature extraction branch to generate color-edge enhancement features. The feature integration layer adds the output of the receptive field enhancement branch with the output of the color-edge enhancement feature layer, and then adds them to a dual-branch parallel structure consisting of depthwise convolutional layers and CBS convolutional layers.

[0011] The improved SPPF module embeds a GLS module after each max pooling layer of the conventional SPPF module. Except for the last GLS module which is concatenated with the corresponding max pooling layer, the remaining GLS modules are connected to the skip connections of the corresponding max pooling layer to the Concat layer. The GLS module is composed of grouped convolution, layer normalization and SiLU activation function concatenated.

[0012] The improved feature fusion Neck module is as follows: all C3K2 modules of the conventional feature fusion Neck module are replaced with MGA modules, and a downsampling branch is added at the 5th layer of the feature extraction Backbone module to perform max pooling operation, which is then used as the input to the first Concat layer and the third Concat layer in the improved feature fusion Neck module, respectively.

[0013] Step 3: Construct an improved YOLO11n target recognition model that includes an improved feature extraction Backbone module, an improved feature fusion Neck module, and a recognition Head module; the improved feature extraction Backbone module replaces all C3K2 modules of the conventional feature extraction Backbone module with MGA modules; the latter three MGA modules in the improved feature fusion Neck module output feature maps of road signs at three different size levels, which are then input into the recognition Head module, which outputs the road sign recognition results;

[0014] Step 4: Using the training and validation sets obtained from the data preprocessing in Step 1, train and validate the improved YOLO11n target recognition model to obtain the final improved YOLO11n target recognition model for recognizing road signs.

[0015] Preferably, the preprocessing involves sequentially performing Mosaic data augmentation, random scaling, horizontal flipping, and color dithering.

[0016] Preferably, in the improved feature extraction Backbone module, the second MGA module, the third MGA module, and the C2PSA module output feature maps of road signs at three different size levels, which are used as the three inputs to the improved feature fusion Neck module, and are respectively denoted as "Input 1", "Input 2", and "Input 3".

[0017] More preferably, the processing flow of "Input 1", "Input 2" and "Input 3" in the improved feature fusion Neck module is as follows:

[0018] (1) Split "Input 1" into a branch, pass through a max pooling layer with a kernel size of 2×2, and output the feature map as "pooled Input 1";

[0019] (2) Input "Input Three" to input the first upsampling module;

[0020] (3) The "pooled input one" and "input two" are spliced ​​together with the feature map after upsampling in step (2) in three channels;

[0021] (4) The output of step (3) is processed by the first MGA module for feature aggregation to obtain “neck mid-section features”;

[0022] (5) After the "neck middle feature" is processed by the second upsampling module, it is concatenated with the original "input one" channel;

[0023] (6) The output of step (5) enters the second MGA module, and the aggregated feature map is used as the first input of the Head recognition module;

[0024] (7) The output of step (6) is split into another path and input to a CBS module for downsampling;

[0025] (8) The output of step (7), the "neck middle feature" and the "pooled input 1" are spliced ​​together in three channels;

[0026] (9) The output of step (8) is sent to the third MGA module, and the aggregated feature map is used as the second input of the Head recognition module;

[0027] (10) The output of step (9) is split into another path and input to the next CBS module for downsampling;

[0028] (11) The output of step (10) is concatenated with “Input 3” and then aggregated through the fourth MGA module. The aggregated feature map is used as the third input of the Head recognition module.

[0029] Preferably, a transfer learning strategy is adopted during training, and the SGD optimizer is selected to jointly train all weight parameters of the improved feature extraction Backbone module, the improved feature fusion Neck module, and the recognition Head module of the improved YOLO11n target recognition model throughout the entire process; the training rounds, batch size, number of iterations, initial learning rate, and loss function are set; after each round of training, a validation set is used for verification, and training is terminated when the preset number of training rounds is reached or the loss function converges, and all weight parameters corresponding to the optimal value of the validation set are saved.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] 1. By constructing an MGA module, this invention improves the ability to identify low-resolution, occluded small signs in complex scenes, especially in complex terrain or densely built-up areas, and can more accurately identify road signs, reducing false detections and missed detections.

[0032] 2. This invention proposes an improved feature fusion Neck module. It introduces a branch from the improved feature extraction Backbone module and preserves the features of small-scale targets well through multi-scale interactive operations of the max pooling layer. This introduces high-resolution information into the Neck layer without introducing additional parameters, alleviates the loss of small-scale labels in high-level feature maps, and improves the robustness and accuracy of small target recognition.

[0033] 3. This invention proposes an improved SPPF module. By adding a designed GLS module after each max pooling operation at each level of the original SPPF module, the model's ability to perceive long-distance imaging and blurred labels is effectively enhanced.

[0034] 4. While maintaining high accuracy, this invention preserves real-time recognition performance through a lightweight convolutional structure and optimized network design. Especially in resource-constrained embedded systems and edge computing environments, the model of this invention can respond quickly and adapt to applications with high real-time requirements, such as road sign recognition and intelligent driving assistance scenarios.

[0035] 5. Because this invention is specifically designed and structurally optimized for road sign target recognition scenarios, it strictly controls the number of model parameters and computational complexity, resulting in lightweight and high-efficiency features that facilitate deployment and implementation on diverse hardware platforms. Whether on high-performance computing servers or low-power devices such as vehicle-mounted embedded systems and edge computing, this invention can achieve stable and reliable operation. It also supports flexible expansion and adaptation based on the recognition scenario and task scale, effectively meeting the needs of complex road conditions, multiple types of signs, and real-time road sign recognition. Attached Figure Description

[0036] Figure 1 These are four types of typical road sign images collected by this invention.

[0037] Figure 2 This is a flowchart of the data preprocessing process of the present invention.

[0038] Figure 3 This is a schematic diagram of the MGA module in this invention.

[0039] Figure 4 This is a comparative diagram of the SPPF module and the improved SPPF module of this invention.

[0040] Figure 5 This is the overall architecture diagram of the improved YOLO11n target recognition model of this invention.

[0041] Figure 6 This is the overall architecture diagram of the original YOLO11n model.

[0042] Figure 7 This is a comparison chart of the mAP (memory accuracy) of the five models for different road sign categories in Table 1.

[0043] Figure 8 This is a visual comparison chart of the five models in Table 1 for recognizing different road sign categories. Detailed Implementation

[0044] The present invention will be further described below with reference to embodiments and accompanying drawings.

[0045] The road sign recognition method based on a lightweight target recognition network is as follows:

[0046] Step 1: Construct and preprocess the road sign dataset

[0047] ① Collect images

[0048] A total of 4,969 road sign images were collected from different shooting angles and distances on the Kaggle platform. The road signs include single and multiple small targets, complex backgrounds, backlighting, tilted and occluded signs, and normal road signs. Figure 1 The image showcases four typical road sign images: single and multiple small targets, complex backgrounds, backlighting, tilting, and occlusion.

[0049] ② Adjust the size of the road sign image

[0050] The collected road sign images were captured using FSCapture software and scaled to 800×800 pixels.

[0051] ③ Marking road signs

[0052] LabelMe is used to label the road signs in all the road sign images after step ②.

[0053] ④ A road sign dataset was constructed using the 4969 road sign images processed in step ③, of which 3530 images were used as the training set, 801 images as the validation set, and 638 images as the test set.

[0054] ⑤ Preprocess the training set to overcome the problem of imbalanced samples and enhance the model's generalization ability.

[0055] To enhance the model's generalization ability in complex road environments, four data augmentation operations were performed on the training set, such as... Figure 2 As shown, the four data augmentation methods are Mosaic data augmentation, random scaling, horizontal flipping, and color dithering, in that order. The logic for using these four data augmentation methods is as follows: For the images in the training set, the system performs Mosaic data augmentation; subsequently, random scaling, horizontal flipping, and color dithering are performed sequentially on the generated images.

[0056] a. Mosaic Data Augmentation: Mosaic data augmentation involves segmenting multiple images (usually four) at random proportions and then stitching them together to create a new image. This method greatly enriches the background context of the image, effectively improving the model's ability to identify distant, small-scale road signs in congested traffic scenes, overcoming the problems of low proportion of small target samples and complex background environments.

[0057] b. Random Scaling: Without altering the original physical aspect ratio of the road signs, the image is randomly scaled within a range of ±50% with a gain factor of 0.5. This operation simulates the visual size difference caused by changes in shooting distance as the vehicle moves, significantly enhancing the model's adaptability to targets of different scales.

[0058] c. Horizontal Flip: The image is randomly flipped horizontally with a 50% probability. This operation increases the spatial diversity of lane, road sign positions, and environmental background, enhancing the model's robustness in recognizing different road orientations and viewpoints (Note: Considering the physical gravity laws of real road scenes, vertical flipping is not enabled to prevent the introduction of interfering features).

[0059] d. Color jitter: Randomly adjusts the hue, saturation, and brightness of the image within the HSV color space. This operation effectively simulates the light distribution and color shift under different weather conditions (such as sunny days, rainy days, and fog) and different time periods (such as early morning, noon, dusk, or nighttime headlight illumination), significantly improving the model's adaptability to the complex and variable lighting conditions of real roads.

[0060] Through the above data augmentation operations, a large number of diverse training samples that closely match real road conditions were generated, which effectively improved the model's generalization ability under different driving scenarios and conditions. To a large extent, it overcame the problems of high target false negative rate and low recognition accuracy caused by multi-scale road signs, strong background interference, and unbalanced sample distribution.

[0061] ⑥ Adjust the input image of the road sign dataset after processing with the four data augmentation operations to a size of 640×640 pixels.

[0062] Step 2: Construct the MGA module (Multi-path Gated Aggregated Feature Extraction Module), the improved SPPF module, and the improved feature fusion Neck module;

[0063] Analysis of the characteristics of the road sign dataset revealed that:

[0064] First, road signs are characterized by complex background interference, low resolution, and susceptibility to obstruction.

[0065] like Figure 1 As shown, traffic sign recognition faces many serious challenges in real natural road environments. Figure 1 (a) shows the phenomenon of single and multiple small targets that are common in roads. The distant signs have extremely low resolution and carry very few pixel features. Figure 1 (b) reveals the complex urban street scene environment, where traffic signs are often blended into the background noise of vehicles, pedestrians and shop signs of similar colors, making it extremely difficult to distinguish the boundary between the target and the background. Figure 1 (c) reflects the degradation of target imaging under poor lighting conditions such as backlighting, with severe color distortion and loss of internal texture details in the markings; Figure 1 The middle (d) shows the tilt caused by the external force, and the severe obscuring by the dense branches on both sides of the road, which caused the target to lose its complete geometric outline and semantic features.

[0066] Therefore, it is necessary to construct multi-scale feature, color feature, and detail feature enhancement mechanisms to suppress the interference of complex background noise and improve the ability to extract key features of incomplete, occluded, and low-resolution targets.

[0067] Secondly, road signs exhibit degradation characteristics such as extremely small scale at long distances and motion blur after imaging.

[0068] In the road sign dataset, distant signs with both length and width less than 32 pixels are classified as extremely small targets, those between 32 and 96 pixels as medium-sized targets, and those larger than 96 pixels as large targets. Signs with edge distortion due to high-speed vehicle movement and signs in low-light imaging are categorized as blurred targets (or low-resolution targets). Statistical analysis of datasets collected from actual roads reveals many extremely small distant targets. Due to pixel loss caused by distance, feature distortion due to motion, and low image resolution caused by low light, conventional networks are prone to missed detections. Therefore, multi-scale feature fusion for distant targets and the introduction of receptive field enhancement or anti-blurring feature recovery strategies to address target degradation are essential.

[0069] Step 2 will be explained in detail below:

[0070] ① Construct an MGA module to adapt to complex backgrounds, low resolution, and easy occlusion.

[0071] To address the challenges of extracting effective features from road sign targets due to complex background interference, low resolution, and susceptibility to occlusion, this invention designs an MGA module. By decoupling the feature extraction process, it parallelizes channel semantic interaction with multi-scale local feature capture to enhance key salient features of the target and suppress background noise, thereby improving the model's ability to represent low-resolution and occluded targets. Figure 3 As shown, the MGA module includes:

[0072] a. Receptive field enhancement pathways: such as Figure 3 The topmost structure shown is the receptive field enhancement branch, used to extract road sign information from multiple scales to adapt to the multi-scale changes after road sign imaging. It consists of four cascaded convolutional units. The input features first pass through a 1×1 CBS convolutional layer. This layer uses C / 2 kernels to adjust the number of channels from C to C / 2. This operation achieves lightweight design by reducing the number of channels while fusing features to represent small-sized targets. Next, parallel processing and summation are performed through a dual-branch structure. A 3×3 CBS convolutional layer and a 3×3 DW depthwise convolutional layer enhance spatial and channel features in a lower receptive field, respectively, to perceive medium-sized targets. Then, a 5×5 DW depthwise convolutional layer increases the receptive field to recognize larger-sized road signs while reducing computational cost. Finally, a 1×1 CBS convolutional layer restores the number of feature channels to C while fully fusing the channel information of the aforementioned multi-scale features. This design aims to extract local details and contextual information without significantly increasing computational overhead, thus achieving high-quality feature representation.

[0073] In the dual-branch structure, a 3×3DW depthwise convolutional layer is introduced, primarily to address the challenge of feature extraction from road signs in complex scenes. Because road signs are relatively uniform in color, their features are inherently sparse. When image resolution is low, or when images are blurred or occluded, target features become even more difficult to extract. The depthwise convolutional operation independently enhances the features of each channel based on the training weights. The enhanced features are then added to the 3×3 CBS convolutional channels with the same receptive field through point-to-point addition. This design effectively reduces false negatives for road signs, improves recall, and makes vehicle decisions based on road signs safer.

[0074] The 5×5DW depthwise convolutional layer, besides significantly reducing computational cost, also fully considers the characteristics of road signs within a larger receptive field. When the contrast between signage and background colors is strong and features are sparse, the dependency between channels for feature fusion or complementarity is relatively weak, and the marginal benefit of direct channel fusion is limited. Therefore, by independently enhancing each channel through depthwise convolution operations and then performing feature fusion via a 1×1 CBS convolutional layer, a lightweight design can be achieved while maintaining or even improving feature representation.

[0075] b. Color feature extraction branch: such as Figure 3 The intermediate structure shown is a color feature extraction branch, used to extract the color features of road signs, thereby improving the detection rate of low-resolution, occluded road signs in complex backgrounds. This branch first divides the input features into two sub-feature maps along the channel dimension using the Split channel, and then performs an element-wise product operation on the corresponding channels of the two sub-feature maps. Figure 3 middle This indicates element-wise multiplication.

[0076] This is primarily because the features formed by the colors of road signs have relatively regular shapes, representing large-scale information compared to details such as edges. By performing a dot product operation on the corresponding channel feature points after splitting the features at the same scale by channel, the difference in the magnitude of the feature point values ​​can be enhanced, thereby highlighting the large-scale, relatively uniform color features of road signs. At the same time, inconsistent redundancy or noise is suppressed, thus improving the discriminative power of the feature representation. Compared to details, this large-scale feature can still capture low-resolution and occluded conditions. Therefore, this structure can help the model separate road signs from the background, reducing the risk of missed detections and false detections.

[0077] c. Detail feature extraction branch: such as Figure 3The lower structure shown is a branch for extracting detailed features (such as edges), used to preserve the original information of road sign details, thereby improving the detection rate of road signs in complex backgrounds. Road signs have obvious shape features and relatively uniform color, and edges can effectively reflect this. Therefore, this branch processes the C-channel features of relevant scales through a 1×1 CBS convolutional layer to achieve channel fusion, thereby enhancing edge detail information. At the same time, for the purpose of matching in subsequent operations and lightweighting of the structure, the number of channels is reduced to C / 2.

[0078] d. Color-Edge Enhancement Feature Layer: The output of the color feature extraction branch and the output of the detail feature extraction branch are concatenated (Concat) to generate the color-edge enhancement feature of the C channel.

[0079] e. Feature integration layer: such as Figure 3 As shown, the feature integration layer adopts a two-level fusion structure, aiming to deeply fuse multi-scale features with color and edge spatial information to improve the detection capability of road signs in complex scenes. First, the C channel output of the receptive field enhancement branch is added element-wise with the output of the color-edge enhancement feature layer, achieving preliminary fusion of multi-scale features with color and edge spatial information. Subsequently, the fused features enter a dual-branch parallel structure: one is a 1×1 DW depthwise convolutional layer, and the other is a 1×1 CBS convolutional layer. The outputs of the two are added element-wise to complete the final feature integration.

[0080] In the dual-branch parallel structure, the 1×1 DW depthwise convolutional layer branch independently enhances the features of each channel. The enhanced features are then added to the 1×1 CBS convolutional layer branch with the same receptive field by adding them point-to-point along the channel. This design organically combines multi-scale feature representation with the color and edge spatial information of road signs, achieving full fusion of diverse features while deepening the feature flow, thereby effectively improving recall. This structure is particularly crucial for road sign detection in challenging scenarios such as small size, low resolution, and occlusion—it captures detailed features while preserving necessary contextual information, enabling the model to maintain stable detection performance under complex conditions.

[0081] ② Construct an improved SPPF module to adapt to the degradation characteristics of extremely small scales and motion blur.

[0082] Conventional SPPF modules achieve feature fusion across multiple receptive fields by cascading multiple 5×5 max pooling operations. However, for extremely small targets with blurred edges, continuous max pooling can easily lead to the complete loss of key pixel features. Conventional SPPF modules, such as... Figure 4As shown in (a). To address the degradation characteristics of road sign imaging, which exhibits extremely small scale at long distances and motion blur, leading to difficulties in target localization and the easy loss of fine-grained information in deeper network layers, this invention designs an improved SPPF module. The improved SPPF module is as follows: Figure 4 As shown in (b), the differences and advantages of the improved SPPF module compared to the regular SPPF module are as follows:

[0083] This invention constructs an innovative feature recombination and anti-degradation correction mechanism by embedding a GLS module after each max-pooling layer of a conventional SPPF module. The first two GLS modules are connected to skip connections between the corresponding max-pooling layers and the concat layer, respectively, while the remaining GLS module is concatenated with the corresponding max-pooling layer. The GLS module consists of group convolution (i.e., 1×1 GConv convolution, with g=2 groups), layer normalization, and a SiLU activation function concatenated, as shown below. Figure 4 As shown in (c), in this module, the number of channels remains unchanged after the feature map is processed by grouped convolution. This effectively reduces the number of model parameters and computational complexity while preserving complete feature information. Subsequently, layer normalization and the SiLU activation function are applied sequentially, keeping the number of channels unchanged. This further enhances the nonlinear expressive power of the network and improves training stability while maintaining equal-dimensional feature mapping. This design aims to achieve cross-channel information interaction and semantic mapping through the GLS module, thereby continuously expanding the receptive field while providing real-time compensation for feature loss during pooling. This structure not only retains the efficiency advantage of the SPPF module in multi-scale fusion but also effectively curbs the weakening trend of extremely small target features in deep networks and performs feature repair for edge blurring caused by motion blur, significantly enhancing the robustness of the model under harsh imaging conditions.

[0084] ③ Construct an improved feature fusion Neck module to solve the problem of small-scale target recognition.

[0085] like Figure 5 As shown, the improved feature fusion Neck module, compared to the conventional feature fusion Neck module, replaces all C3K2 modules with MGA modules and adds a downsampling branch. The new downsampling branch is located at the MGA module in layer 5 of the improved feature extraction Backbone module (where all C3K2 modules of the original feature extraction Backbone module are also replaced with MGA modules). It first performs a max-pooling operation with a stride of 2 on the output of this MGA module, transforming the feature map from... Become These are used as inputs to the first and third concat operations in the improved feature fusion Neck module, respectively. The MaxPooling operation effectively preserves local extreme responses and is more sensitive to high-contrast regions and edge information. This structure can provide a downsampling path for the feature fusion Neck module that is more conducive to preserving small-scale structural information without introducing additional computation, forming a multi-path feature representation mechanism, which is particularly suitable for road sign recognition tasks with many small targets.

[0086] Step 3: Construct an improved YOLO11n target recognition model

[0087] The improved YOLO11n target recognition model constructed in this invention is as follows: Figure 5 As shown, it includes an improved feature extraction Backbone module, an improved feature fusion Neck module, and a head recognition module. (Similar to...) Figure 6 Compared to the original YOLO11n model shown, the improved YOLO11n target recognition model of this invention uses a newly designed MGA module to replace the original C3K2 module in the backbone module and neck module; the SPPF module at the end of the backbone module is replaced with an improved SPPF module, and cross-layer max pooling operation is innovatively introduced in the neck module.

[0088] The detailed steps of the improved YOLO11n target recognition model constructed in this invention for processing data are as follows:

[0089] ① The data flow in the improved feature extraction Backbone module is as follows:

[0090] a. Input image after preprocessing in step 1 Figure 5 The improved feature extraction Backbone module is shown. The input image size is 640×640 pixels, and the number of channels is 3 (i.e., RGB three channels).

[0091] b. The input image is sequentially processed through two cascaded CBS modules (containing Conv2d convolution, BatchNorm2d batch normalization, and SiLU activation function). The first CBS module adjusts the feature map to 320×320 pixels, and the number of channels becomes 16; the second CBS module further downsamples the feature map to 160×160 pixels, and the number of channels becomes 32.

[0092] c. Feature map input to the first MGA module. This module adopts an innovative multi-branch parallel architecture, including a receptive field enhancement branch, a color feature extraction branch, and a detail feature extraction branch: The receptive field enhancement branch fuses features through a 1×1 CBS convolutional layer and reduces the number of channels from C to C / 2, then it is cascaded in parallel with a 3×3 CBS convolutional layer and a 3×3 DW depthwise convolutional layer, and then sequentially passes through a 5×5 DW depthwise convolutional layer (DWConv) and a 1×1 CBS convolutional layer, restoring the number of feature channels to C. This operation achieves lightweighting while extracting and enhancing road signage information at multiple scales, hence the name "Receptive Field Enhancement Branch." The color feature extraction branch sequentially uses Split channels and element-wise multiplication to strengthen the expression of uniform color features, obtaining a C / 2 channel color enhancement feature. This is then concatenated with the edge detail features from the detail feature extraction branch (after dimensionality reduction by a 1×1 CBS convolutional layer) to generate a C channel color-edge enhancement feature. Next, an Add operation is used to add the C channel output of the receptive field enhancement branch to the color-edge enhancement feature channel-wise. Finally, a dual-branch parallel structure containing a 1×1 DW depthwise convolutional layer and a 1×1 CBS convolutional layer performs element-wise addition channel-wise to achieve deep interweaving and full fusion of multi-dimensional features. After this module, the feature map size remains unchanged, but the number of channels increases to 64.

[0093] d. The data is downsampled by the third CBS module, resulting in a feature map of 80×80 pixels with 64 channels; it is then input into the second MGA module, where the feature map size remains at 80×80 pixels, but the number of channels doubles to 128. The output here extracts the features of small targets and is denoted as "Input 1".

[0094] e. Input 1 is downsampled by the fourth CBS module, resulting in a feature map of 40×40 pixels with 128 channels. It is then input into the third MGA module, maintaining the same size and number of channels (40×40 pixels, 128 channels). The output here extracts features of a medium-sized target, denoted as Input 2.

[0095] f. Input 2 is downsampled by the fifth CBS module, and the feature map becomes 20×20 pixels with the number of channels doubled to 256; then it is input into the fourth MGA module for deep feature extraction, and the output feature map is still 20×20 pixels with 256 channels.

[0096] g. The improved SPPF module for feature map input enables cross-channel information interaction and semantic mapping, thereby expanding the receptive field. At the same time, it provides real-time compensation for feature loss during pooling, while keeping the spatial size and number of channels of the feature map unchanged. This design can effectively prevent feature degradation.

[0097] h. Feature map input C2PSA module. This module contains a cross-stage local network design and multiple PSABlock modules (including an Attention mechanism module and an FNN feedforward network module) cascaded internally to extract deep global context information. The output feature map of this module is 20×20 pixels in size and has 256 channels. The output here extracts features of large-sized targets and is denoted as "Input Three".

[0098] ② The "Input 1" (80×80, 128 channels), "Input 2" (40×40, 128 channels), and "Input 3" (20×20, 256 channels) extracted from the improved feature extraction Backbone module are jointly input into the improved feature fusion Neck module for multi-scale feature fusion. The data flow in the improved feature fusion Neck module is as follows:

[0099] 2.1 Cross-layer pooling: A branch is created from "Input 1" and passed through a 2×2 max pooling layer. The output feature map size is halved to 40×40 pixels, while the number of channels remains unchanged at 128. This is denoted as "pooled Input 1".

[0100] 2.2 Input 3 is processed by the upsampling module, and the feature map is enlarged to 40×40 pixels, while the number of channels remains unchanged at 256.

[0101] 2.3 Concat the pooled input 1 (128 channels), input 2 (128 channels), and the upsampled feature map (256 channels) from step 2.2. The concatenated feature map is 40×40 pixels in size, and the number of channels is added together to become 512 (i.e., the sum of 128, 128, and 256).

[0102] 2.4 The output of step 2.3 is processed by an MGA module for feature aggregation. The size of the output feature map remains unchanged, but the number of channels is reduced to 128. This output is denoted as the "neck middle feature".

[0103] 2.5 The “neck center feature” is upsampled to 80×80 pixels using the Upsample module, while the number of channels remains unchanged at 128. Then, it is concatenated with the original “input one”. After concatenation, the number of channels becomes 256, while the number of pixels remains unchanged.

[0104] 2.6 The output of step 2.5 enters the MGA module. The aggregated feature map size is 80×80 pixels, and the number of channels is reduced to 64. This output is used as the first input to the Head recognition module (corresponding to small target recognition).

[0105] 2.7 The output of step 2.6 is downsampled by the CBS module, and the feature map becomes 40×40 pixels, while the number of channels remains unchanged (64 channels).

[0106] 2.8 The feature map (64 channels) downsampled in step 2.7, the "neck middle feature" (128 channels), and the "pooled input one" (128 channels) obtained in step 2.1 are concatenated into three channels (Concat). The size of the concatenated feature map is 40×40 pixels, and the number of channels becomes 320.

[0107] 2.9 The output of step 2.8 is fed into the MGA module. After aggregation, the feature map size is 40×40 pixels, and the number of channels is reduced to 128. This output is used as the second input to the Head module (corresponding to medium-sized target recognition).

[0108] 2.10 The output of step 2.9 is downsampled by the CBS module (step size 2), and the feature map becomes 20×20 pixels, while the number of channels remains unchanged (128 channels).

[0109] 2.11 The output of step 2.10 is concatenated with the deep "Input Three" (20×20, 256 channels), resulting in 384 channels. After aggregation by the MGA module, the number of channels is reduced to 256. This output is then used as the third input to the Head recognition module (corresponding to large target recognition).

[0110] ③ The data flow of the Head module is identified as follows:

[0111] The improved feature fusion Neck module outputs three multi-scale feature maps, which are then input into the decoupled recognition Head module. The small target recognition branch of the recognition Head module receives the 80×80 pixel, 64-channel feature map from the first input, the medium target recognition branch receives the 40×40 pixel, 128-channel feature map from the second input, and the large target recognition branch receives the 20×20 pixel, 256-channel feature map from the third input.

[0112] Finally, the Head module calculates and outputs the bounding box coordinates (center point x and y coordinates, width, and height), confidence score, and target category of the target in parallel on the feature maps of these three different resolutions, thus completing the entire target recognition task.

[0113] Step 4: Using the training and validation sets obtained from the data preprocessing in Step 1, train and validate the improved YOLO11n target recognition model to obtain the final improved YOLO11n target recognition model for recognizing road signs. This embodiment also uses a test set to evaluate the performance of the final improved YOLO11n target recognition model.

[0114] During training, a transfer learning strategy was employed, using the SGD optimizer to jointly train all weight parameters of the improved feature extraction Backbone module, the improved feature fusion Neck module, and the recognition Head module of the improved YOLO11n target recognition model throughout the entire process. The training epochs were set to 200, the batch size to 16, and the iteration count to 44200. The initial learning rate was set to 1e. -2 The loss function is CIoU. Mosaic data augmentation is enabled during training. After each training epoch, the data is validated using a validation set. Training stops when the number of training epochs reaches a set value (200 epochs) or the loss function tends to converge, and all weight parameters corresponding to the optimal validation set are saved.

[0115] To better reflect the performance of the improved YOLO11n target recognition model of this invention, the average precision (mAP), frames per second (FPS), precision (P), and recall (R) are selected to evaluate the model's performance during the performance evaluation process.

[0116] The following tests the five models using a test set:

[0117] (1) Original YOLO11n model.

[0118] (2) Model (1) combined with MGA module.

[0119] (3) Model (1) combines the improved feature fusion Neck module (Improved Neck).

[0120] (4) Model (1) Combined with Improved SPPF module.

[0121] (5) The present invention improves the YOLO11n target recognition model.

[0122] The test results are shown in Table 1.

[0123] Table 1

[0124] Model (1) √ 96.24 88.77 95.54 193 Model (2) √ √ 95.16 92.31 96.24 145 Model (3) √ √ 95.12 90.73 96.16 196 Model (4) √ √ 97.01 87.33 95.67 192 Model (5) √ √ √ √ 94.80 93.67 96.57 124

[0125] As shown in Table 1, compared with the original YOLO11n model, model (2) shows the most significant improvement in recall, reaching 92.31%, an increase of 3.54 percentage points compared to the baseline, and mAP@0.5 also increased to 96.24%, indicating that this combination has an outstanding effect in improving the completeness of target detection. Model (3) also achieved a dual improvement in recall and mAP, while model (5) focuses more on improving precision, reaching 97.01%, but the recall rate decreased slightly.

[0126] The improved YOLO11n target recognition model of this invention exhibits a significant optimization trend in overall performance, achieving the best recall rate (93.67%) and mAP@0.5 (96.57%), representing improvements of 4.90 and 1.03 percentage points respectively compared to the baseline. Although the precision rate (94.80%) decreased by 1.44 percentage points compared to the baseline, the improvement in recall rate far outweighed the loss in precision. In terms of inference speed, the improved YOLO11n target recognition model achieved an FPS of 124, which, although lower than the baseline of 193, is still higher than the typical threshold for real-time detection (usually 30 FPS), meeting the requirements for practical deployment.

[0127] Considering the application scenarios of road sign detection, the cost of false negatives far outweighs that of false positives. Missing critical signs (such as stop signs and speed limit signs) can cause vehicles to miss important traffic information, leading to safety accidents; while false positives can be filtered and compensated for to some extent through subsequent decision-making logic. Therefore, prioritizing the improvement of recall rate while maintaining acceptable accuracy has clear engineering rationale and safety orientation. In summary, the improved YOLO11n target recognition model of this invention achieves a better balance between detection completeness and real-time performance, making it more suitable for road sign detection tasks.

[0128] The recognition performance of the above five models for different road sign categories is as follows: Figure 7 As shown in the figure. To more intuitively illustrate the contribution of each improvement, typical image samples were selected for comparison, and the comparison results are as follows. Figure 8 As shown. By Figure 8 As shown in (a)-(e), the recognition performance of the five typical road scenes in this invention is significantly improved under conditions of small targets, occlusion, backlighting, tilting and complex backgrounds.

[0129] exist Figure 8 In the scenario shown in (a), the original YOLO11n model struggled to recognize road markings in small target images, resulting in false detections (identifying a speed limit of 10 as a speed limit of 20). Introducing the MGA module significantly improved the recognition of distant small targets by enhancing multi-scale feature fusion. The improved Neck module mitigated the loss of small-scale marker features. Furthermore, the improved SPPF module reduced background interference and enhanced global feature extraction, making the markings more prominent. Ultimately, the improved YOLO11n target recognition model effectively reduced false detections and achieved more accurate recognition.

[0130] exist Figure 8In the scenario shown in (b), the original YOLO11n model performed poorly in handling dark backgrounds, occlusions, and blurred road signs, resulting in missed detections. The MGA module successfully identified the "Stop" sign, while the improved feature fusion Neck module enhanced the ability to identify small targets by preserving detailed information. The improved SPPF module successfully separated the sign from the background, further improving the model's localization accuracy and robustness in complex environments.

[0131] exist Figure 8 In the scenario shown in (c), all five models accurately identified the road signs.

[0132] exist Figure 8 In the scenario shown in (d), road signs at multiple angles significantly impact model performance, making it difficult for the original YOLO11n model to accurately identify them under such conditions. The MGA module effectively improves the ability to recognize road signs from different angles, while the improved feature fusion Neck module enhances global feature fusion, resulting in more accurate feature representations from multiple angles. The improved SPPF module achieves fine-grained feature extraction, maintaining stable performance even when the sign angles differ significantly.

[0133] exist Figure 8 In the scenario shown in (e), the original YOLO11n model struggles to efficiently identify small-scale road signs in complex backgrounds and is prone to misidentifying background noise as road signs. The MGA module significantly improves the model's ability to extract road signs in complex backgrounds through lightweight convolutional feature enhancement and multi-scale feature fusion. The improved feature fusion Neck module, by introducing a branch from the improved feature extraction Backbone module through multi-scale interaction operations of the max pooling layer, effectively preserves the features of small-scale targets. Furthermore, the improved SPPF module enhances global information extraction and improves the feature representation of road signs in complex backgrounds.

[0134] In summary, the gradual integration of these improved modules significantly enhances the performance of the YOLO11n model in complex road scenarios, especially under conditions of small targets, occlusion, tilt, and complex backgrounds, demonstrating more accurate, stable, and real-time road sign recognition performance.

[0135] In addition, to clearly demonstrate the advantages of the improved YOLO11n target recognition model, comparative experiments were conducted using classic models such as the original YOLO11n, YOLOv10n, YOLOv8n, YOLOv6n, and RT-DETR. The experiment also incorporated models based on two-stage detection, such as Faster R-CNN, as well as typical novel road sign detection methods proposed by other researchers. Among them, Improved YOLOv8n is from P. Tian and J. Xiong, “A lightweight real-time traffic sign detection model based on improved YOLOv8n,” in Proc. 44th Chin.Control Conf. (CCC), Chongqing, China, 2025, pp. 8501–8506, doi: 10.23919 / CCC64809.2025.11178591; Improved YOLOv5s is from Z. Wang, W. Luo, X. Li, and W. Hao, “Traffic sign detection based on improved YOLOv5-S,” in Proc. ChinaAutom. Congr. (CAC), Qingdao, China, 2024, pp. 740–745, doi: 10.1109 / CAC63892.2024.10865409. The experimental results are shown in Table 2.

[0136] Table 2

[0137] Faster RCNN 84.40 87.60 86.35 70.74 22 YOLO11n 96.24 88.77 95.54 83.17 193 YOLOv10n 92.64 88.07 94.59 82.05 221 YOLOv8n 94.77 91.14 95.78 83.13 200 YOLOv6n 94.76 86.61 94.07 81.70 211 YOLOv5n 95.24 91.79 96.18 83.31 202 RT-DETR 93.41 92.28 94.76 81.95 38 Improved YOLOv8n 95.30 90.26 97.01 83.97 69 Improved YOLOv5s 95.21 91.52 96.55 82.83 82 Ours 94.80 93.67 96.57 83.82 124

[0138] First, through comparative analysis of the YOLO series models within the YOLO family, and between the YOLO family and the Transformer-based RT-DETR and the two-stage detection model Faster R-CNN, the rationality of choosing the original YOLO11n model as the baseline model can be verified. As shown in Table 2, among the lightweight YOLO models, the original YOLO11n model generally performs well in terms of precision (96.24%), mAP@0.5 (95.54%), and mAP@0.5-0.95 (83.17%), while maintaining a leading inference speed (193 FPS). For Faster R-CNN, although its performance in some metrics is acceptable, its real-time performance of 22 FPS cannot meet the deployment requirements of traffic monitoring scenarios. RT-DETR is slightly better than the original YOLO11n model (88.77%) in recall (92.28%), but its precision, mAP, and inference speed are all inferior to the latter. Taking into account the balance between detection performance and real-time performance, the original YOLO11n model has clear rationality and engineering applicability as a baseline model.

[0139] Secondly, the improved YOLO11n target recognition model of this invention significantly outperforms ImprovedYOLOv8n (90.26%) and Improved YOLOv5s (91.52%) in recall (93.67%), exceeding them by 3.41 and 2.15 percentage points respectively. In road sign detection scenarios, improved recall means a lower risk of missed detections, which is crucial for ensuring driving safety. Although Improved YOLOv8n slightly outperforms the improved YOLO11n target recognition model of this invention (96.57%) in mAP@0.5 (97.01%), its recall and inference speed (69 FPS) are both inferior. While Improved YOLOv5s has a 0.41% higher precision, its recall and speed are also inferior to that of this invention. Considering recall, real-time performance, and overall detection performance, the model of this invention has greater practicality and robustness in complex traffic scenarios.

[0140] This invention addresses the problem of blurred road sign features caused by factors such as drastic changes in lighting, inclement weather, and occlusion / dirt, as well as the high false negative rate and low classification accuracy caused by factors such as the wide variety of sign types, minimal scale differences (e.g., small targets at long distances), and deformation. It proposes effective improvement solutions. Extensive experimental results show that the method of this invention significantly improves recall and real-time performance while maintaining high accuracy, exhibiting good robustness in complex and dynamic traffic scenarios. This invention not only positively promotes autonomous driving environmental perception and high-precision map construction technologies but also provides practical technical support for industrial applications such as intelligent transportation systems, assisted driving, urban asset management, automated road maintenance, and traffic violation evidence collection.

Claims

1. A road sign recognition method based on a lightweight target recognition network, characterized in that: Specifically as follows: Step 1: Collect road sign images, adjust pixel size, label road signs, construct a road sign dataset and divide it into training and validation sets, and preprocess the training set; Step 2: Construct the MGA module, improve the SPPF module, and improve the feature fusion Neck module; The MGA module includes a receptive field enhancement branch, a color feature extraction branch, a detail feature extraction branch, a color-edge enhancement feature layer, and a feature integration layer; the color-edge enhancement feature layer concatenates the output of the color feature extraction branch with the output of the detail feature extraction branch to generate color-edge enhancement features; The feature integration layer adds the output of the receptive field enhancement branch to the output of the color-edge enhancement feature layer, and then adds them to a dual-branch parallel structure consisting of depthwise convolutional layers and CBS convolutional layers. The improved SPPF module embeds a GLS module after each max pooling layer of the conventional SPPF module. Except for the last GLS module which is concatenated with the corresponding max pooling layer, the remaining GLS modules are connected to the skip connections of the corresponding max pooling layer to the Concat layer. The GLS module is composed of grouped convolution, layer normalization and SiLU activation function concatenated. The improved feature fusion Neck module is as follows: all C3K2 modules of the conventional feature fusion Neck module are replaced with MGA modules, and a downsampling branch is added at the 5th layer of the feature extraction Backbone module to perform max pooling operation, which is then used as the input to the first Concat layer and the third Concat layer in the improved feature fusion Neck module, respectively. Step 3: Construct an improved YOLO11n target recognition model that includes an improved feature extraction Backbone module, an improved feature fusion Neck module, and a recognition Head module; the improved feature extraction Backbone module replaces all C3K2 modules of the conventional feature extraction Backbone module with MGA modules; the latter three MGA modules in the improved feature fusion Neck module output feature maps of road signs at three different size levels, which are then input into the recognition Head module, which outputs the road sign recognition results; Step 4: Using the training and validation sets obtained from the data preprocessing in Step 1, train and validate the improved YOLO11n target recognition model to obtain the final improved YOLO11n target recognition model for recognizing road signs.

2. The road sign recognition method based on a lightweight target recognition network according to claim 1, characterized in that: The preprocessing involves sequentially performing Mosaic data augmentation, random scaling, horizontal flipping, and color dithering.

3. The road sign recognition method based on a lightweight target recognition network according to claim 1, characterized in that: In the improved feature extraction Backbone module, the second MGA module, the third MGA module, and the C2PSA module output feature maps of road signs at three different size levels, which serve as the three inputs to the improved feature fusion Neck module, denoted as "Input 1", "Input 2", and "Input 3" respectively.

4. The road sign recognition method based on a lightweight target recognition network according to claim 3, characterized in that: The processing flow of "Input 1", "Input 2", and "Input 3" in the improved feature fusion Neck module is as follows: (1) Split "Input 1" into a branch, pass through a max pooling layer with a kernel size of 2×2, and output the feature map as "pooled Input 1"; (2) Input "Input Three" to input the first upsampling module; (3) The "pooled input one", "input two" and the feature map after upsampling in step (2) are spliced ​​together in three channels; (4) The output of step (3) is processed by the first MGA module for feature aggregation to obtain the "neck middle feature"; (5) After the "neck middle feature" is processed by the second upsampling module, it is concatenated with the original "input one" channel; (6) The output of step (5) enters the second MGA module, and the aggregated feature map is used as the first input of the Head recognition module; (7) The output of step (6) is split into another path and input to a CBS module for downsampling; (8) The output of step (7), the "neck intermediate features" and the "pooled input 1" are spliced ​​together in three channels; (9) The output of step (8) is sent to the third MGA module, and the aggregated feature map is used as the second input of the Head recognition module; (10) The output of step (9) is split into another path and input to the next CBS module for downsampling; (11) The output of step (10) is concatenated with "input three", and then aggregated through the fourth MGA module. The aggregated feature map is used as the third input of the recognition Head module.

5. The road sign recognition method based on a lightweight target recognition network according to claim 1, characterized in that: During training, a transfer learning strategy was adopted, and the SGD optimizer was selected. All weight parameters of the improved feature extraction Backbone module, the improved feature fusion Neck module, and the recognition Head module of the improved YOLO11n target recognition model were jointly trained throughout the entire process. The training rounds, batch size, number of iterations, initial learning rate, and loss function were set. After each training round, the model was validated using a validation set. Training was terminated when the preset number of training rounds was reached or the loss function converged, and all weight parameters corresponding to the optimal validation set were saved.

Images