Speed limit sign detection and identification method based on YOLO and PPOCR models

Abstract

The invention discloses a speed limit sign detection and identification method based on YOLO and PPOCR. The method comprises the following steps: capturing a traffic scene image; adjusting the size of the original image, and scaling the image to a set resolution by adopting a nearest neighbor interpolation algorithm; the adjusted image with the set resolution is input into an improved YOLO model for speed limit sign detection, and detection frame position information of the speed limit sign is obtained; cutting the original image according to the detection frame information; carrying out image preprocessing on the cut image; adjusting the size of the preprocessed image, and scaling the image to a set resolution by adopting a nearest neighbor interpolation algorithm; and carrying out digital identification on the image with the adjusted size by using the improved PPOCR model. According to the method, the precision and adaptability of speed limit sign detection and recognition are remarkably improved, efficient and accurate detection and recognition of the speed limit sign and the speed limit value of the speed limit sign are achieved, and the method has wide application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent transportation systems, and specifically provides a speed limit sign detection and recognition method based on the YOLO and PPOCR models. Background Art

[0002] Driven by intelligent transportation systems and artificial intelligence (AI), traffic management and vehicle safety technologies are undergoing profound changes. Autonomous driving, as a core driver, places higher demands on vehicles' environmental perception capabilities. As a key element of road safety management, the accurate detection and recognition of speed limit signs is crucial for driving safety, compliance with traffic regulations, and intelligent road property inspections. While the deep learning-based two-stage processing paradigm (a separate detection-recognition architecture) is widely used in traffic sign recognition, its inherent flaws severely limit its practical effectiveness in complex road scenarios.

[0003] Existing technical solutions typically use object detection algorithms such as YOLO to initially locate speed limit signs, and then use optical character recognition models such as PPOCR to analyze the sign text information. Although such methods can meet basic functional requirements in ideal scenarios, they expose the following key technical bottlenecks in engineering deployment:

[0004] (1) Feature space mismatch: The detection phase prioritizes visual saliency as the optimization target, ignoring the character structure and semantic relevance required for OCR recognition. This feature fragmentation makes it difficult for the cropped area of ​​the detection box to meet the OCR model's requirements for character integrity and clarity, resulting in an increased misjudgment rate in the subsequent recognition phase.

[0005] (2) Error coupling propagation: The serial processing mechanism of the two-stage model causes the positioning deviation of the detection box to produce a cascade amplification effect in the recognition stage. Even if the detection box is slightly offset or deformed, it will significantly reduce the OCR model's ability to parse the character sequence, resulting in systematic error accumulation.

[0006] (3) Insufficient scene robustness: Existing methods have limitations in adaptability to complex road environments, specifically: ① Light sensitivity causes detection confidence to fluctuate dramatically under extreme lighting conditions; ② Partially occluded scenes easily lead to missegmentation of character sequences; ③ There is a significant problem of missed detection in the detection of small targets at long distances; ④ The recognition and generalization capabilities of multi-language and special-shaped signs are insufficient.

[0007] It's worth noting that current improvements are largely limited to optimizing single modules, such as enhancing the detection model's feature extraction capabilities or improving the character decoding strategy of OCR. While these approaches can improve module performance locally, they fail to establish a deep synergy between detection and recognition tasks, leading to inherent issues such as feature fragmentation and error propagation in real-world scenarios. Therefore, there is an urgent need to develop a joint optimization framework for integrated detection and recognition, leveraging multimodal feature fusion and task-driven learning to achieve breakthroughs in system-level performance. Summary of the Invention

[0008] This invention innovatively achieves deep coupling of the two modules through joint feature optimization of detection and recognition, dynamic adjustment guided by physical priors, and design optimization loss function, breaking through the limitations of traditional serial architecture.

[0009] The technical solution to implement the present invention is: a speed limit sign detection and recognition method based on YOLO and PPOCR models, comprising:

[0010] Capturing a traffic scene image as an original image;

[0011] Resize the original image and use the nearest neighbor interpolation algorithm to scale the image to the set resolution;

[0012] The adjusted image with the set resolution size is input into the improved YOLO model for speed limit sign detection to obtain the detection frame position information of the speed limit sign;

[0013] Crop the original image based on the detection box information output by the improved YOLO model;

[0014] Perform image preprocessing on the cropped image;

[0015] Resize the preprocessed image and use the nearest neighbor interpolation algorithm to scale the image to the set resolution;

[0016] The improved PPOCR model is used to perform digit recognition on the resized images.

[0017] Compared with the prior art, the present invention has the following significant advantages:

[0018] The present invention effectively improves the automatic detection and recognition capability of autonomous driving vehicles for road speed limit signs during driving, thereby enhancing the safety and reliability of the autonomous driving system.

[0019] By combining the YOLO and PPOCR models, this paper achieves rapid detection and accurate recognition of speed limit signs, improving recognition speed and accuracy.

[0020] The image preprocessing and resizing techniques of the present invention optimize image quality and enhance the model's ability to recognize speed limit signs in complex environments.

[0021] The present invention introduces a logical judgment mechanism to ensure the accuracy and rationality of the recognition results.

[0022] The present invention is particularly suitable for intelligent road and road property inspections. It can identify speed limits on road sections in real time while vehicles are driving, provide accurate speed limit information for road inspection vehicles or systems, help promptly discover and correct speed limit signs that do not comply with regulations, and ensure the driving safety of road users.

[0023] In addition, the method of the present invention is highly adaptable and flexible, and can adapt to different traffic environments and lighting conditions. It is expected to be widely used in the fields of intelligent transportation systems, autonomous driving vehicle research and development, and has important application value in intelligent road and property inspections.

[0024] Other features of the present invention will be described in the following description, and part of them will become obvious from the description. The purpose of the present invention can be realized and obtained by the structure particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0026] Figure 1 It is a workflow diagram of the present invention.

[0027] Figure 2 This is the improved YOLO network structure diagram.

[0028] Figure 3 It is the improved PPOCR network structure diagram.

[0029] Figure 4 It is a schematic diagram of the internal vehicle installation position of the device of the present invention.

[0030] Figure 5 It is a schematic diagram of the vehicle-mounted external working mechanism of the device of the present invention.

[0031] Figure 6 It is a detection effect diagram of the present invention. DETAILED DESCRIPTION

[0032] A speed limit sign detection and recognition method based on YOLO and PPOCR, specifically including:

[0033] Step 1: Use a 1080P resolution RGB camera to capture road scene images with traffic speed limit signs. Receive 1920×1080 pixel raw RGB image input in the form of a 3D tensor (height × width × channels). This provides the necessary visual data for object detection in the YOLO model and text recognition in the PPOCR model.

[0034] Step 2: For the original image I src Resize to fit the input size requirement of the YOLO model. Use the nearest neighbor interpolation algorithm to scale the input image to a resolution of 640×640 pixels. This operation is done by the coordinate mapping formula Implementation, outputs a normalized tensor of 640×640×3.

[0035] Step 3: Input the adjusted 640×640×3 normalized tensor into the improved YOLO model for speed limit sign detection to obtain the detection box position information of the speed limit sign.

[0036] The improved YOLO network structure is attached Figure 2 The improved YOLO model mainly consists of three parts: backbone network, neck network and detection head, specifically:

[0037] After the input tensor passes through two basic convolutional layers for spatial downsampling and channel expansion, it realizes cross-layer feature reuse through the C2f module, passes through another convolutional layer, and then branches out into two paths after passing through the second C2f module - branch one outputs the first feature A1 directly into the neck network, and after splicing with the shallow features, it is compressed by the C2f module for channel redundancy, and then the bidirectional context attention module (BiCAM) is used to perform spatial-channel bidirectional weight calibration on the text directional features, and finally passed to the detection head Detect1; branch two outputs the second feature A2, which continues to be deepened in the backbone network, and passes through the convolution layer, C2f module (here the sub-branch is separated to feed back the cross-stage features to the neck, and the output is A3), secondary convolution and C2f module, and then the SPPF pyramid pooling module fuses the multi-scale context information and outputs it to the neck, forming a two-way fusion mechanism in the neck: the first output A4 combines the deep features output at the end of the trunk with the output of branch two The feature concatenation of the sub-branch output A5 of the second feature A2 is refined by the C2f module, and the long-distance dependency is enhanced by the BiCAM module to directly generate the detection feature of the large-scale target and input it into the detection head; the second output A5 reconstructs the resolution through two progressive upsampling operations - the first 2x upsampling is then spliced ​​with the middle-level feature A3 of the third stage of the trunk, and the cross-layer semantic association is extracted by the C2f module. After the second upsampling, it is spliced ​​with the high-resolution shallow feature of the second stage of the trunk. When the spatial details are fused through the third C2f module in the Neck network, a part of the feature output A7 is directly enhanced by the BiCAM module and enters the detection head Detect2. The same feature output then passes through a convolutional layer output A8 to interact with the main path feature across resolutions. The C2f+BiCAM combination generates a multi-scale feature map that takes into account positioning accuracy, and inputs the detection head Detect3 together with the first feature output A4. The YOLO model network is shown in the attached figure. Figure 2 shown.

[0038] The BiCAM module is a feature enhancement module designed specifically for traffic sign character detection. Its core consists of three parallel pathways: a residual backbone, channel attention, and spatial bidirectional scanning. The modules maintain the same dimensionality of the input and output tensors during processing. The residual pathway extracts basic contour features through a 1×1→3×3→1×1 convolution chain, ensuring that the original shape information is not lost. The channel pathway first performs global pooling on the input features to generate channel weight vectors, which, like a "smart filter," enhances the text-related channels and suppresses background interference. The spatial pathway scans the image using horizontal (1×5) and vertical (5×1) strip convolution kernels, capturing the regularity of character rows and columns like a "text tracker." Finally, the three feature paths are superimposed pixel by pixel and output. This module significantly improves the detection accuracy of tilted, blurred, or occluded traffic sign characters through channel enhancement to focus on character color contrast, bidirectional convolution to understand arrangement direction, and residual structure to maintain feature stability.

[0039] Step 4: Crop the original image based on the detection box information output by the improved YOLO model, specifically:

[0040] The conversion formula for converting the center coordinates of the detection box to the coordinates of the upper left corner and lower right corner is:

[0041]

[0042] In the formula, (x c ,y c ) is the center coordinate of the detection box, (x1, y1) and (x2, y2) are the coordinates of the upper left corner and lower right corner of the cropping box respectively, and w and h are the width and height of the detection box.

[0043] Step 5: Perform image preprocessing on the cropped image, including denoising, illumination equalization, morphological opening, binarization, and morphological closing to improve image quality. The specific steps are as follows:

[0044] (a) Convolve the cropped image with a Gaussian filter to obtain a denoised image. The Gaussian filter used is:

[0045]

[0046] Where,

[0047] (x,y) represents the coordinate offset of the pixel in the filter kernel

[0048] σ is the standard deviation of the Gaussian kernel, which controls the smoothness of the filter

[0049] The calculation formula for the standard deviation σ is:

[0050] σ=0.5+0.1·SNR(I raw )

[0051] Among them, SNR(I raw ) is the cropped image I raw The signal-to-noise ratio estimate.

[0052] (b) Perform illumination equalization on the denoised image

[0053] First, the denoised image is evenly divided into multiple local sub-blocks (e.g., 8×8 pixels). The grayscale histogram of each sub-block is calculated and a contrast limit threshold is set. Pixels exceeding the threshold are cropped to suppress noise amplification. The cropped histogram is then adaptively equalized to generate a local grayscale mapping function to expand the dynamic range and enhance the contrast. Based on the bilinear interpolation algorithm, the mapping results of the surrounding sub-blocks are fused according to the distance weight between the current pixel and the center of the adjacent sub-block to eliminate the block boundary traces. Finally, the enhanced image with uniform global illumination distribution, clear local details and no block artifacts is output.

[0054] (c) Perform morphological opening operation on the image after illumination equalization

[0055] The processing flow of the morphological opening operation is divided into two steps: erosion and dilation. First, a structural element of a specific shape and size is defined as a scanning template, and an erosion operation is performed on the input image. That is, the template is moved through the image, and only the positions where the template completely covers the area are target pixels are retained, thereby eliminating small noise and isolated points. Then, a dilation operation is performed on the eroded image, and the same template is moved through the image again to restore the positions where there are any target pixels in the area covered by the template to their original size. Finally, an image with small interference removed and the main shape kept smooth is output.

[0056] (d) Binarize the image after morphological opening operation

[0057] A grayscale threshold is set for the image after the morphological opening operation. The areas above the threshold are set to white and the areas below the threshold are set to black through pixel-by-pixel comparison. If the image illumination is uneven, the block adaptive threshold method is used to divide the image into local areas and calculate the threshold independently to ensure accurate contrast separation of each area. After binarization, residual noise can be filtered out or holes can be filled through secondary morphological operations, and finally a binary image with clear black and white and clear target contours is output.

[0058] (e) Perform morphological closing operation on the binarized image

[0059] The closing operation is the dual operation of the opening operation. The processing flow is divided into two steps: dilation and erosion. First, the structural element is defined, and the dilation operation is performed on the binarized image. That is, the template is moved through the image, and the positions where there are any target pixels in the template-covered area are expanded to the original size, filling small holes and connecting adjacent objects. Then, the erosion operation is performed on the dilated image, and the same template is moved through the image again, retaining only the positions where the template-completely covered area is the target pixel, restoring the main shape while maintaining the filling effect, and finally outputting an image with closed holes, smooth boundaries and coherent targets.

[0060] Step 6: Resize the image after morphological closing to fit the input size requirement of the PPOCR model. Use the nearest neighbor interpolation algorithm to scale the input image to a resolution of 320×48 pixels.

[0061] Step 7: Call the improved PPOCR model to perform digit recognition on the resized image. This model is based on the PaddlePaddle deep learning platform. It extracts image features through CNN and then uses RNN or attention mechanism for sequence modeling to identify text information in the image. The following is the reasoning process of the improved PPOCR model:

[0062] The specific process of the PPOCR model for digital recognition of resized images is as follows:

[0063] First, the resized image is passed through a lightweight MobileNetV3 backbone network to extract multi-scale features;

[0064] The multi-scale features extracted by the lightweight MobileNetV3 backbone network are fused by the dynamic multi-scale feature fusion module. The specific process is as follows:

[0065] The Dynamic Multi-Scale Feature Fusion Module (DMSF) achieves dynamic fusion of multi-scale features that are adaptive to image content by embedding a lightweight SE attention variant behind each layer of features and combining it with a cross-layer feature interaction mechanism. Its working mechanism first completes information complementation between layers through bidirectional cross-scale feature alignment: deep high-semantic features (such as 1 / 16 scale) are upsampled to shallow resolution (such as 1 / 4 scale) through transposed convolution and channel-wise spliced ​​with shallow high-resolution features to form a fusion base with enhanced details; at the same time, shallow features are downsampled and spliced ​​with deep features, using deep global semantics to enhance the contextual relevance of shallow features and avoid the boundary blur caused by traditional bilinear interpolation. On this basis, DMSF generates dynamic weights in real time through an improved SE attention mechanism. It performs global average pooling (GAP) on the concatenated fusion features to compress global information. After generating channel statistical vectors, it uses a lightweight network composed of two fully connected layers (FC) (including ReLU activation and Sigmoid function) to generate channel attention weights. The channel compression ratio (such as 16:1) significantly reduces the number of parameters while retaining the ability to model nonlinear relationships between channels. The generated weight matrix is ​​then multiplied by the original feature map channel by channel to dynamically adjust the fusion ratio of features at different levels: for high-resolution layers (1 / 4 scale), the weights focus on enhancing edge texture details and improving the stroke recognition ability of small-scale text; while for high-semantic layers (1 / 16 scale), the weights enhance the overall semantic consistency of characters and suppress complex background noise. Ultimately, after dynamic weighted superposition, cross-layer features achieve complementary optimization of details and semantics. At the same time, through lightweight designs such as depth-wise separable convolution, weight sharing, and redundant channel pruning, the computational complexity is compressed to 25% of traditional FPN, and the recall rate of small objects in text detection tasks is increased by more than 15%, taking into account the advantages of both efficiency and accuracy.

[0066] The fused feature maps then enter the attention-guided context modeling module (CAA). The spatial attention branch of the context modeling module uses deformable convolution to adaptively capture the geometric structure of distorted text, while the channel attention branch suppresses background texture noise through channel weight recalibration, jointly enhancing the feature expression of the text area.

[0067] After the feature sequence processed by the context modeling module is modeled by BiLSTM context dependency, the CTC decoder outputs the recognition result in the form of digital character text. The improved PPOCR network structure is shown in the attached figure. Figure 3 shown.

[0068] Perform logical judgment on the digital text recognized by the improved PPOCR model. First, verify whether the text is a pure number. Then, confirm the compliance of the number by judging its divisibility by 5, and output the speed limit value or "speed limit value not recognized".

[0069] During the YOLO model's detection process, this paper uses a specific dataset and hyperparameters for targeted training, improving the model's generalization and prediction accuracy for speed limit signs under different conditions. During the PPOCR model's recognition process, a training set specifically for black bold numbers and letters is used to improve the accuracy of recognizing numbers and letters commonly seen on speed limit signs.

[0070] Example 1: System construction and configuration

[0071] S1.1: Select and install a 1080P resolution RGB camera with high dynamic range (HDR) and good night vision capabilities to adapt to different lighting conditions.

[0072] S1.2: Configure the connection between the camera and the backend server or edge computing device to ensure real-time transmission of image data and efficient data processing capabilities.

[0073] Example 2: Image Acquisition and Resizing

[0074] S2.1: Start the camera to collect real-time images and obtain the original images of the traffic scene, ensuring that the image clarity and quality meet the requirements of subsequent processing.

[0075] S2.2: Resize the original image to 640×640×3 to meet the input requirements of the YOLO model. During this process, the nearest neighbor interpolation algorithm is used to minimize image quality loss and ensure that the image maintains high accuracy after scaling.

[0076] Example 3: Preparation and training of YOLO model

[0077] S3.1: Configure the YOLO training environment. Install Python 3.8 or later, NVIDIA CUDA Toolkit 11.8, and cuDNN 8.9.1. Select the PyTorch framework as the running environment for the YOLO model. The PyTorch version should be no less than 1.8.0.

[0078] S3.2: Build a rich and diverse dataset of speed limit signs and use image annotation software such as LabelImg to accurately annotate them.

[0079] S3.3: Download the pre-trained model and source code from the YOLO official website, and design an improved YOLO model based on the improvements mentioned above. The main operations are to add a bidirectional cross-scale attention module, a dynamic deformable ROI adjustment mechanism, and set a dual-task joint loss function.

[0080] S3.4: Use random allocation to divide the dataset, set the ratio of training set, validation set, and test set to 8:1:1, adjust the training hyperparameters including learning rate, batch size, and activation function, etc., train the obtained YOLO model pt file, and convert it into the corresponding API-loadable model file format according to deployment requirements.

[0081] Example 4: Preparation and training of PPOCR model

[0082] S4.1: Build the PPOCR model training environment, install PaddlePaddle and its dependent libraries, use the PPOCRLabel semi-automatic data annotation tool to generate a special training set of bold characters for bold alphanumeric characters, and download the official dictionary file corresponding to the PPOCR model and the inference model en_PP-OCRv3_rec.

[0083] S4.2: Write a YAML configuration file to adjust the training parameters of the PPOCR model. After system training and evaluation, the obtained model file will be converted into an API-loadable format suitable for different platforms according to deployment requirements, and the model will be efficiently called through the API service interface.

[0084] Example 5: Detection and Recognition of Speed ​​Limit Signs

[0085] S5.1: Deploy the YOLO model on an edge computing device to receive image input in real time, detect speed limit signs, and output the location of the detection box.

[0086] S5.2: Crop the image based on the detection results, perform preprocessing, and then input it into the PPOCR model for digit recognition.

[0087] S5.3: After the PPOCR model recognizes the text information in the image, it performs a series of logical verifications, including checking whether the text is pure numeric and whether it complies with the numerical rules of speed limit signs, whether the number is divisible by 5, and finally outputs a valid speed limit value or outputs "unrecognized speed limit value" if it does not comply with the rules.

[0088] Example 6: System Integration and Deployment

[0089] S6.1: Integrate image acquisition, preprocessing, model detection and recognition, and result verification modules into a unified software platform to automate the speed limit sign detection and recognition process.

[0090] S6.2: Deploy the system in a real traffic environment and conduct field tests to ensure that the system operates stably and provides accurate recognition results.

[0091] By innovatively integrating the YOLO and PPOCR models, this paper proposes an efficient, economical, and easy-to-operate automatic speed limit sign detection and recognition technology. With its excellent detection accuracy, environmental adaptability, real-time processing capability, and wide applicability, it provides a strong technical support for intelligent traffic management, autonomous vehicle integration, and road inspection, greatly improving the recognition efficiency of speed limit signs and road safety.

[0092] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

[0093] It should be understood that in order to simplify the present invention and help those skilled in the art understand the various aspects of the present invention, in the above description of the exemplary embodiments of the present invention, various features of the present invention are sometimes described in a single embodiment or described with reference to a single figure. However, the present invention should not be interpreted as if all the features included in the exemplary embodiments are essential technical features of the claims of this patent.

[0094] It should be understood that the modules, units, components, etc. included in the device of one embodiment of the present invention can be adaptively changed to be installed in a device different from the embodiment. The different modules, units, or components included in the device of the embodiment can be combined into a single module, unit, or component, or they can be divided into multiple sub-modules, sub-units, or sub-components.

Claims

1. A speed limit sign detection and recognition method based on YOLO and PPOCR model, characterized in that: include: Capturing a traffic scene image as an original image; Resize the original image and use the nearest neighbor interpolation algorithm to scale the image to the set resolution; The adjusted image with the set resolution size is input into the improved YOLO model for speed limit sign detection to obtain the detection frame position information of the speed limit sign; Crop the original image based on the detection box information output by the improved YOLO model; Perform image preprocessing on the cropped image; Resize the preprocessed image and use the nearest neighbor interpolation algorithm to scale the image to the set resolution; The improved PPOCR model is used to perform digit recognition on the resized images.

2. The speed limit sign detection and recognition method based on the YOLO and PPOCR model according to claim 1 is characterized in that: The improved YOLO model includes a backbone network, a neck network, and a detection head. The specific process of the improved YOLO model for speed limit sign detection is as follows: The adjusted image of the set resolution size first enters the backbone network. After passing through two basic convolutional layers in the backbone network for spatial downsampling and channel expansion, it is reused through the first C2f module for cross-layer feature reuse. It then passes through a convolutional layer and then the second C2f module. After further processing in the backbone network, it passes through the convolutional layer, the third C2f module, the secondary convolution, and the fourth C2f module. The SPPF pyramid pooling module fuses multi-scale context information and outputs it to the neck. In the neck, the features output by the SPPF pyramid pooling module are reconstructed through two progressive upsampling operations to reconstruct the resolution. After 2 times upsampling, they are spliced ​​with the third feature A3 output by the third C2f module. The cross-layer semantic association is extracted by the fifth C2f module. After two upsamplings, they are spliced ​​with the first feature A1 output by the second C2f module. After the spatial details are fused through the sixth C2f module in the Neck network, one feature is directly enhanced by the first BiCAM module and enters the detection head 1. The other feature is spliced ​​with the output feature of the sixth C2f module after passing through the convolution layer and then fused with the spatial details through the seventh C2f module. One feature path output by the 2f module is directly enhanced by the second BiCAM module and enters detection head 2. The other feature path output by the seventh C2f module is concatenated with the feature output by the main SPPF pyramid pooling module after a convolutional layer. It is then combined with the eighth C2f module and the third BiCAM module to generate a multi-scale feature map, which is input into detection head 3 together with the first feature output A4. The three-level feature map output by the neck network is processed separately by three independent detection heads. Each detection head synchronously outputs the target bounding box coordinates, detection confidence, and classification probability. Multi-scale fusion is performed through non-maximum suppression to output the speed limit sign detection result.

3. The speed limit sign detection and recognition method based on the YOLO and PPOCR model according to claim 1 is characterized in that: According to the detection box information output by the improved YOLO model, the specific method for cropping the original image is: The conversion formula for converting the center coordinates of the detection box to the coordinates of the upper left corner and lower right corner is: In the formula, (x c ,y c ) is the center coordinate of the detection box, (x1, y1) and (x2, y2) are the coordinates of the upper left corner and lower right corner of the cropping box respectively, and w and h are the width and height of the detection box.

4. The speed limit sign detection and recognition method based on the YOLO and PPOCR model according to claim 3 is characterized in that: The BiCAM module includes three parallel pathways: residual pathway, channel attention, and spatial bidirectional scanning. The dimensions of the input and output tensors remain unchanged during processing: the residual pathway extracts basic contour features through residual blocks containing skip connections to ensure that the original shape information is not lost; the channel attention pathway simultaneously performs global average pooling and global maximum pooling on the input features, adds the two-way features, and then generates a channel weight vector through two fully connected layers. After activation by the Sigmoid function, it is multiplied with the original features channel by channel; the spatial pathway uses depth-separable 7×1 horizontal and 1×7 vertical strip convolution kernels to bidirectionally scan the image to capture the row and column arrangement of characters; finally, the three-way features are aligned through channels, fused through weighted summation, and then output with a unified dimension through 1×1 convolution.

5. The speed limit sign detection and recognition method based on the YOLO and PPOCR model according to claim 1 is characterized in that: Perform image preprocessing on the cropped image, including denoising, illumination equalization, morphological opening, binarization, and morphological closing.

6. The speed limit sign detection and recognition method based on the YOLO and PPOCR model according to claim 1 is characterized in that: The specific process of the PPOCR model for digital recognition of resized images is as follows: The resized image is passed through a lightweight MobileNetV3 backbone network to extract multi-scale features; The multi-scale features extracted by the lightweight MobileNetV3 backbone network are fused through the dynamic multi-scale feature fusion module; The fused feature maps enter the attention-guided context modeling module. The spatial attention branch of the context modeling module uses deformable convolution to adaptively capture the geometric structure of distorted text, while the channel attention branch suppresses background texture noise through channel weight recalibration. After the feature sequence processed by the context modeling module is modeled with context dependencies by BiLSTM, the CTC decoder outputs the recognition result in the form of digital character text.

7. The speed limit sign detection and recognition method based on the YOLO and PPOCR model according to claim 1 is characterized in that: The specific process of the multi-scale features being fused by the dynamic multi-scale feature fusion module is as follows: Inter-level information complementation is achieved through bidirectional cross-scale feature alignment. That is, deep high-semantic features are upsampled to shallow resolution through transposed convolution and channel-joined with shallow high-resolution features to form a fusion base with enhanced details. At the same time, shallow features are downsampled and joined with deep features, and the deep global semantics are used to enhance the contextual relevance of shallow features. Perform global average pooling on the concatenated fusion features to compress global information and generate a channel statistical vector; A lightweight network consisting of two fully connected layers is used to generate the channel attention weight matrix; The weight matrix is ​​then multiplied with the original feature map channel by channel to dynamically adjust the fusion ratio of features at different levels to obtain multi-scale fused features.

Images