Traffic sign detection method and device and electronic equipment

Abstract

The invention provides a traffic sign detection method and device and electronic equipment, and relates to the technical field of target detection, and the method comprises the steps: carrying out the multi-mode feature extraction of a to-be-detected image through a plurality of first extraction modules, and carrying out the element-by-element multiplication of the features of a plurality of modes, and obtaining a plurality of feature maps; multiple modes of feature extraction are carried out on the multiple feature maps through multiple first extraction modules, multiple modes of features are multiplied element by element, feature fusion is carried out on the processed feature maps, and multiple fused feature maps are obtained; and carrying out differential convolution processing on the plurality of fusion feature maps through the plurality of detection modules, and carrying out traffic sign identification on the processed fusion feature maps to obtain a detection result. According to the embodiment of the invention, the small target recognition precision can be improved while the parameter quantity and the calculation quantity are reduced by optimizing the structure of the model core module.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to a traffic sign detection method, device and electronic equipment. BACKGROUND

[0002] Under the background of rapid development of intelligent transportation system, traffic sign detection, as the core perception link of automatic driving and vehicle auxiliary driving system, directly affects driving safety and traffic efficiency in terms of detection accuracy and deployment adaptability. Currently, traffic sign detection technology has been widely applied in vehicle terminal, road intelligent monitoring equipment and other scenarios, and needs to realize real-time recognition of multiple types of traffic signs such as no entry, speed limit and direction indication, and accurately capture small target traffic signs at a long distance and small size to provide reliable data support for driving decision.

[0003] Existing traffic sign detection schemes are mostly based on traditional target detection models, but there are problems in actual application: the model parameter quantity and calculation quantity are too large, and it is difficult to adapt to vehicle deployment: the backbone network of the traditional YOLOv11 model adopts C3K2 module (containing Bottleneck structure), and the detection head is a conventional multi-branch structure, which has high overall parameter quantity and long calculation time. Vehicle terminal devices (such as vehicle embedded terminal and automatic driving perception module) are limited by hardware computing power and power consumption, and it is difficult to meet the real-time inference requirements of traditional models, which may cause detection delay and affect the timeliness of driving assistance decision. SUMMARY

[0004] The present application provides a traffic sign detection method, device and electronic equipment, which can optimize the structure of the model core module to reduce the parameter quantity and calculation quantity while improving the small target recognition accuracy.

[0005] In a first aspect, the present application provides a traffic sign detection method based on YOLOv11, comprising: constructing a detection model based on YOLOv11; wherein the detection model includes a plurality of first extraction modules containing C3K2_Star structure and a plurality of detection modules containing DSCDH structure; inputting a to-be-detected image into the detection model, and outputting a detection result from the detection model; wherein the operation process of the detection model comprises: performing multi-mode feature extraction on the to-be-detected image through a plurality of first extraction modules, and performing element-by-element multiplication processing on the multi-mode features to obtain a plurality of feature maps; performing multi-mode feature extraction on the plurality of feature maps through a plurality of first extraction modules, performing element-by-element multiplication processing on the multi-mode features, and performing feature fusion on the processed feature maps to obtain a plurality of fused feature maps; The plurality of detection modules perform differential convolution processing on the plurality of fusion feature maps, and traffic signs are recognized from the processed fusion feature maps to obtain a detection result.

[0006] The method can extract features in multiple ways from the to-be-detected image through the C3K2_Star structure, and process the features extracted in multiple ways through element-by-element multiplication, thereby reducing the amount of calculation and strengthening the expression of small target features. On the other hand, the DESCDH is proposed to recognize traffic signs from the plurality of fusion feature maps through differential convolution to obtain a detection result. This design reduces parameter redundancy. The present application optimizes the structure of the core module of the model, thereby reducing the amount of parameters and calculation while improving the small target recognition accuracy.

[0007] In a possible implementation, the detection model includes a backbone network, a neck network, and a detection head network; wherein the first extraction module containing the C3K2_Star structure is in the backbone network, the first extraction module containing the C3K2_Star structure is in the neck network, and the plurality of detection modules containing the DESCDH structure are in the detection head network.

[0008] In a possible implementation, the backbone network further includes a plurality of second extraction modules; the plurality of second extraction modules and the plurality of first extraction modules containing the C3K2_Star structure are alternately arranged in the backbone network. The plurality of first extraction modules extract features in multiple ways from the to-be-detected image, and perform element-by-element multiplication processing on the features in multiple ways to obtain a plurality of feature maps, including: The following process is performed multiple times: The second extraction module extracts basic features from the input feature map to obtain a plurality of basic feature maps. The first extraction module containing the C3K2_Star structure extracts features in multiple ways from the basic feature map output by the second extraction module, and performs element-by-element multiplication processing on the features in multiple ways to obtain a plurality of feature maps.

[0009] In a possible implementation, the C3K2_Star structure includes a Star_Block structure; the Star_Block structure includes a plurality of convolution layers and an element-by-element multiplication module; a part of the convolution layers is arranged in a two-way parallel structure; another part of the convolution layers is arranged in a serial structure. The first extraction module containing the C3K2_Star structure extracts features in multiple ways from the basic feature map output by the second extraction module, and performs element-by-element multiplication processing on the features in multiple ways to obtain a plurality of feature maps, including: The two parallel structures are used for respectively performing feature extraction on the basic feature maps output by the second extraction module, to obtain two feature maps; The element-by-element multiplication module is used for performing element-by-element multiplication processing on the two feature maps; The serial structure is used for performing feature extraction on the processed features, and residual processing is performed on the extracted features and the basic feature maps output by the second extraction module, to obtain a plurality of feature maps.

[0010] In a possible implementation, the neck network further includes a plurality of concatenation modules and a plurality of third extraction modules; the plurality of third extraction modules and the plurality of first extraction modules including the C3K2_Star structure form a plurality of two parallel structures; one of the two parallel structures includes the third extraction module, and the other of the two parallel structures includes the first extraction module including the C3K2_Star structure.

[0011] The plurality of first extraction modules are used for performing multi-mode feature extraction on the plurality of feature maps, and element-by-element multiplication processing is performed on the multi-mode features, to obtain a plurality of fused feature maps, including: The following process is performed multiple times: The third extraction module of one of the two parallel structures is used for performing feature extraction on the input feature map; The first extraction module including the C3K2_Star structure of the other of the two parallel structures is used for performing multi-mode feature extraction on the basic feature maps output by the second extraction module, and element-by-element multiplication processing is performed on the multi-mode features; The concatenation module is used for performing concatenation processing on the outputs of the two parallel structures, to obtain the fused feature maps; The third extraction module includes a convolution layer or an up-sampling.

[0012] In a possible implementation, the neck network outputs feature maps of multiple scales at different layer structures, and the feature maps of the multiple scales are respectively input into detection modules including the DSCDH structure corresponding to the scales. The DSCDH structure includes a plurality of shared convolution layers, a recognition convolution layer, and a plurality of output modules. The plurality of detection modules are used for performing differential convolution processing on the plurality of fused feature maps, and traffic signs are recognized from the processed fused feature maps, to obtain detection results, including: The plurality of shared convolution layers are used for performing differential convolution processing on the input feature maps; The recognition convolution layer is used for performing convolution processing on the feature maps output by the plurality of shared convolution layers, to recognize features corresponding to traffic signs and features corresponding to positions of the traffic signs. For each scale of feature map, the output module of the DESCDH structure that processes the feature map of that scale is used to identify traffic signs and their location by processing the output of the recognition convolutional layer, and output the traffic signs and their location.

[0013] In one possible implementation, the shared convolutional layer includes multiple differential convolutional layers, an addition module, a regular convolutional layer, and an equivalent convolutional layer; The input feature map is processed by differential convolution through multiple shared convolutional layers, including: The input feature map is processed by differential convolution and basic convolution through multiple differential convolution layers and the ordinary convolution layer to obtain weight values ​​and biases. The weight value and the bias are added together by the addition module; The output of the addition module is convolved by the equivalent convolutional layer.

[0014] In one possible implementation, the method further includes: Obtain a training set; wherein, the actual result of the images in the training set is the identification of traffic signs in the images and their positions in the images; In each training process, the images in the training set are input into the base model, and the base model outputs the detection results. Based on the detection results and the actual results, the parameters of the base model are updated using an optimizer, and the parameters in the DESCDH structure of the base model are updated using the distributed focus loss function. After training, the base model with the last updated parameters is used as the detection model.

[0015] Secondly, embodiments of the present invention provide a traffic sign detection device based on YOLOv11, comprising: A building module is used to build a detection model based on YOLOv11; wherein, the detection model includes multiple first extraction modules containing C3K2_Star structures and multiple detection modules containing DESCDH structures; The processing module is used to input the image to be detected into the detection model, and the detection model outputs the detection result; The operation of the detection model includes: Multiple first extraction modules are used to extract features from the image to be detected in various ways, and the features from these multiple methods are multiplied element-wise to obtain multiple feature maps. Multiple feature maps are extracted using multiple first extraction modules in various ways, and the features from these multiple methods are multiplied element-wise. The processed feature maps are then fused to obtain multiple fused feature maps. Multiple detection modules perform differential convolution processing on multiple fused feature maps, and then use the processed fused feature maps to identify traffic signs, thus obtaining the detection results.

[0016] Thirdly, embodiments of the present invention provide an electronic device, comprising: processor; A processor for executing a computer program or instructions in the memory, such that a YOLOv11-based traffic sign detection method as described in any of the first aspects is performed.

[0017] Fourthly, embodiments of the present invention provide a computer-readable storage medium that, when instructions in the storage medium are executed by a processor, enables the processor to perform a YOLOv11-based traffic sign detection method as described in any of the first aspects.

[0018] Fifthly, embodiments of the present invention provide a computer program product, the computer program product comprising: computer program code, which, when run on a computer, causes the computer to execute the traffic sign detection method based on YOLOv11 as described in any of the first aspects.

[0019] Furthermore, the technical effects of any of the implementation methods in the second to fifth aspects can be found in the technical effects of different implementation methods in the first aspect, and will not be repeated here.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of a detection model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating an application scenario of a traffic sign detection method based on YOLOv11 provided in an embodiment of the present invention. Figure 3 A schematic diagram illustrating an application scenario of another traffic sign detection method based on YOLOv11 provided in an embodiment of the present invention; Figure 4A A flowchart illustrating a traffic sign detection method based on YOLOv11 provided in an embodiment of the present invention; Figure 4B A flowchart illustrating the computation process of a detection model provided in an embodiment of the present invention; Figure 5 A flowchart illustrating a method for feature extraction from a backbone network provided in an embodiment of the present invention; Figure 6A schematic diagram of a detection model provided in an embodiment of the present invention; Figure 7 A flowchart illustrating a method for feature extraction using a Star_Block structure, provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of a Star_Block structure provided in an embodiment of the present invention; Figure 9 A structural diagram of a C3K2-Star structure provided in an embodiment of the present invention; Figure 10 A structural diagram of a C3K-Star structure provided in an embodiment of the present invention; Figure 11 A flowchart of a method for feature fusion via a neck network is provided in an embodiment of the present invention; Figure 12 This is a flowchart of a method for detection using a detection head network, provided by an embodiment of the present invention. Figure 13 A structural diagram of a DESCDH structure provided in an embodiment of the present invention; Figure 14 A flowchart illustrating a method for convolution processing using shared convolutional layers, provided by an embodiment of the present invention; Figure 15 This is a structural diagram of a shared convolutional layer provided in an embodiment of the present invention; Figure 16 A structural diagram of a traffic sign detection device based on YOLOv11 provided in an embodiment of the present invention; Figure 17 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0023] In the description of the embodiments of this application, unless otherwise stated, "multiple" refers to two or more.

[0024] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0025] As mentioned earlier, the current model built using YOLOv11 detects small-target traffic signs, combined with... Figure 1 As shown, the basic structure of YOLOv11 includes a backbone network 100, a neck network 110, and a detection head network 120; The backbone network is the foundation of the model and is responsible for extracting multi-level deep feature maps from the input image.

[0026] The YOLOv11 backbone network typically undergoes a series of complex convolutional bottleneck modules. Specifically, multiple convolutional bottleneck modules extract features from the image to be detected, resulting in multiple feature maps. The role of the neck network is to efficiently fuse and enhance multiple feature maps extracted from the backbone network to generate a more representative feature pyramid.

[0027] Core Structure: YOLOv11 employs a structure combining a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN), namely PANet. Fusion Mechanism: (1) FPN (Top-down path): High-level semantic information from the deep layers of the backbone network is passed and fused into the high-resolution feature maps of the shallow layers through upsampling operations, thereby supplementing the shallow features with rich contextual information.

[0028] (2) PAN (bottom-up path): Based on FPN, a bottom-up path is added to transfer and fuse the precise positioning information of the shallow layer into the feature map of the deep layer through downsampling operation, which further enhances the positioning capability of features at all scales.

[0029] Output: After thorough fusion by PANet, the neck network finally outputs multiple fused feature maps, which are then used by the subsequent detection head for prediction.

[0030] The head network is the final output layer of the model, responsible for performing dense classification and bounding box regression prediction based on the feature maps fused from the neck network.

[0031] However, the backbone network of the aforementioned YOLOv11 model contains a Bottleneck structure. The Bottleneck structure processes features through a simple process of "dimensionality reduction-feature extraction-dimensionality increase," lacking the ability to nonlinearly combine subtle features of small targets. It cannot effectively capture crucial information such as edges and textures of distant, low-resolution traffic signs. The Bottleneck structure relies on increasing the number of channels or stacking layers to improve performance, resulting in high parameter and computational costs, making it difficult to adapt to the computing power and power consumption constraints of in-vehicle embedded devices. During dimensionality reduction and increase, the hierarchical features of small targets are easily diluted, especially during multi-scale feature fusion, making it difficult to retain effective information of small-sized traffic signs and affecting detection accuracy. The detection head is a conventional multi-branch structure, with high overall parameter count and long computation time. When traditional YOLOv11 models are loaded onto in-vehicle devices (such as in-vehicle embedded terminals and autonomous driving perception modules), the hardware computing power and power consumption limitations make it difficult to meet the real-time inference requirements of traditional models, easily leading to detection delays and affecting the timeliness of driving assistance decisions.

[0032] Based on this, this invention provides a traffic sign detection scheme based on YOLOv11. In the detection model built with YOLOv11, the C3K2_Star structure replaces the Bottleneck structure, and the DESCDH structure replaces the detection head. The improved detection model detects traffic signs. This invention uses the C3K2_Star structure to extract features from the image to be detected in multiple ways, and processes the features extracted in multiple ways by multiplying them element-wise, thereby reducing the amount of computation and enhancing the feature representation of small targets. On the other hand, the invention proposes DESCDH to identify traffic signs by differential convolution of multiple fused feature maps, thereby reducing parameter redundancy. Thus, this invention improves the accuracy of small target recognition while reducing the number of parameters and the amount of computation.

[0033] The objectives, functional features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0034] The traffic sign detection method based on YOLOv11 provided in this embodiment of the invention is applicable to electronic devices, which use the above method to complete the detection of traffic signs.

[0035] For example, the electronic device in this embodiment of the invention can be an in-vehicle terminal. An in-vehicle terminal is a core hardware device for intelligent transportation and vehicle intelligence, integrating functions such as data collection, processing, communication, and interaction. It directly serves scenarios such as driver assistance, travel management, and infotainment, and is a key hub connecting vehicles, people, roads, and the cloud. An in-vehicle terminal refers to an embedded device or intelligent module installed inside a vehicle, possessing hardware computing power and software system compatibility. Its core positioning is the "intelligent hub of the vehicle," requiring real-time collection of data such as vehicle status and road environment, and data interaction through algorithm processing and network transmission, providing underlying hardware support for scenarios such as autonomous driving, in-vehicle assistance, ride-hailing dispatching, and traffic sign recognition.

[0036] For example, the electronic device in the embodiments of the present invention can also be a server. For instance, the server can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The server can be connected to a device with display function via wired or wireless means, and the present invention does not limit the connection method.

[0037] It is understood that the present invention does not limit the specific type of the above-mentioned electronic device.

[0038] The following describes, with reference to the accompanying drawings, possible application scenarios of the traffic sign detection method based on YOLOv11 provided in this embodiment of the invention: like Figure 2 The diagram illustrates an application scenario of this invention, using an electronic device, specifically a vehicle-mounted device, as an example. (See attached image.) Figure 2 As shown, this application scenario includes an in-vehicle device 210 and a camera 220. The in-vehicle device 210 executes the flow of the traffic sign detection method based on YOLOv11 provided in this embodiment of the invention (the specific flow will be described later), identifies the traffic sign, and the camera 220 is used to capture the image to be detected.

[0039] In one or more embodiments, after the camera 220 acquires road images, it directly transmits them to the vehicle terminal 210. The vehicle terminal 210 executes the detection method of the present invention and outputs the name and location information of traffic signs in real time, providing data support for driving assistance systems or autonomous driving decisions.

[0040] like Figure 3 As shown, taking an electronic device as a server as an example, this illustrates another application scenario diagram in an embodiment of the present invention. (See attached diagram.) Figure 3As shown, this application scenario includes a server 310, an in-vehicle device 210, and a camera 220. The camera 220 is connected to the server 310 via the in-vehicle device 210. The server 310 and the in-vehicle device 210 are connected. The server 310 executes the flow of the traffic sign detection method based on YOLOv11 provided in this embodiment of the invention (the specific flow will be described later) to determine the traffic sign. The camera 220 is used to capture the image to be detected, and the in-vehicle device 210 is used for assisted driving.

[0041] In one or more embodiments, after the camera 220 acquires the image to be detected, it transmits it to the vehicle-mounted device 210. The vehicle-mounted terminal 210 transmits the image to be detected to the server 310. The server 310 executes the detection method of the present invention, outputs the name and location information of the traffic sign in real time, and sends the name and location information of the traffic sign to the vehicle-mounted terminal 210. The vehicle-mounted terminal 210 provides driving assistance based on the name and location information of the traffic sign.

[0042] Of course, the methods provided in the embodiments of the present invention are not limited to those described above. Figure 2 and Figure 3 The application scenarios shown can also be used in other possible application scenarios, and the embodiments of the present invention do not impose any limitations.

[0043] After introducing the application scenarios of the embodiments of the present invention, the preferred embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only for illustration and explanation of the present invention and are not intended to limit the present invention. Furthermore, the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0044] The following is a detailed description of the traffic sign detection method based on YOLOv11 provided by the embodiments of the present invention, with reference to the accompanying drawings: Figure 4A and Figure 4B This diagram illustrates the workflow of a traffic sign detection method based on YOLOv11 provided by an embodiment of the present invention. Figure 4A and Figure 4B As shown, the specific process of this method is as follows: S410: A detection model is built based on YOLOv11; the detection model includes multiple first extraction modules containing C3K2_Star structures and multiple detection modules containing DESCDH structures; S420: The image to be detected is input into the detection model, and the detection model outputs the detection result; the calculation process of the detection model includes: S420-1: Multiple first extraction modules extract features from the image to be detected in various ways, and multiply the features in various ways element by element to obtain multiple feature maps; S420-2: Multiple feature maps are extracted using multiple first extraction modules in various ways, and the features from these multiple methods are multiplied element-wise. The processed feature maps are then fused to obtain multiple fused feature maps. S420-3: Multiple detection modules perform differential convolution processing on multiple fused feature maps, and then use the processed fused feature maps to identify traffic signs to obtain detection results.

[0045] Furthermore, when the detection results need to be displayed to the user, if the results include the identifier of the traffic sign in the image to be detected and its position within the image, then the identifier of the traffic sign in the image to be detected is used to determine the name of the traffic sign, and the name and position of the traffic sign in the image to be detected are displayed to the user. Of course, if there are no traffic signs in the image to be detected, the detection result outputs 0. The detection result can be converted into a preset message, i.e., no traffic sign was detected, and displayed to the user on the screen. Alternatively, if there are no traffic signs in the image to be detected, the detection result does not need to be displayed, and detection can continue.

[0046] In some embodiments, the detection model includes a backbone network, a neck network, and a detection head network; wherein a first extraction module containing a C3K2_Star structure is partially in the backbone network, a first extraction module containing a C3K2_Star structure is partially in the neck network, and multiple detection modules containing a DESCDH structure are in the detection head network.

[0047] Step 420-1 is implemented through a backbone network. Multiple first extraction modules containing C3K2_Star structures in the backbone network extract features from the image to be detected in various ways, and perform element-wise multiplication of the features in various ways to obtain multiple feature maps. Step 420-2 is implemented through the neck network. Multiple feature extraction modules containing C3K2_Star structures in the neck network extract features from multiple feature maps in various ways, and perform element-wise multiplication of the features in various ways. The processed feature maps are then fused to obtain multiple fused feature maps. Step 420-3 is implemented through a detection head network. Multiple detection modules containing DESCDH structures in the detection head network perform differential convolution processing on multiple fused feature maps, and then perform traffic sign recognition on the processed fused feature maps to obtain the detection results.

[0048] The backbone network also includes multiple second extraction modules; multiple second extraction modules and multiple first extraction modules containing C3K2_Star structures are alternately set in the backbone network; Multiple secondary extraction modules include multiple convolutional layers.

[0049] Based on the above structure, in some embodiments, multiple first extraction modules perform feature extraction on the image to be detected in various ways, and multiply the features in these various ways element-wise to obtain multiple feature maps, including: Execute the following process multiple times, combined with Figure 5 As shown: S510: The second extraction module extracts basic features from the input feature map to obtain multiple basic feature maps; S520: The first extraction module, which contains the C3K2_Star structure, performs feature extraction on the basic feature map output by the second extraction module in multiple ways, and performs element-wise multiplication on the features in multiple ways to obtain multiple feature maps.

[0050] In the above embodiments of the present invention, the second extraction module is a conventional convolution, which excels at quickly extracting basic features (such as edges and color gradients) and lays the foundation for features through linear fusion of channel dimensions. However, its ability to express high-dimensional nonlinear features is limited. The first extraction module enhances high-dimensional nonlinear features through element-wise multiplication using C3K2_Star, and is particularly effective at capturing subtle textures of small targets (such as text and symbols on traffic signs). Thus, in the YOLOv11-based detection model of the present invention, the first and second extraction models are alternately set, forming a collaborative mechanism of "basic feature extraction - high-dimensional feature enhancement". This balances computational costs while improving the hierarchical nature of feature expression, and is particularly suitable for the multi-scale feature capture needs of small target traffic signs.

[0051] For example, the number of times is 4, combined with Figure 6 As shown, the backbone network consists of 11 layers: the first layer (1) is a convolutional layer, the second layer (2) is a second extraction module, the third layer (3) is a first extraction module, the fourth layer (4) is a second extraction module, the fifth layer (5) is a first extraction module, the sixth layer (6) is a second extraction module, the seventh layer (7) is a first extraction module, the eighth layer (8) is a second extraction module, the ninth layer (9) is a first extraction module, the tenth layer (10) is a pooling layer, and the eleventh layer is a convolutional layer.

[0052] For example, the first layer is Conv, which is the most basic and core feature extraction operation in a Convolutional Neural Network (CNN). The tenth layer is SPPF (Fast Spatial Pyramid Pooling), the core of which is "serialized repetitive pooling instead of parallel multi-scale pooling," which effectively achieves receptive fields of different sizes by reusing pooling kernels of the same size multiple times. The eleventh layer is C2PSA (Cross-Channel Position-aware Spatial Attention).

[0053] In detail, the image to be detected undergoes basic feature extraction through the convolutional layer of layer 1. The input of layer 2 is the output of layer 1. Layer 2 performs basic feature extraction on the output of layer 1. Layer 3 performs feature extraction in multiple ways on the output of layer 3 and performs element-wise multiplication on the features from these multiple ways. Layer 4 performs basic feature extraction on the output of layer 3. Layer 5 performs feature extraction in multiple ways on the output of layer 4 and performs element-wise multiplication on the features from these multiple ways. Layer 6 performs basic feature extraction on the output of layer 5. Layer 7 performs feature extraction in multiple ways on the output of layer 6 and performs element-wise multiplication on the features from these multiple ways. Layer 8 performs basic feature extraction on the output of layer 7. Layer 9 performs feature extraction in multiple ways on the output of layer 8 and performs element-wise multiplication on the features from these multiple ways. Layer 10 performs pooling on the output of layer 9. Layer 11 performs attention processing on the output of layer 10.

[0054] In some embodiments, the C3K2_Star structure includes: a Star_Block structure; the Star_Block structure includes multiple convolutional layers and an element-wise multiplication module; a portion of the convolutional layers are configured as a two-way parallel structure; and another portion of the convolutional layers form a serial structure. The first extraction module, which includes the C3K2_Star structure, performs feature extraction on the basic feature map output by the second extraction module using various methods. Then, it performs element-wise multiplication on these features to obtain multiple feature maps. Figure 7 As shown, it includes: S710: The basic feature map output by the second extraction module is extracted using a two-way parallel structure to obtain two feature maps; S720: Performs element-wise multiplication on the feature maps of the two paths through the element-wise multiplication module; S730: The processed features are extracted through a serial structure, and the extracted features are compared with the basic feature map output by the second extraction module to obtain multiple feature maps.

[0055] For example, in combination Figure 8 As shown, a convolutional layer 81 (DWConv) can also be set between the two parallel structures and the second extraction module to extract features, for example, DWConv7. 7. DWConv (Depthwise Convolution) is a lightweight operation that breaks down ordinary convolution into "Depthwise Convolution" and "Pointwise Convolution." DWConv7×7 specifically refers to the depthwise convolution component using a 7×7 kernel. Through DWConv7... 7. Feature extraction is performed, and the output features are fed into two parallel structures. These two parallel structures perform feature extraction in two different ways. One parallel structure consists of a convolutional layer 82 (Conv2d) and an activation function 83 (ReLU6). The convolutional layer 82 is a Conv2d1... 1, Conv2d1 1 represents a 1×1 convolution, with ReLU6 as the activation function, meaning the number of channels is tripled before applying the ReLU6 activation function. One parallel path consists of 84 convolutional layers (Conv2d), for example, Conv2d1. 1. Increase the number of channels to three times the original number, and multiply the features output from one parallel structure element-wise with the features output from another parallel structure. The serial structure includes convolutional layer 85 (Conv2d) and convolutional layer 86 (DWConv). For example, convolutional layer 85 is Conv2d1. 1. Convolutional layer 86 is DWConv7×7. The feature map obtained by element-wise multiplication is input into the serial structure. The Conv2d1×1 convolutional layer of the serial structure is used to restore the original number of channels. Then, the convolutional layer 86DWConv7×7 of the serial structure is used again for processing. The processed output is then compared with the original input, i.e., convolutional layer 81DWConv7. The output of 7 is used for residual connection.

[0056] This invention proposes to replace the original BottleNeck structure with the Star_Block module, aiming to reduce the computational cost of the model while enhancing its feature representation capabilities. Star_Block is a module based on StarNet, and its structure is as follows: Figure 8As shown, the Star_Block module maps input data to a higher-dimensional nonlinear feature space. This means it can capture more complex and subtle feature combinations, thereby improving the model's object detection accuracy in complex scenes. Its working principle is not simply increasing the network width (i.e., the number of channels), but rather introducing a mechanism similar to a multinomial kernel function. This module achieves second-order nonlinear combinations of features by multiplying different feature channels in pairs, effectively simulating the interaction of high-dimensional features within a compact computational space. This characteristic allows Star_Block to achieve powerful feature representation capabilities while maintaining computational efficiency, solving the problem that traditional structures often experience a dramatic increase in computational cost when improving performance.

[0057] In some embodiments, combined with Figure 9 As shown, this embodiment of the invention provides a C3K2_Star structure that includes a Star_Block structure. The C3K2_Star structure includes six layers. The first layer 91 is Conv, and the second layer 92 is Split. Split (feature splitting) refers to the operation of dividing the input feature map into multiple sub-feature maps according to preset rules (such as channel ratio, spatial region).

[0058] The third layer (93) and fourth layer (94) of the C3K2_Star structure are C3K-Star, combined with... Figure 10 As shown, C3K-Star consists of four layers. The first layer 101 is Conv, the second layer 102 is a parallel structure with one channel being Star_Block and the other being Conv, the third layer 103 is Concat (channel splicing), which refers to the operation of splicing and fusing multiple feature maps of the same spatial size (height × width) in the channel dimension, and the fourth layer is Conv104.

[0059] The fifth layer 95 of the C3K2_Star structure is Concat, which concatenates the outputs of the second layer 92, the third layer 93, and the fourth layer 94. The sixth layer 96 of the C3K2_Star structure is Conv, which inputs the features output from the fifth layer 95 into the sixth layer 96 for feature extraction.

[0060] In some embodiments, the neck network further includes multiple splicing modules and multiple third extraction modules; the multiple third extraction modules and multiple first extraction modules containing the C3K2_Star structure form multiple two-way parallel structures; one path of the two-way parallel structure contains the third extraction module, and the other path of the parallel structure contains the first extraction module of the C3K2_Star structure.

[0061] Multiple feature extraction modules containing the C3K2_Star structure are used to extract features from multiple feature maps using various methods. The features from these methods are then multiplied element-wise. Finally, the processed feature maps are fused to obtain multiple fused feature maps, including: Execute the following process multiple times, combined with Figure 11 As shown: S1110: The third extraction module of the input feature map is used to extract features through a two-way parallel structure with one path. S1120: The first extraction module, which has a C3K2_Star structure in one of the two parallel paths, performs feature extraction on the basic feature map output by the second extraction module in multiple ways, and performs element-wise multiplication on the features in multiple ways; S1130: The outputs of the two parallel structures are spliced ​​together by the splicing module to obtain a fused feature map.

[0062] The third extraction module includes convolutional layers or upsampling.

[0063] Combination Figure 6 As shown, the neck network includes layers 12 to 23. Layer 12 is Upsample, which refers to the operation of upsampling a low-resolution feature map (small size) to a high-resolution (large size) using a specific algorithm. Layer 13 is Concat. Layer 14 is the first extraction module containing the C3K2_Star structure. Layer 15 is Upsample. Layer 16 is Concat. Layer 17 is the first extraction module containing the C3K2_Star structure. Layer 18 is Conv. Layer 19 is Concat. Layer 20 is the first extraction module containing the C3K2_Star structure. Layer 21 is Conv. Layer 22 is Concat. Layer 23 is the first extraction module containing the C3K2_Star structure.

[0064] The concatenation module of the thirteenth layer Concat receives two parallel inputs, one of which is the twelfth layer Upsample, and the other is the output feature of the seventh layer in the backbone network. That is, the thirteenth layer Concat receives features from the seventh and twelfth layers and performs feature fusion according to the channel dimension.

[0065] The sixteenth-layer Concat's splicing module takes two parallel inputs, one of which is from the twelfth, thirteenth, fourteenth, and fifteenth layers, and the other is from the features output from the fourth layer of the backbone network. In other words, the sixteenth-layer Concat receives features from the fourth and fifteenth layers and performs feature fusion according to the channel dimension.

[0066] The concatenation module of the nineteenth layer Concat takes two parallel inputs, one of which is layers twelve, thirteen, and fourteen, and the other is layers twelve, thirteen, fourteen, fifteen, sixteen, seventeen, and eighteen. That is, the nineteenth layer Concat receives features from layers fourteen and eighteen and performs feature fusion according to the channel dimension.

[0067] The concatenation module of layer 22, Concat, takes into account two parallel inputs: one from layer 11 of the backbone network, and the other from layers 12 through 21. In other words, layer 22's Concat receives features from layers 11 and 21 and fuses them according to the channel dimension.

[0068] The output of the seventeenth layer in the neck network is input to the twenty-sixth layer 26 of the detection head network, the output of the twentieth layer is input to the twenty-fifth layer 25 of the detection head network, and the output of the twenty-third layer is input to the twenty-fourth layer 24 of the detection network.

[0069] In some embodiments, the outputs of different layers in the neck network are feature maps of multiple scales, and the feature maps of multiple scales are respectively input into the detection module containing the DESCDH structure of the detection head network that processes the corresponding scale. The DESCDH structure includes multiple shared convolutional layers, a recognition convolutional layer, and multiple output modules; Step 420-3 involves performing differential convolution processing on multiple fused feature maps using multiple detection modules, and then using the processed fused feature maps to identify traffic signs to obtain the detection results. The specific implementation method is as follows: combining... Figure 12 As shown: S1210: Differential convolution processing is performed on the input feature map through multiple shared convolutional layers; S1220: By recognizing convolutional layers, convolution processing is performed on the feature maps output by multiple shared convolutional layers to identify the features corresponding to traffic signs and the features corresponding to the location of traffic signs; S1230: For each scale of feature map, the output module of the DESCDH structure that processes the feature map of that scale is used to identify traffic signs and their location by processing the output of the recognition convolutional layer, and output the traffic signs and their location.

[0070] Optionally, the DESCDH structure also includes multiple group normalized convolutional layers before multiple shared convolutional layers. The recognition convolutional layers include classification convolutional layers and regression convolutional layers. The multiple output modules include multiple scaling modules, multiple classification output modules, and multiple localization output modules. That is, the DESCDH structure can also include multiple group normalized convolutional layers, multiple shared convolutional layers, classification convolutional layers, regression convolutional layers, multiple scaling modules, multiple classification output modules, and multiple localization output modules. For each scale of feature map, group normalization convolution processing is performed on the feature map of that scale by processing the group normalization convolution layer of the DESCDH structure. By processing the feature maps at this scale through shared convolutional layers of the DESCDH structure, differential convolution processing is performed on the feature maps output by the group normalized convolutional layers. By processing the feature map at this scale using a DESCDH structure classification convolutional layer, classification convolution processing is performed on the feature map output by the shared convolutional layer. By processing the feature map at this scale using a DESCDH structure of a regression convolutional layer, regression convolution processing is performed on the feature map output by the shared convolutional layer. The scaling module of the DESCDH structure, which processes the feature map at this scale, performs scale scaling operations on the outputs of the classification convolutional layer and the regression convolutional layer. By processing the DESCDH structure classification output module of the feature map at this scale, the output of the scaling module is used to identify traffic signs, and the traffic sign identifiers in the detection results are obtained. By processing the DESCDH structure of the feature map at this scale, the localization output module performs traffic sign localization on the output of the scaling module, thus obtaining the position of the traffic sign in the detection result.

[0071] For example, in combination Figure 13 As shown, there are three scale types. The DESCDH structure includes three group normalized convolutional layers, two shared convolutional layers, one classification convolutional layer, one regression convolutional layer, six scaling modules, three classification output modules, and three localization output modules. The same two shared convolutional layers, classification convolutional layers, and regression convolutional layers are used for processing feature maps at all three scales. However, different group normalized convolutional layers, different scaling modules, different classification output modules, and different localization output modules are used for processing feature maps at all three scales.

[0072] For the feature map P3 at the first scale, feature extraction is performed using a group normalized convolutional layer 131 (Conv_GN, composed of Conv3×3+GN+SiLU). The output of the group normalized convolutional layer 131 serves as the input to a shared convolutional layer 132 (DEConv). The output of the shared convolutional layer 132 serves as the input to a classification convolutional layer 133. The output of the classification convolutional layer 133 (Conv_Cls) serves as the input to a scaling module 134 (Scale). The output of the scaling module 134 (Scale) serves as the input to a classification output module 135, which outputs the traffic sign identifier. The output of the shared convolutional layer 132 serves as the input to a regression convolutional layer 136 (Conv_Reg). The output of the regression convolutional layer 136 (Conv_Reg) serves as the input to a scaling module 137 (Scale). The output of the scaling module 137 (Scale) serves as the input to a localization output module 138, which outputs the location of the traffic sign.

[0073] For the feature map P4 at the second scale, feature extraction is performed using a group normalized convolutional layer 141 (Conv_GN, composed of Conv3×3+GN+SiLU). The output of the group normalized convolutional layer 141 serves as the input to a shared convolutional layer 132 (DEConv). The output of the shared convolutional layer 132 serves as the input to a classification convolutional layer 133 (Conv_Cls). The output of the classification convolutional layer 133 (Conv_Cls) serves as the input to a scaling module 144 (Scale). The output of the scaling module 144 (Scale) serves as the input to a classification output module 145, which outputs the traffic sign identifier. The output of the shared convolutional layer 132 serves as the input to a regression convolutional layer 136 (Conv_Reg). The output of the regression convolutional layer 136 (Conv_Reg) serves as the input to a scaling module 147 (Scale). The output of the scaling module 147 (Scale) serves as the input to a localization output module 148, which outputs the location of the traffic sign.

[0074] For the feature map P5 at the third scale, feature extraction is performed using a group normalized convolutional layer 151 (Conv_GN, composed of Conv3×3+GN+SiLU). The output of the group normalized convolutional layer 151 serves as the input to a shared convolutional layer 132 (DEConv). The output of the shared convolutional layer 132 serves as the input to a classification convolutional layer 133. The output of the classification convolutional layer 133 (Conv_Cls) serves as the input to a scaling module 154 (Scale). The output of the scaling module 154 (Scale) serves as the input to a classification output module 155, which outputs the traffic sign identifier. The output of the shared convolutional layer 132 serves as the input to a regression convolutional layer 136 (Conv_Reg). The output of the regression convolutional layer 136 (Conv_Reg) serves as the input to a scaling module 157 (Scale). The output of the scaling module 157 (Scale) serves as the input to a localization output module 158, which outputs the location of the traffic sign.

[0075] In some embodiments, the shared convolutional layer includes multiple differential convolutional layers, an addition module, a regular convolutional layer, and an equivalent convolutional layer; The inputs of multiple differential convolutional layers and ordinary convolutional layers are the outputs of multiple group normalized convolutional layers; the outputs of multiple differential convolutional layers and ordinary convolutional layers are weight values ​​and biases. The input to the equivalent convolutional layer is the sum of the weight values ​​and the bias.

[0076] The specific implementation process of performing differential convolution processing on the input feature map through multiple shared convolutional layers is as follows: combining... Figure 14 As shown: S1410: The input feature map is processed by differential convolution and basic convolution through multiple differential convolution layers and ordinary convolution layers to obtain weight values ​​and biases. S1420: The weight value and bias are added together by the addition module; S1430: The output of the addition module is convolved by an equivalent convolutional layer.

[0077] Furthermore, the shared convolutional layer also includes group normalization and an activation function, wherein the input of the group normalization is the output of the equivalent convolutional layer; the input of the activation function is the output of the group normalization; and the output of the activation function is the output of the shared convolutional layer.

[0078] For example, in combination Figure 15 As shown, the differential convolutional layer 151 in the shared convolutional layer includes Conv_cd3 3. Conv_hd3 3. Conv_vd3 3. Conv_ad3 3. Among them, Conv_cd3 3. Conv_hd3 3. Conv_vd3 3. Conv_ad3 The core logic of 3 is to enhance the feature response in a specific dimension through direction-specific differential shifting: Conv_cd3×3: cd=Central Differential, which focuses on the detailed features of the target's central region and enhances the extraction of core semantic information; Conv_hd3×3: hd=Horizontal Differential, which focuses on capturing horizontal features (such as the horizontal lines of text in traffic signs and horizontal edges); Conv_vd3×3: vd=Vertical Differential, which focuses on capturing vertical features (such as the vertical lines and borders of traffic signs). Conv_ad3×3: ad=All-directional Differential, which integrates differences in multiple directions such as horizontal, vertical, and diagonal lines to adapt to complex shape features (such as the outlines of triangular and circular traffic signs).

[0079] A typical convolutional layer is Conv3. 3; The equivalent convolutional layer 152 is Conv2d3 3; Group normalized layer 153, activation function 154 is SiLU.

[0080] Combination Figure 6 and Figure 13 As shown, the inputs at layers 17, 20, and 23 are processed using independent Conv_GN modules (composed of Conv3×3+GN+SiLU). Then, the outputs at the three scales are processed by a shared convolutional module consisting of two DEConv modules. In DEConv, the input is processed by four differential convolutions and one regular convolutional kernel to obtain corresponding weights and biases. These five weights and biases are summed and applied to an equivalent convolutional kernel, then processed through GN and SiLU activation functions to obtain the final output of DEConv. Finally, the outputs of the shared convolutional module are fed into Conv_Cls and Conv_Reg for classification and localization tasks, respectively, and scaled using independent Scale modules to obtain classification and localization outputs at different scales.

[0081] The DEConv structure provided by this invention is as follows: Figure 15As shown, DEConv uses four types of difference convolutions plus one ordinary convolution, which means it considers not only the pixel's own value but also the difference (gradient) between the pixel and its surrounding pixels during computation. This makes DEConv very adept at capturing detailed information such as the edges, contours, and textures of objects. Since using five parallel convolutions for feature extraction inevitably increases the number of parameters and inference time, DEConv is designed to use convolution kernels with the same kernel size, stride, and padding size to operate on the same input, calculate the corresponding weights and biases, and then sum all the weights and biases to generate an equivalent convolution kernel. This equivalent convolution kernel produces the final output, calculated using the following formula:

[0082] in, denoted by SiLU activation function, and GN denotes group normalization. Indicates input, This represents the convolution kernel parameters.

[0083] In some embodiments, the method further includes: Obtain a training set; wherein, the actual result of the images in the training set is the identification of traffic signs in the images and their positions in the images; For each training process, the images in the training set are input into the base model, and the base model outputs the detection results. Based on the detection results and the actual results, the optimizer updates the parameters of the base model, and the distributed focus loss function is used to update the parameters in the DESCDH structure of the base model. After training, the base model with the last updated parameters is used as the detection model.

[0084] For example, the CCTSDB2021 traffic sign dataset was selected, which contains 17,856 images and annotation information, divided into training, validation, and test sets in a 7:2:1 ratio. The YOLOv11-Star-DESCDH model (i.e., the base model) was trained using this dataset. The training parameters were: 300 epochs, 16 batch size, 0.01 initial learning rate, AdamW optimizer, and a patience parameter of 50.

[0085] The DESCDH architecture employs Distribution Focal Loss (DFL) regression, which improves localization accuracy by predicting the probability distribution of bounding boxes. It decouples classification and regression tasks into independent branches and uses a learnable scale module to balance these two tasks. Furthermore, the detection head separates the training and inference paths, integrating necessary decoding steps and specialized model derivation optimizations during inference to ensure high efficiency during deployment.

[0086] like Figure 16 As shown, the present invention also provides a traffic sign detection device based on YOLOv11, comprising: Module 1610 is used to build a detection model based on YOLOv11; wherein, the detection model includes multiple first extraction modules containing C3K2_Star structures and multiple detection modules containing DESCDH structures; Processing module 1620 is used to input the image to be detected into the detection model, and the detection model outputs the detection result; The operation of the detection model includes: Multiple first extraction modules are used to extract features from the image to be detected in various ways, and the features from these multiple methods are multiplied element-wise to obtain multiple feature maps. Multiple feature maps are extracted using multiple first extraction modules in various ways, and the features from these multiple methods are multiplied element-wise. The processed feature maps are then fused to obtain multiple fused feature maps. Multiple detection modules perform differential convolution processing on multiple fused feature maps, and then use the processed fused feature maps to identify traffic signs, thus obtaining the detection results.

[0087] Optionally, the detection model includes a backbone network, a neck network, and a detection head network; wherein, a first extraction module containing a C3K2_Star structure is partially in the backbone network, a first extraction module containing a C3K2_Star structure is partially in the neck network, and multiple detection modules containing a DESCDH structure are in the detection head network.

[0088] Optionally, the backbone network further includes multiple second extraction modules; in the backbone network, multiple second extraction modules and multiple first extraction modules containing the C3K2_Star structure are alternately arranged; Processing module 1620 is specifically used for: Execute the following process multiple times: The second extraction module extracts basic features from the input feature map to obtain multiple basic feature maps. The first extraction module, which includes the C3K2_Star structure, performs feature extraction on the basic feature map output by the second extraction module in multiple ways, and performs element-wise multiplication on the features in multiple ways to obtain multiple feature maps.

[0089] Optionally, the C3K2_Star structure includes: a Star_Block structure; the Star_Block structure includes multiple convolutional layers and element-wise multiplication modules; a portion of the convolutional layers are configured as a two-way parallel structure; and another portion of the convolutional layers form a serial structure. Processing module 1620 is specifically used for: The two parallel structures are used to extract features from the basic feature map output by the second extraction module to obtain two feature maps. The element-wise multiplication module performs element-wise multiplication on the feature maps of the two paths. The processed features are extracted using the serial structure, and the extracted features are then compared with the basic feature map output by the second extraction module to obtain multiple feature maps.

[0090] Optionally, the neck network further includes multiple splicing modules and multiple third extraction modules; the multiple third extraction modules and multiple first extraction modules containing the C3K2_Star structure form multiple two-way parallel structures; one path of the two-way parallel structure contains the third extraction module, and the other path of the parallel structure contains the first extraction module of the C3K2_Star structure; Processing module 1620 is specifically used for: Execute the following process multiple times: The third extraction module of one path in the two-way parallel structure performs feature extraction on the input feature map; The first extraction module, which contains a C3K2_Star structure in the other path of the two-way parallel structure, performs feature extraction on the basic feature map output by the second extraction module in multiple ways, and performs element-wise multiplication on the features in multiple ways; The splicing module splices the outputs of the two parallel structures to obtain a fused feature map. The third extraction module includes a convolutional layer or upsampling.

[0091] Optionally, the outputs of different layers in the neck network are feature maps of multiple scales, and the feature maps of multiple scales are respectively input into the detection module containing the DESCDH structure of the detection head network that processes the corresponding scale. The DESCDH structure includes multiple shared convolutional layers, a recognition convolutional layer, and multiple output modules; Processing module 1620 is specifically used for: The input feature map is processed by differential convolution through multiple shared convolutional layers; The recognition convolutional layer is used to perform convolution processing on the feature maps output by multiple shared convolutional layers to identify the features corresponding to traffic signs and the features corresponding to the location of traffic signs. For each scale of feature map, the output module of the DESCDH structure that processes the feature map of that scale is used to identify traffic signs and their location by processing the output of the recognition convolutional layer, and output the traffic signs and their location.

[0092] Optionally, the shared convolutional layer includes multiple differential convolutional layers, an addition module, a regular convolutional layer, and an equivalent convolutional layer; Processing module 1620 is specifically used for: The input feature map is processed by differential convolution and basic convolution through multiple differential convolution layers and the ordinary convolution layer to obtain weight values ​​and biases. The weight value and the bias are added together by the addition module; The output of the addition module is convolved by the equivalent convolutional layer.

[0093] Optionally, the device further includes: A training unit is used to acquire a training set; wherein, the actual result of the images in the training set is the identification of traffic signs in the images and their positions in the images; In each training process, the images in the training set are input into the base model, and the base model outputs the detection results. Based on the detection results and the actual results, the parameters of the base model are updated using an optimizer, and the parameters in the DESCDH structure of the base model are updated using the distributed focus loss function. After training, the base model with the last updated parameters is used as the detection model.

[0094] In addition, combined Figures 1-16 The YOLOv11-based traffic sign detection method and apparatus described in this embodiment of the invention can be implemented by electronic devices.

[0095] Electronic devices, including: processors; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the YOLOv11-based traffic sign detection method as described above.

[0096] Based on the above introduction, an example is proposed. Figure 17 The structure of electronic devices.

[0097] The electronic device may include a processor 1710 and a memory 1720 storing computer program instructions.

[0098] Specifically, the processor 1710 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0099] Memory 1720 may include mass storage for data or instructions. For example, and not limitingly, memory 1720 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 1720 may include removable or non-removable (or fixed) media. Where appropriate, memory 1720 may be internal or external to a data processing device. In a particular embodiment, memory 1720 is a non-volatile solid-state memory. In a particular embodiment, memory 1720 includes read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0100] The processor 1710 reads and executes computer program instructions stored in the memory 1720 to implement any of the YOLOv11-based traffic sign detection methods in the above embodiments.

[0101] In one example, the electronic device may also include a communication interface 1730 and a bus 1740. For example, Figure 17 As shown, the processor 1710, memory 1720, and communication interface 1730 are connected through bus 1740 and complete communication with each other.

[0102] The communication interface 1730 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of the present invention.

[0103] Bus 1740 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 1740 may include one or more buses. While specific buses are described and illustrated in embodiments of the invention, the invention contemplates any suitable bus or interconnect.

[0104] The electronic device can execute the YOLOv11-based traffic sign detection method in this embodiment of the invention based on the received task, thereby achieving a combination of... Figures 1-16 The method and apparatus for detecting traffic signs based on YOLOv11 are described.

[0105] In addition, in conjunction with the electronic devices in the above embodiments, the present invention can provide a storage medium that, when the instructions in the storage medium are executed by the processor of the electronic device, enables the electronic device to perform the YOLOv11-based traffic sign detection method as described in any of the above embodiments.

[0106] 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.

[0107] 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 1The function specified in one or more boxes.

[0108] 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 Figure 1 The steps of the function specified in one or more boxes.

[0109] 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 the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0110] 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 detecting traffic signs based on YOLOv11, characterized in that, include: A detection model is constructed based on YOLOv11; wherein, the detection model includes multiple first extraction modules containing C3K2_Star structures and multiple detection modules containing DESCDH structures; The image to be detected is input into the detection model, and the detection model outputs the detection result; The operation of the detection model includes: Multiple first extraction modules are used to extract features from the image to be detected in various ways, and the features from these multiple methods are multiplied element-wise to obtain multiple feature maps. Multiple feature maps are extracted using multiple first extraction modules in various ways, and the features from these multiple methods are multiplied element-wise. The processed feature maps are then fused to obtain multiple fused feature maps. Multiple detection modules perform differential convolution processing on multiple fused feature maps, and then use the processed fused feature maps to identify traffic signs, thus obtaining the detection results.

2. The traffic sign detection method based on YOLOv11 according to claim 1, characterized in that, The detection model includes a backbone network, a neck network, and a detection head network; wherein, a first extraction module containing a C3K2_Star structure is partially in the backbone network, a first extraction module containing a C3K2_Star structure is partially in the neck network, and multiple detection modules containing a DESCDH structure are in the detection head network.

3. The traffic sign detection method based on YOLOv11 according to claim 1 or 2, characterized in that, The backbone network also includes multiple second extraction modules; in the backbone network, multiple second extraction modules and multiple first extraction modules containing C3K2_Star structures are alternately arranged; Multiple first extraction modules perform feature extraction on the image to be detected using various methods, and then perform element-wise multiplication of these features to obtain multiple feature maps, including: Execute the following process multiple times: The second extraction module extracts basic features from the input feature map to obtain multiple basic feature maps. The first extraction module, which includes the C3K2_Star structure, performs feature extraction on the basic feature map output by the second extraction module in multiple ways, and performs element-wise multiplication on the features in multiple ways to obtain multiple feature maps.

4. The traffic sign detection method based on YOLOv11 according to claim 3, characterized in that, The C3K2_Star structure includes a Star_Block structure; the Star_Block structure includes multiple convolutional layers and element-wise multiplication modules; some convolutional layers are configured as a two-way parallel structure; the other part of the convolutional layers form a serial structure. The first extraction module, which includes the C3K2_Star structure, performs feature extraction on the basic feature map output by the second extraction module using various methods. Then, it performs element-wise multiplication on the features obtained from these methods to produce multiple feature maps, including: The two parallel structures are used to extract features from the basic feature map output by the second extraction module to obtain two feature maps. The element-wise multiplication module performs element-wise multiplication on the feature maps of the two paths. The processed features are extracted using the serial structure, and the extracted features are then compared with the basic feature map output by the second extraction module to obtain multiple feature maps.

5. The traffic sign detection method based on YOLOv11 according to claim 4, characterized in that, The neck network also includes multiple splicing modules and multiple third extraction modules; the multiple third extraction modules and multiple first extraction modules containing the C3K2_Star structure form multiple two-way parallel structures; one path of the two-way parallel structure contains the third extraction module, and the other path of the parallel structure contains the first extraction module of the C3K2_Star structure. Multiple feature maps are extracted using various methods through multiple first extraction modules. The features from these methods are then multiplied element-wise. The processed feature maps are then fused to obtain multiple fused feature maps, including: Execute the following process multiple times: The third extraction module of one path in the two-way parallel structure performs feature extraction on the input feature map; The first extraction module, which contains a C3K2_Star structure in the other path of the two-way parallel structure, performs feature extraction on the basic feature map output by the second extraction module in multiple ways, and performs element-wise multiplication on the features in multiple ways; The splicing module splices the outputs of the two parallel structures to obtain a fused feature map. The third extraction module includes a convolutional layer or upsampling.

6. The traffic sign detection method based on YOLOv11 according to claim 2, characterized in that, The different layers of the neck network output feature maps of multiple scales, and these feature maps of multiple scales are respectively input into the detection module containing the DESCDH structure to process the corresponding scale. The DESCDH structure includes multiple shared convolutional layers, a recognition convolutional layer, and multiple output modules; Multiple fused feature maps are processed by differential convolution using multiple detection modules, and traffic sign recognition is performed on the processed fused feature maps to obtain detection results, including: The input feature map is processed by differential convolution through multiple shared convolutional layers; The recognition convolutional layer is used to perform convolution processing on the feature maps output by multiple shared convolutional layers to identify the features corresponding to traffic signs and the features corresponding to the location of traffic signs. For each scale of feature map, the output module of the DESCDH structure that processes the feature map of that scale is used to identify traffic signs and their location by processing the output of the recognition convolutional layer, and output the traffic signs and their location.

7. The traffic sign detection method based on YOLOv11 according to claim 6, characterized in that, The shared convolutional layer includes multiple differential convolutional layers, an addition module, a regular convolutional layer, and an equivalent convolutional layer; The input feature map is processed by differential convolution through multiple shared convolutional layers, including: The input feature map is processed by differential convolution and basic convolution through multiple differential convolution layers and the ordinary convolution layer to obtain weight values ​​and biases. The weight value and the bias are added together by the addition module; The output of the addition module is convolved by the equivalent convolutional layer.

8. The traffic sign detection method based on YOLOv11 according to claim 1, characterized in that, The method further includes: Obtain a training set; wherein, the actual result of the images in the training set is the identification of traffic signs in the images and their positions in the images; In each training process, the images in the training set are input into the base model, and the base model outputs the detection results. Based on the detection results and the actual results, the parameters of the base model are updated using an optimizer, and the parameters in the DESCDH structure of the base model are updated using the distributed focus loss function. After training, the base model with the last updated parameters is used as the detection model.

9. A traffic sign detection device based on YOLOv11, characterized in that, include: A building module is used to build a detection model based on YOLOv11; wherein, the detection model includes multiple first extraction modules containing C3K2_Star structures and multiple detection modules containing DESCDH structures; The processing module is used to input the image to be detected into the detection model, and the detection model outputs the detection result; The operation of the detection model includes: Multiple first extraction modules are used to extract features from the image to be detected in various ways, and the features from these multiple methods are multiplied element-wise to obtain multiple feature maps. Multiple feature maps are extracted using multiple first extraction modules in various ways, and the features from these multiple methods are multiplied element-wise. The processed feature maps are then fused to obtain multiple fused feature maps. Multiple detection modules perform differential convolution processing on multiple fused feature maps, and then use the processed fused feature maps to identify traffic signs, thus obtaining the detection results.

10. An electronic device, characterized in that, include: Memory is used to store computer programs or instructions; A processor for executing a computer program or instructions in the memory such that the YOLOv11-based traffic sign detection method as described in any one of claims 1-8 is executed.

Images