Multi-target detection method and device based on improved YOLOV3 algorithm, and storage medium
By using the GhostNet network and Soft-NMS algorithm, combined with the SPP module, the YOLOv3 algorithm was improved, solving the problems of insufficient detection speed and real-time performance of YOLOv3, and achieving more efficient multi-target detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-05-01
AI Technical Summary
The existing YOLOv3 target detection algorithm has shortcomings in terms of detection speed and real-time performance. In particular, the detection speed is not ideal on devices with weak computing power, and there is a problem of missing detection of highly overlapping targets.
The GhostNet network is used as the backbone feature extraction network. It is combined with the Soft-NMS algorithm and the Spatial Pyramid Pooling module SPP to replace the traditional Non-Maximum Suppression algorithm NMS, thereby reducing the computational load and improving the detection accuracy.
While ensuring detection accuracy, it significantly improves detection speed and real-time performance, reduces the false negative rate of highly overlapping targets, and enhances the effect of multi-target detection.
Smart Images
Figure CN121962592A_ABST
Abstract
Description
A method, apparatus, and storage medium for multi-target detection based on an improved YOLOv3 algorithm Technical Field
[0001] This invention relates to the field of multi-target detection technology, and in particular to a multi-target detection method, apparatus and storage medium based on an improved YPLOV3 algorithm. Background Technology
[0002] With scientific advancements and the continuous development of artificial intelligence, multi-target tracking technology has also seen rapid development in fields such as surveillance. However, current multi-target tracking technologies still face challenges such as target uncertainty and occlusion. The YOLOv3 object detection algorithm holds a crucial position within the YOLO series of object detection algorithms, employing novel methods to improve the detection speed and accuracy of the YOLO series. However, while the YOLOv3 object detection algorithm demonstrates excellent performance in detection speed and accuracy, it places high demands on computational resources. Due to this high computational cost and inefficiency, the YOLOv3 object detection algorithm's real-time performance is unsatisfactory in practical applications, especially on devices with limited computing power, where its detection speed is less than ideal. The YOLOv3 object detection algorithm uses traditional non-maximum suppression to remove redundant predicted boxes, but this also leads to missed detections of many highly overlapping targets. To address the shortcomings of YOLOv3, such as high computational cost and slow speed, YOLOv3-tiny was later proposed. YOLOv3-tiny has a very fast detection speed, which well meets the requirements of real-time detection, but its detection accuracy is significantly lower than that of YOLOv3.
[0003] Therefore, how to improve detection speed and real-time performance while ensuring detection accuracy has become one of the urgent existing technical problems to be solved. Summary of the Invention
[0004] This invention provides a multi-target detection method, apparatus, and storage medium based on an improved YPLOV3 algorithm, which improves detection speed and real-time performance while ensuring detection accuracy.
[0005] Firstly, a multi-target detection method based on an improved YPLOV3 algorithm is provided, including:
[0006] Based on the received image, features are extracted using a feature extraction network with GhostNet as the backbone to obtain the corresponding feature map.
[0007] Based on the obtained feature map, the feature is mapped to a feature vector of S×S×(B×5+C), where S×S is the number of grids in the image segmentation, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected.
[0008] The Soft-Number Suppression algorithm (Soft-NMS) is used to select suitable bounding boxes based on the set confidence threshold and Intersection over Union (IOU) value, and then multi-target detection is performed using these suitable bounding boxes.
[0009] In one implementation, the feature map is generated by using a linear transformation based on a set of original feature maps to generate a Ghost feature map at the lowest possible cost.
[0010] Suppose we have a set of original feature maps Y∈R h*w*c Where h and w are the height and width of the space, and m is the number of channels less than c, a feature map Y'∈R with m channels is generated through convolution. h*w*m For the feature maps of the remaining cm channels, a linear transformation Ф is performed on the feature map in Y'. i,j (y i The feature maps are generated using the following method: i = 1, ..., m; j = 1, ..., s. The two feature maps are then combined.
[0011] In one implementation, the GhostNet network is composed of stacked Ghost feature map modules.
[0012] In one implementation, an SPP module is added after each output of the feature extraction network, for a total of 3 SPP modules.
[0013] In one implementation, the SPP module consists of 5x5, 9x9, 13x13 max pooling and a jump connection.
[0014] In one implementation, a suitable prediction box is obtained by using the Soft Non-Maximum Suppression (Soft-NMS) algorithm based on a set confidence threshold and Intersection over Union (IOU) value. Specifically, this includes:
[0015] Based on the candidate box set B and its corresponding confidence score set, obtain the highest confidence score m and the corresponding candidate box bm, delete bm from B and put it into the filtered candidate box set D;
[0016] Iterate through all elements in set B. If the Intersection over Union (IOU) value is lower than the Ni value, add it to set D. If the IOU value is not lower than the Ni value, introduce the corresponding decay function si = si * f[iou(M,bi)]. If the confidence level after introducing the decay function is greater than the probability threshold δ, add it to set D. Finally, output the filtered set.
[0017] Secondly, a multi-target detection system based on the improved YPLOV3 algorithm is provided, including:
[0018] The feature map acquisition module is used to extract features from the received image using a feature extraction network with GhostNet as the backbone, and obtain the corresponding feature map.
[0019] The mapping module is used to map the features into a feature vector of S×S×(B×5+C) based on the obtained feature map, where S×S is the number of grids in the image segmentation, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected.
[0020] The detection module is used to select suitable prediction boxes based on the set confidence threshold and intersection-union ratio (IOU) value using the Soft-NMS algorithm, and then perform multi-target detection using the suitable prediction boxes.
[0021] In one implementation, the detection module is specifically used to obtain the highest confidence score m and the corresponding candidate box bm based on the candidate box set B and its corresponding confidence score set, and to delete bm from B and put it into the filtered candidate box set D.
[0022] Iterate through all elements in set B. If the Intersection over Union (IOU) value is lower than the Ni value, add it to set D. If the IOU value is not lower than the Ni value, introduce the corresponding decay function si = si * f[iou(M,bi)]. If the confidence level after introducing the decay function is greater than the probability threshold δ, add it to set D. Finally, output the filtered set.
[0023] Thirdly, a computing device is provided, comprising at least one processor and at least one memory, wherein the memory stores a computer program, and the memory is configured to read the computer program from the memory and execute any of the steps described in the multi-target detection method based on the improved YPLOV3 algorithm provided in the first aspect.
[0024] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing computer-executable instructions for causing a computer to perform any step of the multi-target detection method based on the improved YPLOV3 algorithm provided in the first aspect.
[0025] This invention provides a multi-object detection method, apparatus, and storage medium based on an improved YPLOV3 algorithm. The method includes: extracting features from a received image using a feature extraction network with GhostNet as the backbone to obtain a corresponding feature map; mapping the features to an S×S×(B×5+C) feature vector, where S×S is the number of grids in the image segmentation, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected; using the Soft Non-Maximum Suppression (Soft-NMS) algorithm, suitable prediction boxes are selected based on a set confidence threshold and Intersection over Union (IOU) value; and multi-object detection is performed using the suitable prediction boxes. This method achieves improved detection speed and real-time performance while maintaining detection accuracy.
[0026] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and accompanying drawings. Attached Figure Description
[0027] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0028] Figure 1 is a flowchart of a multi-target detection method based on the improved YPLOV3 algorithm according to an embodiment of the present invention;
[0029] Figure 2 is a target detection network architecture diagram according to an embodiment of the present invention;
[0030] Figure 3 is a Ghost model diagram according to an embodiment of the present invention;
[0031] Figure 4 is an SPP module architecture diagram according to an embodiment of the present invention;
[0032] Figure 5 is a diagram of the PANet network structure according to an embodiment of the present invention. Detailed Implementation
[0033] To improve detection speed and real-time performance while ensuring detection accuracy, a multi-target detection method, device, and storage medium based on an improved YPLOV3 algorithm are provided.
[0034] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Furthermore, the embodiments and features in the embodiments of the present invention can be combined with each other without conflict.
[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of the embodiments of the present invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein.
[0036] As shown in Figure 1, this embodiment provides a multi-target detection method based on the improved YPLOV3 algorithm. The specific implementation steps include:
[0037] S11. Based on the received image, feature extraction is performed using a feature extraction network with GhostNet as the backbone to obtain the corresponding feature map.
[0038] In practice, as shown in Figure 2, the backbone network Darknet53 of YOLOV3P is replaced with the lightweight network GhostNet, which significantly reduces the computational parameters and computational load, and increases the real-time performance of the target detection network.
[0039] Find a set of original feature maps, and then use a linear transformation to generate Ghost feature maps at the lowest possible cost:
[0040] Suppose we have a set of original feature maps Y∈R h*w*c Where h and w are the height and width of the space, and m is the number of channels less than c. A partial feature map Y'∈R with m channels is generated through convolution operations. h*w*m For the remaining cm channels of feature maps, instead of generating them through standard convolution, they are generated by performing a linear transformation Ф on the feature maps in Y'. i,j (y i The two feature maps are generated using methods such as linear combination, translation, and scaling, where i = 1,...,m; j = 1,...,s, and finally the two feature maps are combined.
[0041] As shown in Figure 3, the entire GhostNet is mainly composed of stacked Ghost modules. It is a lightweight network with higher accuracy and fewer model computational parameters and costs. Like depthwise separable convolution modules, Ghost modules effectively reduce the large amount of computation generated by ordinary convolution.
[0042] S12. Based on the obtained feature map, map the features to a feature vector of S×S×(B×5+C), where S×S is the number of grids in the image segmentation, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected.
[0043] S13. Using the Soft-Number Suppression algorithm (Soft-NMS), suitable prediction boxes are selected based on the set confidence threshold and Intersection over Union (IOU) value, and multi-target detection is performed using the suitable prediction boxes.
[0044] The predicted bounding boxes are sorted according to their confidence scores, and those below a set threshold are deleted. Non-maximum suppression is then applied, and highly overlapping predicted bounding boxes are removed based on the set Interchange of Units (IOU) values.
[0045] The purpose of Soft-NMS is to preserve a suitable detection window and remove redundant candidate boxes, so that the algorithm can find the optimal target location.
[0046] The specific algorithm design of Soft-NMS is as follows:
[0047] Improved Algorithm: Soft-NMS (Flexible Nonmaximum Suppression)
[0048] Input: Candidate box set B = {b1, b2, ..., bi}, confidence score set S = {s1, s2, ..., si}, probability threshold
[0049] 1: while:
[0050] 2: m = Max(S) / / Calculate the score with the highest confidence level
[0051] 3: B = B - bm
[0052] 4: D = D + bm / / The highest confidence candidate box bm is placed into D.
[0053] 5: for biin B
[0054] 6: if iou(M, bi) < Ni then
[0055] 7: D = D + bi
[0056] 8: if iou(M, bi)≥N then
[0057] 9: si=si*f[iou(M,bi)] / / Introduce the corresponding decay function
[0058] 10: and if si≥δ / / If the confidence level after introducing the decay function is greater than the probability threshold
[0059] 11: D = D + bi / / Then put it into D
[0060] 12: return D / / Output the filtered set
[0061] In convolutional neural networks (CNNs), classification typically occurs in the final fully connected layer. The number of features in a fully connected layer is fixed, which also necessitates a fixed input image size. However, it's often difficult to maintain a consistent input image size. Therefore, image cropping or stretching is commonly performed. Sometimes, these operations can lead to image distortion, affecting the final detection accuracy. SPP effectively avoids the image distortion problem caused by cropping and stretching. Furthermore, it solves the problem of repetitive feature extraction in CNNs, improving the speed of candidate box generation and reducing computational costs. As shown in Figure 4, the SPP module consists of 5x5, 9x9, and 13x13 max pooling layers and a skip connection.
[0062] For complex multi-object detection, the SPP module fuses local and global features, enriching the expressive power of the feature map. This is highly advantageous for detecting targets with significant size differences in the image, greatly improving detection accuracy. To further enhance detection accuracy, an SPP module is added after each output of the feature extraction network, for a total of three SPP modules.
[0063] As shown in Figure 5, PANet adds a bottom-up pyramid to YOLOv3's FPN to pass the low-level positioning information upwards, thus improving the previous deficiency of insufficient positioning information.
[0064] The FPN primarily allocates each proposal based on its size, assigning proposals of corresponding sizes to the corresponding feature layers. However, this approach also has significant problems. For example, proposals of similar sizes may be assigned to adjacent layers. Furthermore, since the allocation is based solely on size, it may fail to correlate the importance of features with their hierarchical relationship.
[0065] PANet, based on the Mask R-CNN approach, uses ROIAlign to provide each proposal with more layers of information, which is more conducive to localization and classification.
[0066] This embodiment provides a multi-target detection method based on an improved YPLOV3 algorithm, the innovations of which include:
[0067] Innovation 1: By using the Soft-NMS algorithm instead of the traditional NMS algorithm, the problem of missed detection caused by the high overlap of candidate boxes in the target detection algorithm is reduced;
[0068] Innovation Point 2: The PANet network is used to fuse high-level and low-level information, replacing the FPN network. This enhances the detection capability and localization and classification effect of small target objects. In addition, an SPP module is added after each feature output of the backbone network to improve detection accuracy.
[0069] Innovation Point 3: The GhostNet network is used to replace the original Darknet-53 as the backbone feature extraction network. At the same time, depthwise separable convolutions are used to replace the traditional convolutions in PANet and SPP, which greatly reduces the amount of computation and parameters, improves the detection speed, and enhances the real-time performance of the object detection network.
[0070] Based on the same technical concept, this application also provides a multi-target detection device based on the improved YPLOV3 algorithm. Since the principle of the device in solving the problem is similar to that of the multi-target detection method based on the improved YPLOV3 algorithm, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.
[0071] The feature map acquisition module is used to extract features from the received image using a feature extraction network with GhostNet as the backbone, and obtain the corresponding feature map.
[0072] The mapping module is used to map the features into a feature vector of S×S×(B×5+C) based on the obtained feature map, where S×S is the number of grids in the image segmentation, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected.
[0073] The detection module is used to filter suitable prediction boxes using the Soft Non-Maximum Suppression algorithm Soft-NMS based on the set confidence threshold and Intersection over Union (IOU) value, and to perform multi-target detection using the suitable prediction boxes. Specifically, it is used to obtain the highest confidence score m and the corresponding candidate box bm based on the candidate box set B and its corresponding confidence score set, delete bm from B and put it into the filtered candidate box set D.
[0074] Iterate through all elements in set B. If the Intersection over Union (IOU) value is lower than the Ni value, add it to set D. If the IOU value is not lower than the Ni value, introduce the corresponding decay function si = si * f[iou(M,bi)]. If the confidence level after introducing the decay function is greater than the probability threshold δ, add it to set D. Finally, output the filtered set.
[0075] For ease of description, the above sections are divided into modules (or units) according to their functional modules and described separately. Of course, in implementing this invention, the functions of each module (or unit) can be implemented in one or more software or hardware components.
[0076] After introducing the multi-target detection method and apparatus based on the improved YPLOV3 algorithm according to exemplary embodiments of the present invention, a computing device according to another exemplary embodiment of the present invention will be introduced next.
[0077] Those skilled in the art will understand that various aspects of the present invention can be implemented as systems, methods, or program products. Therefore, various aspects of the present invention can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, collectively referred to herein as a "circuit," "module," or "system." In some possible implementations, the computing device according to the present invention may include at least one processor and at least one memory. The memory stores program code. When the program code is executed by the processor, the processor performs the steps in the multi-target detection method based on the improved YPLOV3 algorithm according to various exemplary embodiments of the present invention described above. For example, the processor can execute the steps shown in Figure 1: S11, based on the received image, feature extraction is performed using a feature extraction network with GhostNet as the backbone to obtain the corresponding feature map; and S12, according to the obtained feature map, the features are mapped to a feature vector of S×S×(B×5+C), where S×S is the number of grids in the image segmentation, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected; and S13, using the Soft-NMS algorithm, suitable prediction boxes are selected according to the set confidence threshold and the Intersection over Union (IOU) value, and multi-object detection is performed using the suitable prediction boxes.
[0078] In some possible implementations, various aspects of the multi-object detection method based on the improved YPLOV3 algorithm provided by the present invention can also be implemented as a program product, which includes program code. When the program product is run on a computer device, the program code is used to cause the computer device to execute the steps in the multi-object detection method based on the improved YPLOV3 algorithm according to various exemplary embodiments of the present invention described above. For example, the computer device can execute step S11 as shown in FIG1: based on the received image, perform feature extraction through a feature extraction network with GhostNet network as the backbone to obtain a corresponding feature map; and step S12: according to the obtained feature map, map the features into a feature vector of S×S×(B×5+C), where S×S is the number of grids segmented in the image, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected; and step S13: through the Soft Non-Maximum Suppression algorithm Soft-NMS, select suitable prediction boxes according to the set confidence threshold and Intersection over Union (IOU) value, and perform multi-object detection through the suitable prediction boxes.
[0079] The program product can take the form of any combination of one or more readable media. The readable media can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. An embodiment of the present invention for multi-target detection based on the improved YPLOV3 algorithm can take the form of a portable compact disk read-only memory (CD-ROM) and include program code, and can run on a computing device. However, the program product of the present invention is not limited thereto. In the present invention, the readable storage medium can be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus, or device.
[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0081] 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 means for implementing the functions specified in one or more flowchart blocks and / or one or more block diagram blocks.
[0082] These computer program instructions may also be stored in a computer-readable storage medium that directs 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 that implement the functions specified in one or more flowcharts and / or one or more block diagrams.
[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the steps specified in one or more flowcharts and / or one or more block diagrams. 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.
[0084] 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 multi-target detection method based on an improved YPLOv3 algorithm, characterized in that, include: Based on the received image, features are extracted using a feature extraction network with GhostNet as the backbone to obtain the corresponding feature map. Based on the obtained feature map, the features are mapped to a feature vector of S×S×(B×5+C), where S×S is the number of grids in the image segmentation, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected. Using the Soft-NMS algorithm, suitable prediction boxes are selected based on the set confidence threshold and Intersection over Union (IOU) value, and multi-object detection is performed using the suitable prediction boxes.
2. The method according to claim 1, characterized in that, The feature map is generated by using a linear transformation based on a set of original feature maps to generate Ghost feature maps with the lowest possible cost: assuming there is a set of original feature maps Y∈R h*w*c Where h and w are the height and width of the space, and m is the number of channels less than c, a feature map Y'∈R with m channels is generated through convolution. h*w*m For the feature maps of the remaining cm channels, a linear transformation Ф is performed on the feature map in Y'. i,j (y i The feature maps are generated using the following method: i = 1, ..., m; j = 1, ..., s. The two feature maps are then combined.
3. The method according to claim 2, characterized in that, The GhostNet network is composed of stacked Ghost feature map modules.
4. The method according to claim 1, characterized in that, Each output of the feature extraction network is followed by an SPP module, for a total of 3 SPP modules.
5. The method according to claim 4, characterized in that, The SPP module consists of 5x5, 9x9, and 13x13 max pooling and a jump connection.
6. The method according to claim 1, characterized in that, The Soft Non-Maximum Suppression (Soft-NMS) algorithm is used to select suitable prediction boxes based on a set confidence threshold and Intersection over Union (IOU) value. Specifically, it involves: obtaining the highest confidence score m and its corresponding candidate box bm from the candidate box set B and its corresponding confidence score set; removing bm from B and adding it to the selected candidate box set D; iterating through all elements in set B; if the IOU value is lower than Ni, adding it to set D; if the IOU value is not lower than Ni, introducing a corresponding decay function si = si * f[iou(M,bi)]; if the confidence score after introducing the decay function is greater than the probability threshold δ, adding it to set D; and finally outputting the selected set.
7. A multi-target detection system based on an improved YPLOv3 algorithm, characterized in that, include: The feature map acquisition module is used to extract features from the received image using a feature extraction network with GhostNet as the backbone, and obtain the corresponding feature map. The mapping module is used to map the obtained feature map into a feature vector of S×S×(B×5+C), where S×S is the number of grids in the image segmentation, B is the number of bounding boxes to be predicted for each grid, and C is the number of object categories to be detected. The detection module is used to select suitable prediction boxes through the Soft-NMS algorithm based on the set confidence threshold and the Intersection over Union (IOU) value, and to perform multi-object detection using the suitable prediction boxes.
8. The system according to claim 7, characterized in that, The detection module is specifically used to obtain the highest confidence score m and the corresponding candidate box bm based on the candidate box set B and its corresponding confidence score set, delete bm from B and put it into the filtered candidate box set D; iterate through all elements in set B, and if the intersection-union ratio (IOU) is lower than the Ni value, it is put into set D; if the IOU is not lower than the Ni value, the corresponding decay function si = si * f[iou(M,bi)] is introduced; if the confidence score after introducing the decay function is greater than the probability threshold δ, it is put into set D, and finally the filtered set is output.
9. A computing device, characterized in that, It includes at least one processor and at least one memory, wherein the memory stores a computer program, and the processor is configured to read the computer program from the memory and execute the method according to any one of claims 1 to 6.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the method described in any one of claims 1 to 6.