Highway traffic sign detection method and device for difficult area focusing

Abstract

The invention discloses an expressway traffic sign detection method and device for difficult region focusing, and the method comprises the steps: generating a difficult region in an image: predicting the error rate of a target based on a low-resolution image and a high-resolution image, recognizing the position of a potential difficult target, generating a candidate difficult region, and determining the difficult region with the maximum performance influence; establishing an expressway traffic sign detection network framework aiming at difficult region focusing, extracting global features in the expressway traffic sign detection network framework, carrying out difficult region attention learning according to the global features, enhancing local features of the difficult region through context learning, respectively detecting the global features and the enhanced local features, and then fusing the global features and the enhanced local features; and performing supervision based on the generated difficult area, training a network framework, and adopting the trained network framework to realize highway traffic sign detection. The expressway traffic sign detection method is superior to other expressway traffic sign detection methods under qualitative and quantitative evaluation, and accurate expressway traffic sign recognition and positioning results can be obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural image processing, and relates to a highway traffic sign detection method and device for focusing on difficult areas. Background Technique

[0002] Traffic Sign Detection (TSD) is crucial for driving assistance systems and autonomous vehicles to automatically identify and classify traffic signs in real-world scenarios, and it is an important step in understanding and responding to traffic rules. Methods for automatic traffic sign detection can be roughly divided into three categories: traditional techniques, deep learning-based methods, and integrated or hybrid strategies. Traditional algorithms rely on color segmentation, shape recognition, pattern recognition, and machine learning-based detection. Deep learning algorithms, especially Convolutional Neural Network (CNN), have received extensive attention and applications because they can learn robust low-level and high-level features and achieve excellent performance. Hybrid algorithms attempt to combine the advantages of traditional algorithms or feature-based algorithms with deep learning techniques.

[0003] With the significant progress of deep learning techniques, especially deep learning-based object detection algorithms, many major advancements have also been made in traffic sign detection algorithms. Yang et al. proposed a traffic sign detection algorithm that combines an attention network with Faster-RCNN. By identifying Regions of Interest (ROIs) and clustering traffic signs, this algorithm uses a refined region generation network to determine the final candidate regions, aiming to improve the accuracy and efficiency of traffic sign detection algorithms in practical application scenarios. Ma et al. innovatively proposed a rotated region generation network, which uses a region generation-based method to generate inclined candidate boxes based on the text direction. This method effectively improves the fitting effect of the text region and facilitates subsequent correction processing, thereby enhancing the text recognition performance. Wang and Yu used an improved YOLOv4 network, which contains an additional feature layer for detecting and anchor box positioning of small and medium-sized traffic signs. Compared with large models, this improved model has fewer parameters and is faster than large models. Although the above methods have achieved certain success and progress in traffic sign detection, they are difficult to be applied to traffic sign detection in highway scenarios. Because in highway scenarios, traffic signs are sparsely distributed and mostly small targets in panoramic images, only accounting for a very small part of the entire image. Existing detection algorithms generally perform multiple downsampling operations on images, and this operation will cause the features of small targets to be submerged, making it difficult to accurately identify and locate these small targets with less pixel area.

[0004] To solve the above problems, the present invention proposes a highway traffic sign detection method and system focusing on difficult regions. Since traffic annotations in highway scenes are sparsely distributed and occupy a small area, these regions are considered difficult regions for detector recognition and positioning. To solve the problem that the small target features in these difficult regions are easily submerged by the features of other large and simple regions, a difficult region attention module is proposed. To further enrich the features of these difficult regions and make their features more distinguishable, a difficult region context encoding module is proposed to learn the context features of difficult regions in the global context and enhance the local features of difficult regions, thereby improving the discriminability of features in the detector. Summary of the Invention

[0005] Aiming at the problem that existing traffic sign detection algorithms are difficult to accurately identify and locate traffic signs that are sparsely distributed and occupy a small area in highway scenes, the purpose of the present invention is to provide a new end-to-end highway traffic sign detection solution focusing on difficult regions.

[0006] The present invention provides a highway traffic sign detection method focusing on difficult regions, including the following processes:

[0007] Generate difficult regions in the image, including identifying the positions of potential difficult targets based on the error rates of predicting targets in low-resolution and high-resolution images, and generating candidate difficult regions according to the obtained positions, and determining the difficult regions that have the greatest impact on performance from them;

[0008] Establish a highway traffic sign detection network framework focusing on difficult regions, extract the global features of the image in this network framework, and perform difficult region attention learning according to the global features; then enhance the local features of difficult regions through context learning; detect the global features and the enhanced local features respectively, and finally fuse the detection results;

[0009] Supervise based on the generated difficult regions, train the highway traffic sign detection network framework focusing on difficult regions, and use the trained network framework to implement highway traffic sign detection.

[0010] Moreover, after determining the positions of difficult targets based on the error rates of predicting targets in low-resolution and high-resolution images, the Monte Carlo method is used to randomly generate candidate region boxes as candidate difficult regions.

[0011] Moreover, calculate the ratio of the number of difficult-to-detect targets covered by each candidate difficult region to the area of this region as the difficulty score of this region, perform non-maximum suppression according to the difficulty score, and select and determine the difficult regions that have the greatest impact on performance from the suppressed candidate regions.

[0012] Moreover, in the highway traffic sign detection network framework focused on difficult regions, a difficult region attention module is set up for optimized learning of difficult regions, and the mask of the difficult region is further encoded into bounding box coordinates.

[0013] Moreover, a difficult region predictor is set up in the difficult region attention module. The difficult region predictor includes an encoder, a detector, a fusion module, and a decoder. The detector detects the features obtained by the encoder and outputs a significant feature mask, then takes the inverse of this significant feature mask, and inputs the inverse result together with the features obtained by the encoder into the fusion module. The obtained fusion result is input into the decoder, and the decoder outputs a difficult region mask. This difficult region mask is input into a multi-layer perceptron to predict the bounding box coordinates of the difficult region.

[0014] Moreover, in the highway traffic sign detection network framework focused on difficult regions, a difficult region context encoding module is set up. The bounding box coordinates of the difficult region and the global features output by the difficult region attention module are sent together into the difficult region context encoding module to learn the context representation of each difficult region and enhance the local features.

[0015] Moreover, when detecting the global features and the enhanced local features respectively, the global features are input into a global detector to obtain a global prediction result, and the enhanced local features are sent into a local detector to obtain a local prediction result. The two prediction results are fused by using non-maximum suppression to finally determine the detection result.

[0016] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned highway traffic sign detection method focused on difficult regions.

[0017] On the other hand, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the above-mentioned highway traffic sign detection method focused on difficult regions.

[0018] On the other hand, the present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the above-mentioned highway traffic sign detection method focused on difficult regions.

[0019] The present invention realizes the detection of highway traffic signs by using the method of focusing on difficult regions. Different from other invention patents, the present invention designs a difficult region attention module for optimizing the learning of difficult regions and designs a difficult region context encoding module to enhance the local features of difficult regions, so that the detection focuses more on the regions containing many difficult samples. And due to the enhancement of features, the detector can extract distinguishable features of difficult targets in these regions. Also, the present invention is different from other inventions in terms of usage skills. The present invention cleverly mines and learns difficult regions to accurately identify and locate sparse and small-sized traffic signs in difficult regions, making the test cases of the present invention superior to the existing state-of-the-art traffic sign detection methods in both subjective and objective results.

[0020] The solution proposed by the present invention is superior to other traffic sign detection methods in both qualitative and quantitative evaluations, and can generate traffic sign detection frames with accurate categories and accurate positions in highway scenarios. Brief Description of the Drawings

[0021] Figure 1 It is a schematic diagram of the detection process of highway traffic signs focusing on difficult regions according to an embodiment of the present invention.

[0022] Figure 2 It is a framework diagram of the detection network of highway traffic signs focusing on difficult regions according to an embodiment of the present invention.

[0023] Figure 3 It is an architecture diagram of the difficult region attention module and the difficult region context encoding module according to an embodiment of the present invention.

[0024] Figure 4 It is an experimental result diagram on the panoramic dataset of highways in Shanxi Province provided by an embodiment of the present invention. Among them, part (a) is the detection label; part (b) is the detection result diagram using YOLOv5; part (c) is the detection result diagram of the present invention. Detailed Embodiment

[0025] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] Refer to Figures 1 - 3 , a method for detecting highway traffic signs focusing on difficult regions according to an embodiment of the present invention includes the following specific process steps:

[0027] Step a: First, detect the low-resolution version and the high-resolution version of the image to obtain the rough detection and fine detection results. Then, analyze the errors such as false detections and missed detections between the two, and find the "difficult-to-detect targets" that are prone to errors in the rough detection but can be corrected by the fine detection to obtain the positions of potential difficult targets. Subsequently, use the Monte Carlo method to randomly generate a batch of candidate rectangular regions, that is, obtain the candidate difficult regions, and determine the difficult regions that have the greatest impact on performance by calculating the area occupied by the region and the number of difficult targets therein.

[0028] Furthermore, the present invention proposes that as many difficult targets as possible should be covered while balancing the relationship between the region area and the number of targets; select the regions that cover the most difficult targets and have the smallest area as the difficult regions. In the embodiment, it is preferably to calculate the "ratio of the number of difficult-to-detect targets covered by each region to the region area" and perform non-maximum suppression, and finally select several regions with the highest scores as the difficult regions.

[0029] In the embodiment, use the low-resolution detection model and the high-resolution detection model to detect the low-resolution image and the high-resolution image respectively to obtain the rough detection result and the fine detection result. Then, preferably use the object detection error analysis tool TIDE (A General Toolbox for Identifying Object Detection Errors) to perform error analysis on the rough detection result and the fine detection result respectively, and select the set of bounding boxes of the "difficult-to-detect targets" that are prone to errors in the rough detection but can be corrected by the fine detection. To determine the difficult regions, use the Monte Carlo method to randomly generate several candidate rectangular regions in the image, and calculate the ratio of the number of difficult-to-detect targets covered by each region to the area of the region as the difficulty score of the region in turn. Perform maximum suppression on all candidate regions according to this score, and finally select K regions from the suppressed candidate regions as the difficult regions. Specifically in implementation, other object detection error analysis tools can also be used according to needs, and the value of K can be set according to the actual situation.

[0030] Step b: Establish a highway traffic sign detection network framework focused on difficult regions. In this network framework, mainly extract the global features of the image, and perform difficult region learning through the difficult region attention module according to this feature; then use the difficult region context module to enhance the local features of the difficult regions; use the global detector and the local detector to detect the global features and the enhanced local features respectively, and finally fuse the detection results to obtain the prediction result.

[0031] See Figure 2, in the embodiment of the present invention, the highway traffic sign detection network framework set for focusing on difficult areas includes a backbone network for extracting global and local features, a difficult area attention module, a difficult area context encoding module, a global detector, a local detector, and a detection result fusion module. The image is first input into the backbone network to extract global features, and then the extracted global features are input into the difficult area attention module to learn the difficult area mask. Then, the masked difficult area is input into the backbone network to extract local features. The local features and the coordinates of the difficult area mask are input into the difficult area context encoding module together for context learning to obtain enhanced local features. The enhanced local features are input into the local detector, and at the same time, the global features are sent into the global detector. Finally, the results of the two detectors are fused to obtain the final prediction result.

[0032] The specific processing process of this network framework includes the following:

[0033] (1) The backbone network used extracts global features from the image and sends the global features into the difficult area attention module for learning of difficult areas. The learning process is supervised by the ground truth of difficult areas obtained in step a. The learned difficult area mask is further encoded into bounding box coordinates.

[0034] Specifically, the backbone network preferably adopts a Convolutional Neural Network (CNN).

[0035] The difficult area attention module includes a difficult area predictor and a multi-layer perceptron.

[0036] Specifically, the global feature extracted by the backbone network is F glob , and the process of inputting this feature into the difficult area attention module for difficult mask learning is as follows:

[0037]

[0038] Among them, represents the difficult area predictor, and M D is the difficult area mask.

[0039] This embodiment preferably proposes that the implementation method of the difficult area predictor is as follows:

[0040] The difficult area predictor includes an encoder, a detector, a fusion module, and a decoder. The encoder consists of several sequentially arranged 3×3 convolutional layers (preferably three convolutional layers). The detector detects the features downsampled to the smallest size obtained by the encoder and outputs a significant feature mask M S , and then the mask is inverted to obtain MR , input M R and the features downsampled to the minimum size output by the encoder into the fusion module. The fusion module performs a pixel-by-pixel multiplication operation on the two feature maps, and then inputs the obtained fusion result into the decoder. The decoder consists of several sequentially arranged 3×3 convolutional layers (preferably three convolutional layers) and a 1×1 output convolutional layer, and finally obtains the hard region mask M D . This mask is input into a multi-layer perceptron to obtain the bounding box coordinates of the hard region. This process can be expressed as:

[0041] Coord (x,y) = MLP(M D )

[0042] where Coord (x,y) represents the bounding box coordinates of the hard region, and MLP(·) represents the multi-layer perceptron, which consists of three fully connected layers.

[0043] (2) Extract the local hard region according to the bounding box coordinates of the hard region, and apply the same backbone network as in (1) to this region to extract local features. To enhance the local feature context information, send the bounding box coordinates of the hard region and this local feature into the hard region context encoding module to learn the context representation of each hard region.

[0044] The hard region context encoding module includes a context generator and a feature information enhancement module.

[0045] Specifically, input the bounding box coordinates of the hard region and the global features extracted by the backbone network into the context generator in the hard region context encoding module for fusion and encoding to obtain the context representation of each hard region; then fuse this context representation with the local features extracted for the hard region and send it into the feature information enhancement module to enhance the local features.

[0046] The process of inputting the bounding box coordinates of the hard region and the global features into the context generator can be expressed as:

[0047] F context = θ CG (F glob , Coord (x,y) )

[0048] where θ CG (·) is the context generator, which performs operations on the input global feature F globFirst, it is processed through a 1×1 convolutional layer, and the resulting matrix is flattened into a vector, which is concatenated with the coordinates of the difficult region bounding boxes that have passed through multiple linear layers (preferably three linear layers). Then, it is encoded through multiple linear layers (preferably three linear layers) to obtain the context representation F context . This representation can be used to enhance the local features of the difficult regions, and the feature information enhancement process can be expressed as:

[0049] F fused =θ FIE (F context ,F local )

[0050] Among them, F local is the local feature extracted for the difficult region, and θ FIE (·) is the feature information enhancement module. The first branch of this module copies the context representation and copies it into a two-dimensional matrix. The second branch inputs the local feature into a 1×1 convolutional layer to change the number of channels, and separates the result from the channel dimension. Then, the features of each separated channel dimension obtained by the second branch are multiplied element-wise with the two-dimensional matrix of the context representation obtained by the first branch, and after combining the operation results of each channel, they are sequentially input into a 3×3 convolutional layer and a 1×1 convolutional layer to obtain the finally enhanced fused feature F fused .

[0051] (3) Send the extracted global feature into the global detector to obtain the global prediction result, send the enhanced local feature into the local detector to obtain the local prediction result, and then use the non-maximum suppression method to fuse the prediction results of the two to obtain the final detection prediction result.

[0052] In the embodiment, the global feature is input into the global detector to obtain the global prediction result, the enhanced local feature is sent into the local detector to obtain the local prediction result, and the two prediction results are fused by using non-maximum suppression to finally determine the detection result.

[0053] Specifically, the process of using the global and local detectors for result prediction is as follows:

[0054] pred glob =D glob (F glob )

[0055] pred local =D local (F fused )

[0056] Among them, D glob (·) is the global detector, D local(·) is a local detector, pred glob and pred local are the results of global and local predictions respectively.

[0057] Merge the results of global and local predictions, and use the non-maximum suppression algorithm for processing to remove redundant predicted bounding boxes, then the final prediction result can be obtained.

[0058] Step c, training stage: First, use the difficult regions of the images obtained in step a as the ground truth for the supervised difficult region attention module. Then, input the images into the network framework established in step b. In it, first the backbone network extracts global features, and then input the features into the difficult region attention module to learn the difficult region mask. Input the bounding box coordinates of the difficult region mask and the local features extracted from the difficult region into the difficult region context encoding module for context feature learning to enhance the local features. The global features and the enhanced local features are respectively input into the global detector and the local detector, and the detection results output by the two are fused through maximum suppression to obtain the final detection result.

[0059] Step d, testing stage: Based on the trained network framework, input the test input images into the backbone network for global feature extraction. Then, input the global features into the difficult region attention module to automatically extract the difficult region mask. Then, the bounding box coordinates of the difficult region mask and the local features extracted from the difficult region are input into the difficult region context encoding module together to obtain the enhanced local features. Finally, input the global features and the enhanced local features into the global detector and the local detector respectively for detection result prediction, and use non-maximum suppression to fuse the prediction results to obtain the final result.

[0060] In specific implementation, input the real-time captured images into the network framework based on the trained one, and the detection results can be obtained in real time by using the test method.

[0061] In specific implementation, the method proposed by the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. The system devices for implementing the method, such as the computer-readable storage medium storing the corresponding computer program of the technical solution of the present invention and the computer device including running the corresponding computer program, should also be within the protection scope of the present invention.

[0062] The following embodiments describe the establishment of an electronic device for the highway traffic sign detection method focusing on difficult regions provided by the present invention. The establishment of the electronic device for the highway traffic sign detection method focusing on difficult regions described below can be mutually corresponding and referred to the highway traffic sign detection method focusing on difficult regions described above.

[0063] The electronic device may include: a processor, a communications interface, a memory, and a communication bus. Among them, the processor, the communications interface, and the memory complete communication with each other through the communication bus. The processor may call the logical instructions in the memory to execute the highway traffic sign detection method for focusing on difficult areas, mainly including the software processing part in the above steps.

[0064] In addition, when the logical instructions in the above-mentioned memory are implemented in the form of software functional units and sold or used as independent products, they may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, may be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0065] In some possible embodiments, a highway traffic sign detection system for focusing on difficult areas is provided, including the following modules

[0066] The first module is used to generate difficult areas in the image, including identifying the positions of potential difficult targets based on the error rates of predicting targets in low-resolution and high-resolution images, generating candidate difficult areas according to the obtained positions, and determining the difficult area that has the greatest impact on performance from them;

[0067] The second module is used to establish a highway traffic sign detection network framework for focusing on difficult areas, extract the global features of the image in this network framework, and perform difficult area attention learning based on the global features; then enhance the local features of the difficult areas through context learning; detect the global features and the enhanced local features respectively, and finally fuse the detection results;

[0068] The third module is used to perform supervision based on the generated difficult areas, train the highway traffic sign detection network framework for focusing on difficult areas, and use the trained network framework to implement highway traffic sign detection.

[0069] In some possible embodiments, a non-transitory computer-readable storage medium is provided, including a readable storage medium, on which a computer program is stored. When the computer program is executed, a highway traffic sign detection method focusing on difficult areas as described above is implemented.

[0070] In some possible embodiments, a computer program product is provided, including a computer program, which implements the highway traffic sign detection method focusing on difficult areas as described above when executed by a processor.

[0071] Referring to Figure 3 , the application test experiment of the embodiment of the present invention is as follows:

[0072] The experiment uses the panoramic dataset of expressways in Shanxi Province, and 80 panoramic images in the test set are selected. The size of each panoramic image is 4096×8192 pixels.

[0073] The present invention is compared with other highway traffic sign detection methods, and experimental data are provided to demonstrate the effectiveness of the present invention. The evaluation indexes of the comparative experimental results are shown in Table 1 below, and the experimental results are as Figure 4 shown. Part (a) is the detection label; part (b) is the detection result diagram using YOLOv5; part (c) is the detection result diagram of the present invention.

[0074] Table 1

[0075]

[0076] It can be seen from the experimental data in the above table that the method of the present invention has obtained higher accuracy compared with the comparative methods, that is, it is better than the comparative methods.

[0077] In summary, the highway traffic sign detection method and system for focusing on difficult regions proposed by the present invention first define and generate difficult regions in the image, determine the positions of difficult targets by predicting the error rates of targets in low-resolution and high-resolution images, select difficult regions using the Monte Carlo method based on these positions, and determine the difficult regions that have the greatest impact on performance by calculating the area occupied by the regions and the number of difficult targets therein; then use a backbone network to extract global features of the image, and send the global features into the difficult region attention module for optimized learning of difficult regions, and the masks of difficult regions are further encoded as bounding box coordinates; then send the bounding box coordinates of the previously obtained difficult regions and the global features together into the difficult region context encoding module to learn the context representation of each difficult region and enhance local features; finally, send the extracted global features into the global detector to obtain the global prediction result, send the enhanced local features into the local detector to obtain the local prediction result, and then fuse these prediction results to obtain the final detection prediction result. The network proposed by the present invention is superior to other highway traffic sign detection methods in both qualitative and quantitative evaluations.

[0078] Compared with the existing methods, the advantages and positive effects of the present invention are as follows: The present invention proposes a new traffic sign detection method for panoramic images of highway scenes. This method mainly uses the difficult region attention module to make the detection pay more attention to the regions where small-sized traffic signs with sparse distribution are located; at the same time, a new difficult region context encoding module is designed to enhance the feature representation of difficult regions to improve the final prediction result. The superiority of the method proposed by the present invention has been effectively verified from both subjective and objective evaluation indexes on the panoramic dataset of highways in Shanxi Province.

[0079] The specific embodiments described in this article are only examples to illustrate the spirit of the present invention. Those skilled in the art of the present invention can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A freeway traffic sign detection method for focusing on difficult areas, characterized in that, It includes the following processes: Generate difficult regions in the image, including identifying the positions of potential difficult targets based on the error rates of predicting targets from low-resolution and high-resolution images, and generating candidate difficult regions according to the obtained positions, and determining the difficult region with the greatest impact on performance from them; Establish a highway traffic sign detection network framework focused on difficult regions. Extract the global features of the image in this network framework, and perform difficult region attention learning based on the global features; then enhance the local features of the difficult regions through context learning; detect the global features and the enhanced local features respectively, and finally fuse the detection results; Supervise based on the generated difficult regions, train the highway traffic sign detection network framework focused on difficult regions, and use the trained network framework to implement highway traffic sign detection.

2. The highway traffic sign detection method for focusing on difficult areas according to claim 1, wherein: After determining the difficult target positions based on the error rates of predicting targets from low-resolution and high-resolution images, use the Monte Carlo method to randomly generate candidate region boxes as candidate difficult regions.

3. The highway traffic sign detection method for focusing on difficult areas according to claim 1, characterized in that: Calculate the ratio of the number of difficult-to-detect targets covered by each candidate difficult region to the area of this region as the difficulty score of this region, perform maximum suppression according to the difficulty score, and select and determine the difficult region with the greatest impact on performance from the suppressed candidate regions.

4. A method for detecting highway traffic signs focusing on difficult areas according to claim 1, characterized in that: In the highway traffic sign detection network framework focused on difficult regions, a difficult region attention module is set to perform optimization learning of difficult regions, and the mask of the difficult region is further encoded as the bounding box coordinates.

5. A highway traffic sign detection method for focusing on difficult areas according to claim 4, characterized in that: A difficult region predictor is set in the difficult region attention module. The difficult region predictor includes an encoder, a detector, a fusion module, and a decoder. The detector detects the features obtained by the encoder and outputs a significant feature mask, then invert the significant feature mask, input the inverted result and the features obtained by the encoder into the fusion module together, the obtained fusion result is input into the decoder, the decoder outputs a difficult region mask, and this difficult region mask is input into a multi-layer perceptron to predict the bounding box coordinates of the difficult region.

6. The method for detecting highway traffic signs focusing on difficult areas according to claim 4, characterized in that: In the highway traffic sign detection network framework focused on difficult regions, a difficult region context encoding module is set. The bounding box coordinates and global features of the difficult region output by the difficult region attention module are sent into the difficult region context encoding module together to learn the context representation of each difficult region and enhance the local features.

7. A method for detecting highway traffic signs focusing on difficult areas according to claim 1, characterized in that: When detecting the global features and the enhanced local features respectively, input the global features into the global detector to obtain the global prediction result, input the enhanced local features into the local detector to obtain the local prediction result, and fuse the two prediction results through non-maximum suppression to finally determine the detection result.

8. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the highway traffic sign detection method focused on difficult regions as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the highway traffic sign detection method focused on difficult regions as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, it implements the highway traffic sign detection method for focusing on difficult areas according to any one of claims 1 to 7.

Images