Small target traffic sign identification method, medium and computer equipment

Abstract

The invention relates to a small target traffic sign recognition method, a medium and computer equipment, and the method comprises the steps: constructing an improved recognition model with YOLOv11 as a basic model, and enabling the recognition model to comprise a small target detection layer; constructing a data set, training the improved recognition model, and recognizing the small target traffic sign by using the trained improved recognition model; the computer readable storage medium and the computer equipment are realized based on the method. According to the method, the shallow information of the small target traffic sign is better utilized, more small target traffic sign features are reserved, and the positioning precision of the small target traffic sign is improved; important feature channels are automatically concerned, the capturing capability of the network on key features is enhanced, and semantic information and detail features are fused more efficiently; the capturing capability of the model on the shape features of the feature pattern is improved, the calculation efficiency of feature extraction and fusion is optimized, and the feature utilization efficiency and the information richness are effectively improved; the cross-scale feature interaction capability of the model is effectively enhanced, and the detection capability of a detection head on a small target is improved.

Description

Technical Field

[0001] This invention relates to the technical field of image or video recognition or understanding, and in particular to a method, medium, and computer device for recognizing small target traffic signs. Background Technology

[0002] Object detection, a core task in computer vision, aims to identify objects of interest in images or videos and accurately determine their location and size. Unlike simple image classification tasks, object detection not only determines the category of an object but also solves the problem of object localization. It is considered the foundation for many other vision tasks, such as determining object location in instance segmentation, identifying objects to be labeled in image annotation, and continuously monitoring the position of detected targets in consecutive frames in target tracking. Applications of object detection cover a wide range of important areas, including pedestrian detection (used for pedestrian recognition in intelligent security and autonomous driving scenarios), face detection (widely used in face recognition systems), text detection (playing a role in document analysis and image text extraction), traffic sign and traffic light detection (a key component of autonomous driving and intelligent transportation systems), and remote sensing target detection (used in geographic information analysis, military reconnaissance, and other fields).

[0003] With the rapid development of Intelligent Transportation Systems (ITS) and autonomous driving technologies, Traffic Sign Recognition (TSR) has become a key research area in computer vision. Traffic signs, as a core component of road traffic management, provide drivers and autonomous driving systems with crucial information such as traffic rules, road conditions, and safety warnings. However, in real-world driving environments, traffic sign recognition faces numerous challenges, including:

[0004] (1) Changes in lighting; strong direct sunlight during the day may cause reflections, while dim lighting at night will reduce visibility. These will seriously interfere with the clear imaging of traffic signs.

[0005] (2) Weather conditions; Rain and snow cover, reduced visibility on foggy or hazy days, and dust obscuring traffic signs during sandstorms;

[0006] (3) Obstructions; such as roadside trees, buildings or other vehicles, may partially or completely obscure traffic signs;

[0007] (4) Traffic signs are exposed for a long time; aging, fading, and surface wear will also reduce the recognizability of the signs.

[0008] Traditional traffic sign recognition methods primarily rely on image processing techniques and machine learning algorithms, such as edge detection, color segmentation, and template matching. While these methods can be effective in simple environments, they often struggle to achieve high accuracy and robustness in complex scenarios. Edge detection-based recognition is prone to misjudgment if the edges of the traffic sign are unclear due to lighting or occlusion. In complex color backgrounds, color segmentation algorithms struggle to accurately segment the colors of traffic signs, while template matching algorithms are overly dependent on templates. If the traffic sign is deformed or its angle changes, reducing the number of features, the matching performance suffers significantly.

[0009] In recent years, the rise of deep learning technology has opened up new paths for object detection tasks. Among them, the YOLO (YouOnly Look Once) series of models has been widely used in the field of traffic sign recognition due to its high efficiency, speed, and accuracy. However, the recognition of small target traffic signs remains a major challenge in this field. This is because small target traffic signs occupy a very small proportion of pixels in an image, resulting in a lack of feature information. This makes it difficult for models to accurately extract their features, and small targets are easily affected by background noise, making it even more difficult to effectively distinguish them from the background in complex environments. Summary of the Invention

[0010] This invention addresses the problems existing in the prior art and provides a method, medium, and computer device for recognizing small target traffic signs. It improves YOLOv11 based on feature fusion and attention mechanism, thereby effectively improving the recognition accuracy of small target traffic signs.

[0011] The technical solution adopted in this invention is a method for recognizing small target traffic signs. The method is based on YOLOv11 and constructs an improved recognition model, which includes a small target detection layer.

[0012] A dataset is constructed, and the improved recognition model is trained to identify small target traffic signs.

[0013] Preferably, the improved recognition model is based on YOLOv11;

[0014] The backbone network of the model replaces the C3K2 modules with enhanced C3K2 modules;

[0015] The Neck part of the model includes three sets of upsampling layers, a first fully connected layer, and three sets of convolutional modules and a second fully connected layer, which are sequentially placed after the backbone network. The latter two first fully connected layers are set up in conjunction with the first two enhanced C3K2 modules of the backbone network.

[0016] Each first fully connected layer and second fully connected layer is followed by a DCNv3-GSCSP module;

[0017] The last four DCNv3-GSCSP modules introduce the detection layer output of the hybrid attention HAttention module, which includes a small target detection layer.

[0018] Preferably, the enhanced C3K2 module includes a convolutional block, a segmentation block, and several C3K modules modified with SE arranged sequentially. The outputs of the segmentation block and the last C3K module modified with SE are output after passing through an SDI feature fusion block and a convolutional block.

[0019] Preferably, any of the C3K modules modified with SE includes a sequentially connected convolutional layer and two Bottleneck_SE blocks, the outputs of which are connected and output through the convolutional layer.

[0020] Preferably, the Bottleneck_SE block comprises two convolutional blocks, an SE module, and a fusion block connected in sequence.

[0021] Preferably, the DCNv3-GSCSP module includes a first convolutional layer, after which there are parallel GSConv branches and a second convolutional layer. The outputs of the GSConv branches and the second convolutional layer are added together and then output after being convolved with the input of the DCNv3-GSCSP module into a fully connected layer and a convolutional layer connected in sequence.

[0022] The first convolutional layer is a deformable convolutional layer DCNv3;

[0023] The GSConv branch includes two sequentially connected GSConv modules.

[0024] Preferably, the GSConv module includes sequentially arranged convolutional layers and depth-separable convolutional layers, and the outputs of the convolutional layers and depth-separable convolutional layers are output after passing through a fully connected layer and a channel shuffling layer.

[0025] Preferably, the HAttention module includes a shallow feature extraction unit and a deep feature extraction unit arranged in sequence. The outputs of the shallow feature extraction unit and the deep feature extraction unit are added together and then processed and output by the image reconstruction unit.

[0026] A computer-readable storage medium storing a small target traffic sign recognition program thereon, which, when executed by a processor, implements the small target traffic sign recognition method.

[0027] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the small target traffic sign recognition method.

[0028] This invention relates to a method, medium, and computer device for recognizing small target traffic signs. Based on YOLOv11, an improved recognition model is constructed, including a small target detection layer. A dataset is built, and the improved recognition model is trained to recognize small target traffic signs. The method is implemented using a computer-readable storage medium and a computer device.

[0029] The beneficial effects of this invention are as follows:

[0030] (1) Adding a small target detection layer to the recognition model can better utilize the shallow information of small target traffic signs, retain more features of small target traffic signs, and improve the positioning accuracy of small target traffic signs;

[0031] (2) Introduce feature fusion blocks and SE modules into the backbone network based on YOLOv11, construct a new enhanced C3K2 module, so that the recognition model can automatically focus on important feature channels, enhance the network's ability to capture key features, and more efficiently fuse semantic information and detailed features.

[0032] (3) The C3K2 module of the Neck network is replaced by the DCNv3-GSCSP module optimized by deformable convolution DCNv3. This improves the model's ability to capture the shape features of the feature map, while further optimizing the computational efficiency of feature extraction and fusion, effectively improving the efficiency of feature utilization and information richness.

[0033] (4) The HATtention module is introduced into the detection head to effectively enhance the model’s cross-scale feature interaction capability and improve the detection head’s ability to detect small targets. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method of the present invention;

[0035] Figure 2 This is a schematic diagram of the structure of the improved recognition model of the present invention;

[0036] Figure 3 This is a schematic diagram of the enhanced C3K2 module in this invention;

[0037] Figure 4 This is a schematic diagram of the SDI feature fusion block in this invention;

[0038] Figure 5This is a schematic diagram of the DCNv3-GSCSP module in this invention;

[0039] Figure 6 This is a schematic diagram of the detection head in this invention;

[0040] Figure 7 This is a schematic diagram of the HAttention module in this invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to embodiments, but the scope of protection of the present invention is not limited thereto.

[0042] This invention relates to a method for recognizing small target traffic signs, the method comprising the following steps:

[0043] (1) Based on YOLOv11, an improved recognition model is constructed, wherein the improved recognition model includes a small target detection layer;

[0044] (2) Construct the dataset;

[0045] (3) Train the improved recognition model;

[0046] (4) Use the improved recognition model after training to recognize small target traffic signs.

[0047] The following is a detailed explanation of the steps.

[0048] (1) Based on YOLOv11, an improved recognition model is constructed, wherein the improved recognition model includes a small target detection layer;

[0049] To address the issue of unclear shallow features for small targets like traffic signs in deep feature maps, this invention improves the YOLOv11 model by adding a fourth small target detection layer to the original three detection heads. This allows the original detection heads for large, medium, and small targets to be used for targets larger than 32×32, 16×16, and 8×8 pixels, respectively. By utilizing more shallow features, the newly added small target detection layer can detect targets from 4×4 pixels to 8×8 pixels, thus improving the model's accuracy in detecting small targets.

[0050] The improved recognition model is based on YOLOv11;

[0051] The backbone network of the model replaces the C3K2 modules with enhanced C3K2 modules;

[0052] The Neck part of the model includes three sets of upsampling layers, a first fully connected layer, and three sets of convolutional modules and a second fully connected layer, which are sequentially placed after the backbone network. The latter two first fully connected layers are set up in conjunction with the first two enhanced C3K2 modules of the backbone network.

[0053] Each first fully connected layer and second fully connected layer is followed by a DCNv3-GSCSP module;

[0054] The last four DCNv3-GSCSP modules introduce the detection layer output of the hybrid attention HAttention module, which includes a small target detection layer.

[0055] The following details the areas for improvement.

[0056] (1-1) The enhanced C3K2 module includes a convolution block, a segmentation block, and several C3K modules modified with SE arranged in sequence. The outputs of the segmentation block and the last C3K module modified with SE are output after passing through the SDI feature fusion block and the convolution block.

[0057] Any of the C3K modules modified with SE includes sequentially connected convolutional layers and two Bottleneck_SE blocks, the outputs of which are concatenated and output through the convolutional layer.

[0058] The Bottleneck_SE block consists of two concatenated convolutional blocks, an SE module, and a fusion block (add).

[0059] In this invention, all C3K2 modules in the YOLOv11 backbone network are replaced with enhanced C3K2 modules. The original bottleneck module is modified with an SE module (Squeeze-Excitation). The SE module is added to the end of the Bottleneck module of the C3K module in the C3K2 module (i.e., after the second convolutional layer). By explicitly modeling the dependencies between channels, the network automatically focuses on important feature channels, enhancing the network's ability to capture key features. The Concat module in the C3K2 module of the YOLOv11 network is replaced with an SDI feature fusion block to reduce noise introduction and more efficiently fuse semantic information and detailed features.

[0060] In this invention, the purpose of the SE block compression operation is to embed global spatial information into the channel descriptor, aggregate the spatial features of each channel, and convolve the feature map. Perform global average pooling, and then pool the values ​​of each channel. Spatial dimension is compressed into a scalar to obtain the channel descriptor. ,

[0061]

[0062] in It is a feature map The c-th channel, This is the global statistic for that channel; the SE modifier captures the non-linear dependencies between channels, generating weights for each channel to emphasize important features and suppress irrelevant features; the SE module constructs a gating mechanism using a two-layer fully connected network and a non-linear activation function. First, the channel dimension is reduced using a reduction rate *r* (reducing computation), then restored to its original dimension after ReLU activation, and finally, the sigmoid function is used to output the normalized weights. ,

[0063]

[0064] in It is the ReLU activation function. and It is the weight matrix of the fully connected layer. It is the Sigmoid function;

[0065] The scaling operation of the SE module recalibrates the original feature map at the channel level using the weights generated by the activation operation, and then applies the weights s c Feature map u of the corresponding channel c Multiply to obtain the output features , ,in, X Figure 8 =[ x Figure 9 1 , x Figure 10 2 ,…, x Figure 1 C ] This is the calibrated feature map.

[0066] In this invention, the SDI feature fusion block includes an attention mechanism and channel reduction processing, uniform adjustment of feature map size, feature smoothing processing, and Hadamard product fusion features;

[0067] First, the initial feature map of the i-th layer generated by the encoder is processed. By applying spatial attention and channel attention mechanisms, local spatial information and global channel information are integrated to obtain the processed feature map. The formula is: ,in, and These are the parameters for spatial and channel attention, respectively. Then, a 1×1 convolution is performed... The number of channels is reduced to the hyperparameter c to obtain the feature map. (Dimensions are H) i ×Wi ×c);

[0068] Then, the feature map size is uniformly adjusted to match the i-th layer. For reference, all levels (j level) Adjusted to be with For the same resolution, the formula is:

[0069] =

[0070] Where D, I, and U represent adaptive average pooling (shrinking), identity mapping (invariance), and bilinear interpolation (amplification), respectively.

[0071] Then, feature smoothing is performed on the adjusted feature map. Smoothing is achieved by applying a 3×3 convolution, resulting in... , ,in, These are the parameters for smooth convolution;

[0072] Finally, Hadamard product fusion is performed on all smoothed feature maps. Element-wise multiplication is performed to fuse semantic and detail information from multiple levels, generating an enhanced feature map of the i-th layer. , Where H() represents the Hadamard product, It will be passed to the corresponding level's decoder for further processing.

[0073] (1-2) The DCNv3-GSCSP module includes a first convolutional layer, after which there are parallel GSConv branches and a second convolutional layer. The outputs of the GSConv branches and the second convolutional layer are added together and then the convolution result of the input of the DCNv3-GSCSP module is input to a fully connected layer and a convolutional layer connected in sequence before being output.

[0074] The first convolutional layer is a deformable convolutional layer DCNv3; the second convolutional layer is a regular convolutional layer Conv.

[0075] The GSConv branch includes two sequentially connected GSConv modules.

[0076] The GSConv module includes sequentially arranged convolutional layers and depth-separable convolutional layers. The outputs of the convolutional layers and depth-separable convolutional layers are then output after passing through a fully connected layer and a channel shuffling layer.

[0077] In this invention, the C3K2 module of the Neck part of the YOLOv11 network is replaced by the Deformable Convolution DCNv3-optimized VoV-GSCSP module, namely the DCNv3-GSCSP module. This improves the model's ability to capture the shape features of the feature map while further optimizing the computational efficiency of feature extraction and fusion, effectively enhancing feature utilization efficiency and information richness.

[0078] In this invention, the essence of deformable convolution DCNv3 is to introduce a learnable offset for each sampling point on the basis of traditional convolution kernels, so that the sampling position of the convolution operation is no longer fixed as a regular grid, but can be dynamically adjusted according to the content features of the input image, thereby better conforming to the geometric shape of the target object (such as stretching, rotating, twisting, scaling, etc.).

[0079] Deformable convolution introduces a learnable offset Δp on the basis of traditional convolution. The sampling position changes from a fixed "p" to a dynamic "p+Δp". The offset Δp can be learned by the model and supports fine adjustment at the sub-pixel level.

[0080] In this invention, GSConv is a lightweight convolutional structure. For the input feature map, the convolution operation first generates a feature map with C2 / 2 channels, outputting two branches. One branch uses the DWConv module to perform convolution operations on each channel independently, outputting a feature map with C2 / 2 channels. The outputs of the Conv module and the DWConv module are concatenated. Finally, through a shuffle layer operation, the information generated by the convolutional layer module is permeated into the feature map of DWConv. By adopting a uniform mixing strategy, local feature information is exchanged on different channels, so that the GSConv module can maintain the computational efficiency of the DWConv module while also closely approximating the feature representation capability of Conv, effectively preserving the implicit connections between channels, and having low time complexity.

[0081] In this invention, the DCNv3-GSCSP module feeds the input into two separate branches for processing. One branch consists of the DCNv3 module, the GSbottleneck module, and the Conv convolutional module. The DCNv3 module essentially introduces a learnable offset for each sampling point on top of the traditional convolutional kernel, making the sampling positions of the convolutional operation no longer fixed to a regular grid, thus enhancing the model's shape feature capture capability. The GSbottleneck module consists of two GSConv modules and one Conv module. The two GSConv modules are used to enhance the non-linear representation of features and information reuse, while the middle Conv module is used to further process features and enhance the model's learning ability. The other branch consists of a single Conv module; using only one Conv module is to retain more shallow features of the feature map. The two branches are connected by a Concat module, and the output of the Concat module is finally processed by a single Conv module. This fully explores the feature information of the shallow and deep networks, achieving efficient fusion and effectively improving feature utilization efficiency and information richness.

[0082] (1-3) The HAttention module includes a shallow feature extraction unit and a deep feature extraction unit arranged in sequence. The outputs of the shallow feature extraction unit and the deep feature extraction unit are added together and then processed by the image reconstruction unit for output.

[0083] In this invention, a hybrid attention mechanism (HAttention) is introduced into the YOLOv11 detector head to construct a new detector head (HAHead). The HAttention module is used to replace the second Conv module of the first branch in the original YOLOv11 detector head, so that the structure before the first branch reaches the loss function becomes Conv→HAttention→Conv. Then, the same structure is used to replace the structure before the second branch in the original YOLOv11 detector head reaches the loss function.

[0084] The HAttention module processes the input feature map in three parts: shallow feature extraction, deep feature extraction, and image reconstruction.

[0085] First, shallow feature extraction is performed on the input feature map. Then, deep feature extraction is performed on the feature map after shallow feature extraction. The deep features are extracted by multiple residual hybrid attention groups (RHAGs) in series. Each RHAG contains multiple hybrid attention blocks (HABs), parallel processing channel attention blocks (CABs), spatial window attention blocks (W-MSA / SW-MSA), an overlapping cross attention block (OCAB) for cross-window interaction, and finally residual connections.

[0086] Within each RHAG, the input feature map first undergoes processing through multiple HAB (Hybrid Attention Blocks). Within each HAB block, the input feature X is first subjected to Layer Normalization (LN). Perform sub-attention and channel attention calculations, where window attention (W-MSA) will... The code is divided into multiple non-overlapping windows of size M×M (M=16 by default). Self-attention is calculated within each window, and the output is (S)W-MSA(X). N Channel attention block pair The CAB is applied, which contains convolutions and channel attention blocks, and its output is CAB(X). N Then, through weighting factors The default setting is 0.01, which merges the outputs of W-MSA and CAB and adds residual connections.

[0087]

[0088] in Used to balance the contributions of self-attention and channel attention, avoiding optimization conflicts;

[0089] Finally, After layer normalization, the input is given to the MLP, and the output of the HAB is obtained through residual connection.

[0090] After several HAB processes, the input feature X is processed by the Overlapping Cross-Attention Block (OCAB). In the OCAB, the input feature X is divided into query (X... Q ), key (X) K ), value (X) K ), where X Q Divide into non-overlapping windows (size M×M), X K With X V Create an overlapping window with dimensions M0×M0. , The overlap ratio is typically set to 0.5 by default. Then, within each query window, cross-attention is calculated using the key / value pairs of the overlapping windows. The attention output is processed by layer normalization and MLP, and residual connections are added before being used as the output of OCAB.

[0091] After processing through multiple OCAB layers and one Conv layer, the deep feature F is obtained. D Then, a global residual connection is made to combine the deep features F. D Fusion with shallow feature F0,

[0092]

[0093] The fused features contain both shallow details and deep semantic information. Finally, a pixel shuffling method is used to refine the fused features. Upsampling is performed to obtain the final high-resolution image I. HR .

[0094] HAttention's multi-scale attention fusion can "amplify" the feature weights of small targets in low-resolution feature maps through an attention mechanism, while simultaneously using high-resolution features to supplement details. Cross-window information interaction can associate the local features of small targets with their surrounding context (such as the texture region where the small target is located), assisting the model in determining its semantic category. In summary, introducing hybrid attention HAttention to the YOLOv11 detection head and constructing a new detection head HAHead can effectively enhance the model's cross-scale feature interaction capability and improve the detection head's ability to detect small targets.

[0095] (2) Construct the dataset;

[0096] In this embodiment, the TT100K traffic sign dataset from Tsinghua University is selected as the dataset. The TT100K traffic sign dataset contains more than 10,000 traffic sign sample images, and most of the traffic signs are small, which meets the requirement of small target.

[0097] During data preprocessing, this invention first converts the original labeled dataset into a YOLO format dataset, where each row of data identifies a target, specifically including the target's category number, the target's position coordinates x and y in the image, and the target's width w and height h. All images are then cropped or padded to a fixed size of 640×640 pixels while maintaining the aspect ratio of the images, and unfilled areas are treated with zero padding.

[0098] To ensure the effectiveness of model training, the dataset was divided into three parts: 6,793 images for training, 996 images for validation, and 1,949 images for testing. These parts were used for model training, hyperparameter optimization, performance evaluation, and final detection performance testing.

[0099] (3) Train the improved recognition model;

[0100] In this embodiment, the experimental environment was set on the Ubuntu 22.04.5 LTS system, based on the PyTorch 1.11.0 framework and accelerated using CUDA 11.5. The hardware environment was an NVIDIA Corporation GV100GL [Tesla V100DGXS 32GB] with 32GB of video memory.

[0101] The initial learning rate of the model was set to 0.01, the momentum parameter was set to 0.937, the weight decay coefficient was set to 0.0005, and the batch size was set to 32.

[0102] The preprocessed training set is input into the recognition model network. The classification loss, localization loss, and confidence loss are calculated through forward propagation. Then, backpropagation is performed to update the parameters. The training process continues for 600 epochs. After each epoch, the validation set is input into the current model for inference, and the performance metrics are evaluated in real time. After training, the precision, recall, and mean precision of the validation set are used as the main evaluation metrics for model performance. The model weight with the highest mAP on the validation set is selected as the final detection model.

[0103] (4) Use the improved recognition model after training to recognize small target traffic signs.

[0104] The present invention also relates to a computer-readable storage medium having stored thereon a small target traffic sign recognition program, which, when executed by a processor, implements the small target traffic sign recognition method described above.

[0105] The present invention also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the small target traffic sign recognition method.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes ​ The steps of the function specified in one or more boxes.

[0110] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0111] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for recognizing small target traffic signs, characterized in that: The method uses YOLOv11 as the base model to construct an improved recognition model, which includes a small target detection layer. A dataset is constructed, and the improved recognition model is trained to identify small target traffic signs.

2. The method for recognizing small target traffic signs according to claim 1, characterized in that: The improved recognition model is based on YOLOv11; The backbone network of the model replaces the C3K2 modules with enhanced C3K2 modules; The Neck part of the model includes three sets of upsampling layers, a first fully connected layer, and three sets of convolutional modules and a second fully connected layer, which are sequentially placed after the backbone network. The latter two first fully connected layers are set up in conjunction with the first two enhanced C3K2 modules of the backbone network. Each first fully connected layer and second fully connected layer is followed by a DCNv3-GSCSP module; The last four DCNv3-GSCSP modules introduce the detection layer output of the hybrid attention HAttention module, which includes a small target detection layer.

3. The method for recognizing small target traffic signs according to claim 1, characterized in that: The enhanced C3K2 module includes a convolutional block, a segmentation block, and several C3K modules modified with SE arranged in sequence. The outputs of the segmentation block and the last C3K module modified with SE are output after passing through an SDI feature fusion block and a convolutional block.

4. The method for recognizing small target traffic signs according to claim 3, characterized in that: Any of the C3K modules modified with SE includes sequentially connected convolutional layers and two Bottleneck_SE blocks, the outputs of which are concatenated and output through the convolutional layer.

5. The method for recognizing small target traffic signs according to claim 4, characterized in that: The Bottleneck_SE block comprises two convolutional blocks, an SE module, and a fusion block connected in sequence.

6. The method for recognizing small target traffic signs according to claim 1, characterized in that: The DCNv3-GSCSP module includes a first convolutional layer, followed by a parallel GSConv branch and a second convolutional layer. The outputs of the GSConv branch and the second convolutional layer are added together and then convolved with the input of the DCNv3-GSCSP module before being output after being connected to a fully connected layer and a convolutional layer in sequence. The first convolutional layer is a deformable convolutional layer DCNv3; The GSConv branch includes two sequentially connected GSConv modules.

7. The method for recognizing small target traffic signs according to claim 6, characterized in that: The GSConv module includes sequentially arranged convolutional layers and depth-separable convolutional layers. The outputs of the convolutional layers and depth-separable convolutional layers are then output after passing through a fully connected layer and a channel shuffling layer.

8. The method for recognizing small target traffic signs according to claim 1, characterized in that: The HAttention module includes a shallow feature extraction unit and a deep feature extraction unit arranged in sequence. The outputs of the shallow feature extraction unit and the deep feature extraction unit are added together and then processed by the image reconstruction unit before being output.

9. A computer-readable storage medium, characterized in that: It stores a small target traffic sign recognition program, which, when executed by a processor, implements the small target traffic sign recognition method as described in any one of claims 1 to 8.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the small target traffic sign recognition method according to any one of claims 1 to 8.

Images