Abnormal target identification method, apparatus and device, and readable storage medium
Through the design of four feature extraction modules and multiple detection heads, the problem of difficult to balance the efficiency and accuracy of abnormal target recognition in the prior art is solved, efficient and accurate identification results are achieved, and real-time requirements are met.
Patent Information
- Application Number
- CN202411020103.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-12-13
AI Technical Summary
The existing abnormal target recognition methods cannot take into account the accuracy and efficiency of the identification results. As the number of network layers increases, the model parameters and calculation amount increase, resulting in a decrease in efficiency and cannot meet the real-time requirements.
Four connected feature extraction modules are used to deeply extract the initial feature map, and redundant information is reduced through the feature fusion module, and multiple detection heads are used for identification, so as to identify targets at different scales, reduce the calculation amount and improve efficiency.
It is realized that with few feature extraction modules, the accuracy and efficiency of the recognition results are ensured, and the real-time requirements of abnormal target recognition are met.
Smart Images
Figure CN119131449A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of target recognition technology, and in particular to an abnormal target recognition method, device, equipment and computer-readable storage medium. Background Art
[0002] Abnormal target recognition plays a key role in the security field. By using abnormal target recognition technology to identify and analyze image information collected by front-end equipment, people in the monitoring area can be identified and their behavior patterns, such as running and fighting, can be analyzed, thereby achieving effective monitoring of personnel activities, providing strong support for security management and control, and improving the efficiency and timeliness of security prevention.
[0003] Traditional abnormal target recognition methods extract features from the image to be recognized, perform dimensionality reduction and feature selection on the extracted features to reduce redundant information, and finally input the processed features into the fully connected layer to obtain the abnormal targets in the recognized image. However, as the abnormal target recognition scenes become more and more complex, factors such as image occlusion and perspective changes gradually affect the effectiveness of feature extraction. Therefore, in order to improve the effectiveness of feature extraction, existing abnormal target recognition methods begin to continuously deepen the number of layers of the feature extraction network to fully capture feature information of different scales in the image, thereby obtaining more accurate recognition results based on richer feature information. However, as the number of network layers increases, the parameters and computational complexity of the model also increase, which will lead to a decrease in the efficiency of abnormal target recognition and fail to meet the real-time requirements of abnormal target recognition.
[0004] In summary, the existing abnormal target recognition methods have the problem of being unable to balance the accuracy and efficiency of abnormal target recognition results. Summary of the invention
[0005] To this end, the technical problem to be solved by the present invention is to overcome the problem that the abnormal target recognition method in the prior art cannot take into account both the accuracy and recognition efficiency of the abnormal target recognition results.
[0006] In order to solve the above technical problems, the present invention provides an abnormal target identification method, comprising:
[0007] Acquire a pre-trained abnormal target recognition model; the abnormal target recognition model includes an initial convolution module, a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module connected in series in a forward propagation direction, a first feature fusion module, a second feature fusion module, a third feature fusion module, a first detection head, a second detection head, a third detection head and a fourth detection head;
[0008] Input the image to be detected into the initial convolution module and output the initial feature map;
[0009] Inputting the initial feature map to a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module which are connected in series in a forward propagation direction, and obtaining a first feature map, a second feature map, a third feature map and a fourth feature map based on outputs of the feature extraction modules;
[0010] Input the first feature map and the second feature map into a first feature fusion module, and output a first fused feature map; input the second feature map and the third feature map into a second feature fusion module, and output a second fused feature map; input the third feature map and the fourth feature map into a third feature fusion module, and output a third fused feature map;
[0011] Input the first feature map to the first detection head, and output the first target recognition result; input the first fused feature map to the second detection head, and output the second target recognition result; input the second fused feature map to the third detection head, and output the third target recognition result; input the third fused feature map to the fourth detection head, and output the fourth target recognition result;
[0012] Based on the first target recognition result, the second target recognition result, the third target recognition result and the fourth target recognition result, an abnormal target recognition result of the image to be detected is obtained.
[0013] Preferably, the first feature extraction module, the second feature extraction module, the third feature extraction module and the fourth feature extraction module all include a first convolution submodule and a C2F submodule connected in series; wherein the number of bottleneck layers of the C2F submodule is 2 layers.
[0014] Preferably, the fourth feature extraction module also includes a spatial pyramid pooling module for pooling the fourth feature map.
[0015] Preferably, the spatial pyramid pooling module includes a second convolution submodule, a parameter-free attention submodule, a first pooling submodule, a second pooling submodule, a third pooling submodule and a fourth pooling submodule which are sequentially connected in series along the forward propagation direction.
[0016] Preferably, the spatial pyramid pooling module performs pooling on the fourth feature map, including:
[0017] Inputting the fourth feature map into the second convolution submodule for convolution, and outputting a first target feature map;
[0018] Inputting the first target feature map into the parameter-free attention submodule, and outputting a second target feature map;
[0019] Input the second target feature map into the first pooling submodule for pooling, and output a third target feature map;
[0020] Inputting the third target feature map into the second pooling submodule for pooling, and outputting a fourth target feature map;
[0021] Inputting the fourth target feature map into the third pooling submodule for pooling, and outputting a fifth target feature map;
[0022] The fifth target feature map is input into the fourth pooling submodule for pooling, and the pooled fourth feature map is output.
[0023] Preferably, the first detection head, the second detection head, the third detection head and the fourth detection head each include a third convolution submodule, a parallel bounding box prediction submodule and a category prediction submodule.
[0024] Preferably, after obtaining the abnormal target recognition result of the image to be detected, the method further includes: inputting the abnormal target recognition result into a database in the grok-1 model, so as to query the abnormal target recognition result based on the query interface of the grok-1 model.
[0025] The present invention also provides a target recognition device, comprising:
[0026] A model acquisition module is used to acquire a pre-trained abnormal target recognition model; the abnormal target recognition model includes an initial convolution module, a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module connected in series in a forward propagation direction, a first feature fusion module, a second feature fusion module, a third feature fusion module, a first detection head, a second detection head, a third detection head and a fourth detection head;
[0027] The initial feature map acquisition module is used to input the image to be detected into the initial convolution module and output the initial feature map;
[0028] A feature extraction module, configured to input the initial feature map into a first feature extraction module, a second feature extraction module, a third feature extraction module, and a fourth feature extraction module connected in series in a forward propagation direction, and obtain a first feature map, a second feature map, a third feature map, and a fourth feature map based on outputs of the feature extraction modules;
[0029] A feature fusion module, used for inputting the first feature map and the second feature map into a first feature fusion module, and outputting a first fused feature map; inputting the second feature map and the third feature map into a second feature fusion module, and outputting a second fused feature map; inputting the third feature map and the fourth feature map into a third feature fusion module, and outputting a third fused feature map;
[0030] A detection module, configured to input the first feature map into a first detection head and output a first target recognition result; input the first fused feature map into a second detection head and output a second target recognition result; input the second fused feature map into a third detection head and output a third target recognition result; input the third fused feature map into a fourth detection head and output a fourth target recognition result;
[0031] The recognition result acquisition module is used to obtain the abnormal target recognition result of the image to be detected based on the first target recognition result, the second target recognition result, the third target recognition result and the fourth target recognition result.
[0032] The present invention also provides an abnormal target identification device, comprising:
[0033] An image acquisition device, used for acquiring images to be detected;
[0034] A host computer is connected to the image acquisition device for acquiring the image to be detected and performing the steps of the abnormal target recognition method to obtain an abnormal target recognition result of the image to be detected;
[0035] The display device is connected to the host computer for displaying the abnormal target recognition result of the image to be detected.
[0036] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned abnormal target identification method are implemented.
[0037] The abnormal target recognition method provided by the present application uses four serially connected feature extraction modules to perform deep feature extraction on the initial feature map to obtain the first feature map, the second feature map, the third feature map and the fourth feature map. Since each feature map contains feature information of different scales, target information of different scales can be recognized based on different feature maps; at the same time, the first feature fusion module is used to fuse the first feature map and the second feature map, the second feature fusion module is used to fuse the second feature map and the third feature map, and the third feature fusion module is used to fuse the third feature map and the fourth feature map, so as to achieve the complementarity of detail information and semantic information between feature maps of different scales and reduce the single feature map. The errors that may be caused by hierarchical features can be eliminated, and by merging the overlapping information in features of different levels, redundant information can be reduced, the calculation amount of the model can be reduced, and the efficiency of abnormal target recognition can be improved; finally, multiple detection heads are used to identify the first feature map, the first fused feature map, the second fused feature map and the third fused feature map respectively, which can not only realize the recognition of large targets, but also the recognition of small targets. Even in the case of only four feature extraction modules, the extracted feature information can still be fully utilized to ensure the accuracy of the abnormal target recognition results, which not only takes into account the recognition efficiency and meets the real-time nature of abnormal target recognition, but also ensures the accuracy of the recognition results. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:
[0039] Figure 1 Flowchart of the abnormal target identification method provided for this application;
[0040] Figure 2 A structural diagram of a security management system provided for this application;
[0041] Figure 3 This is a structural diagram of the abnormal target identification device provided in this application. DETAILED DESCRIPTION
[0042] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.
[0043] See also Figure 1 , Figure 1 The figure shows a flow chart of the abnormal target identification method provided by the present application, which specifically includes:
[0044] S10: Acquire a pre-trained abnormal target recognition model; the abnormal target recognition model includes an initial convolution module, a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module connected in series in a forward propagation direction, a first feature fusion module, a second feature fusion module, a third feature fusion module, a first detection head, a second detection head, a third detection head and a fourth detection head;
[0045] S20: Input the image to be detected into the initial convolution module and output the initial feature map;
[0046] S30: inputting the initial feature map to a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module which are connected in series in a forward propagation direction, and obtaining a first feature map, a second feature map, a third feature map and a fourth feature map based on outputs of the feature extraction modules;
[0047] S40: Input the first feature map and the second feature map into a first feature fusion module, and output a first fused feature map; input the second feature map and the third feature map into a second feature fusion module, and output a second fused feature map; input the third feature map and the fourth feature map into a third feature fusion module, and output a third fused feature map;
[0048] S50: Input the first feature map to the first detection head, and output the first target recognition result; input the first fused feature map to the second detection head, and output the second target recognition result; input the second fused feature map to the third detection head, and output the third target recognition result; input the third fused feature map to the fourth detection head, and output the fourth target recognition result;
[0049] S60: Obtaining an abnormal target recognition result of the image to be detected based on the first target recognition result, the second target recognition result, the third target recognition result and the fourth target recognition result.
[0050] The abnormal target recognition method provided by the present application uses four serially connected feature extraction modules to perform deep feature extraction on the initial feature map to obtain the first feature map, the second feature map, the third feature map and the fourth feature map. Since each feature map contains feature information of different scales, target information of different scales can be recognized based on different feature maps; at the same time, the first feature fusion module is used to fuse the first feature map and the second feature map, the second feature fusion module is used to fuse the second feature map and the third feature map, and the third feature fusion module is used to fuse the third feature map and the fourth feature map, so as to achieve the complementarity of detail information and semantic information between feature maps of different scales and reduce the single feature map. The errors that may be caused by hierarchical features can be eliminated, and by merging the overlapping information in features of different levels, redundant information can be reduced, the calculation amount of the model can be reduced, and the efficiency of abnormal target recognition can be improved; finally, multiple detection heads are used to identify the first feature map, the first fused feature map, the second fused feature map and the third fused feature map respectively, which can not only realize the recognition of large targets, but also the recognition of small targets. Even in the case of only four feature extraction modules, the extracted feature information can still be fully utilized to ensure the accuracy of the abnormal target recognition results, which not only takes into account the recognition efficiency and meets the real-time nature of abnormal target recognition, but also ensures the accuracy of the recognition results.
[0051] For example, by recognizing the first feature map through the first detection head, small targets larger than 4*4 can be identified, by recognizing the first fused feature map through the second detection head, small targets larger than 8*8 can be identified, by recognizing the second fused feature map through the third detection head, medium-sized targets larger than 16*16 can be identified, and by recognizing the third fused feature map through the fourth detection head, large targets larger than 32*32 can be identified. By designing detection heads corresponding to different feature maps, the recognition of targets of different scales can be realized. Finally, based on the recognition results of all detection heads, the abnormal target recognition result of the image to be detected is obtained.
[0052] Specifically, in some embodiments of the present application, the first feature extraction module, the second feature extraction module, the third feature extraction module and the fourth feature extraction module all include a first convolution submodule and a C2F submodule, wherein the number of bottleneck layers in the C2F submodule is 2.
[0053] Optionally, in some embodiments of the present application, a spatial pyramid pooling module is also included after the fourth feature extraction module, which is used to pool the fourth feature map output by the fourth feature extraction module to reduce the dimension and aggregate the features of the fourth feature map, further reduce the computational complexity of the model, and improve the efficiency of abnormal target recognition.
[0054] Specifically, the spatial pyramid pooling module includes a second convolution submodule, a parameter-free attention submodule, a first pooling submodule, a second pooling submodule, a third pooling submodule and a fourth pooling submodule which are sequentially connected in series along the forward propagation direction.
[0055] Based on the above structure, the spatial pyramid pooling module performs pooling on the fourth feature map including:
[0056] Input the fourth feature map into the second convolution submodule for convolution, and output the first target feature map;
[0057] Input the first target feature map into the parameter-free attention submodule and output the second target feature map;
[0058] Input the second target feature map into the first pooling submodule for pooling, and output a third target feature map;
[0059] Input the third target feature map into the second pooling submodule for pooling, and output a fourth target feature map;
[0060] Input the fourth target feature map into the third pooling submodule for pooling, and output a fifth target feature map;
[0061] The fifth target feature map is input into the fourth pooling submodule for pooling, and the pooled fourth feature map is output.
[0062] In order to further reduce the parameters of the abnormal target recognition model, reduce the amount of calculation, and thus improve the recognition efficiency, in some embodiments of the present application, the first detection head, the second detection head, the third detection head, and the fourth detection head all include a third convolution submodule, a parallel bounding box prediction submodule, and a category prediction submodule;
[0063] Specifically, the feature map is further extracted by the third convolution submodule, and then the output of the third convolution submodule is respectively input into the bounding box prediction submodule and the category prediction submodule, the bounding box prediction submodule is used to predict the bounding box parameters, and the category prediction submodule is used to predict the category probability distribution, so as to obtain the target recognition result of the input feature map based on the bounding box parameters and the predicted category probability distribution;
[0064] The decoupled head architecture is used to realize the independent operation of classification and bounding box prediction, so that the two tasks no longer share parameters. Since there are differences in the areas concerned by target category information and location information, this decoupled head design can further refine the difference between the two, which not only reduces the number of model parameters and computational complexity, but also helps to improve the generalization ability of the model and improve the efficiency of abnormal target recognition.
[0065] Furthermore, after obtaining the abnormal target recognition result of the image to be detected, the abnormal target recognition result can also be input into the database of the grok-1 model so that the abnormal target recognition result can be queried based on the query interface of the grok-1 model.
[0066] Specifically, grok-1 is a large-scale model with powerful natural language processing capabilities. It is built on the basis of Transformer and hybrid expert architecture. It can be widely used in multiple fields such as information retrieval, question answering, code snippet generation, etc. In addition, grok-1 integrates search tools and databases, which can further enhance the capabilities and authenticity of the model and provide users with more accurate and efficient services. Therefore, the abnormal target recognition results are stored in the database of the grok-1 model. Users or staff can retrieve the existing information in the database by entering text information, thereby obtaining the corresponding abnormal target recognition results.
[0067] In combination with the abnormal target identification method provided in the above embodiment, the embodiment of the present application also applies it to the security field and proposes a security management system, such as Figure 2 As shown, the system includes: a front-end perception module, a transcoding module, a business platform, an algorithm module, a database, and an information retrieval module; wherein the algorithm module is used to implement the abnormal target identification method provided in the above embodiment, and the database is a database of the grok-1 model;
[0068] Specifically, the front-end perception module is used to capture road information through cameras and other devices, generate real-time video streams, and then aggregate the video streams to a unified video stream management platform to achieve centralized monitoring and management; the transcoding module is used to transcode the real-time video streams generated by the front-end perception module, thereby reducing bandwidth occupancy and transmission delays during transmission; the business platform is used for front-end screen information display, user-system interaction, docking with other platforms and providing basic operation and maintenance information; the algorithm module is used to identify abnormal targets for each video frame in the transcoded video stream, obtain abnormal target identification results for each video frame, and transmit the results to the database in the grok-1 model; users can send query requests to the security management system through remote devices, and the information retrieval module responds to the request and verifies information such as the mac address of the remote device, the requester's authority and the timeliness of the information, and sends the corresponding abnormal target identification results to the remote device.
[0069] The abnormal target identification method provided in the embodiment of the present application is applied to the security management system, which can not only identify larger violation events such as trucks pulling to the right, key vehicle management, and suspected vehicle inspection, but also identify smaller and difficult-to-identify violation events such as road spillage and license plate obstruction, thereby quickly and accurately detecting and identifying abnormal behaviors or targets in each area, so that relevant personnel can respond and handle them in a timely manner based on the identification results.
[0070] Based on the abnormal target identification method provided in the above embodiment, the embodiment of the present application also provides an abnormal target identification device, such as Figure 3 As shown, the device specifically includes:
[0071] The model acquisition module 10 is used to acquire a pre-trained abnormal target recognition model; the abnormal target recognition model includes an initial convolution module, a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module connected in series in a forward propagation direction, a first feature fusion module, a second feature fusion module, a third feature fusion module, a first detection head, a second detection head, a third detection head and a fourth detection head;
[0072] The initial feature map acquisition module 20 is used to input the image to be detected into the initial convolution module and output the initial feature map;
[0073] A feature extraction module 30 is used to input the initial feature map to a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module which are connected in series in a forward propagation direction, and obtain a first feature map, a second feature map, a third feature map and a fourth feature map based on the output of each feature extraction module;
[0074] The feature fusion module 40 is used to input the first feature map and the second feature map into the first feature fusion module, and output the first fused feature map; input the second feature map and the third feature map into the second feature fusion module, and output the second fused feature map; input the third feature map and the fourth feature map into the third feature fusion module, and output the third fused feature map;
[0075] The detection module 50 is used to input the first feature map into the first detection head and output the first target recognition result; input the first fused feature map into the second detection head and output the second target recognition result; input the second fused feature map into the third detection head and output the third target recognition result; input the third fused feature map into the fourth detection head and output the fourth target recognition result;
[0076] The recognition result acquisition module 60 is used to obtain the abnormal target recognition result of the image to be detected based on the first target recognition result, the second target recognition result, the third target recognition result and the fourth target recognition result.
[0077] The embodiment of the present application also provides an abnormal target identification device, including:
[0078] An image acquisition device, used for acquiring images to be detected;
[0079] A host computer is connected to the image acquisition device for acquiring the image to be detected and executing the steps of the abnormal target recognition method to obtain the abnormal target recognition result of the image to be detected;
[0080] The display device is connected to the host computer for displaying the abnormal target recognition result of the image to be detected.
[0081] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned abnormal target identification method are implemented.
[0082] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0083] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0084] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0085] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0086] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.
Claims
1. A method for identifying abnormal targets, characterized in that: include: Acquire a pre-trained abnormal target recognition model; the abnormal target recognition model includes an initial convolution module, a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module connected in series in a forward propagation direction, a first feature fusion module, a second feature fusion module, a third feature fusion module, a first detection head, a second detection head, a third detection head and a fourth detection head; Input the image to be detected into the initial convolution module and output the initial feature map; Inputting the initial feature map to a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module which are connected in series in a forward propagation direction, and obtaining a first feature map, a second feature map, a third feature map and a fourth feature map based on outputs of the feature extraction modules; Inputting the first feature map and the second feature map into a first feature fusion module, and outputting a first fused feature map; Input the second feature map and the third feature map into a second feature fusion module, and output a second fused feature map; input the third feature map and the fourth feature map into a third feature fusion module, and output a third fused feature map; Input the first feature map to a first detection head, and output a first target recognition result; input the first fused feature map to a second detection head, and output a second target recognition result; Inputting the second fused feature map into a third detection head, and outputting a third target recognition result; Inputting the third fusion feature map into a fourth detection head, and outputting a fourth target recognition result; Based on the first target recognition result, the second target recognition result, the third target recognition result and the fourth target recognition result, an abnormal target recognition result of the image to be detected is obtained.
2. The abnormal target identification method according to claim 1, characterized in that: The first feature extraction module, the second feature extraction module, the third feature extraction module and the fourth feature extraction module all include a first convolution submodule and a C2F submodule connected in series; wherein the number of bottleneck layers of the C2F submodule is 2 layers.
3. The abnormal target identification method according to claim 1, characterized in that: The fourth feature extraction module also includes a spatial pyramid pooling module for pooling the fourth feature map.
4. The abnormal target identification method according to claim 3 is characterized in that: The spatial pyramid pooling module includes a second convolution submodule, a parameter-free attention submodule, a first pooling submodule, a second pooling submodule, a third pooling submodule and a fourth pooling submodule which are sequentially connected in series along a forward propagation direction.
5. The abnormal target identification method according to claim 4, characterized in that: The spatial pyramid pooling module pools the fourth feature map, including: Inputting the fourth feature map into the second convolution submodule for convolution, and outputting a first target feature map; Inputting the first target feature map into the parameter-free attention submodule, and outputting a second target feature map; Input the second target feature map into the first pooling submodule for pooling, and output a third target feature map; Inputting the third target feature map into the second pooling submodule for pooling, and outputting a fourth target feature map; Inputting the fourth target feature map into the third pooling submodule for pooling, and outputting a fifth target feature map; The fifth target feature map is input into the fourth pooling submodule for pooling, and the pooled fourth feature map is output.
6. The abnormal target identification method according to claim 1, characterized in that: The first detection head, the second detection head, the third detection head and the fourth detection head each include a third convolution submodule, a parallel bounding box prediction submodule and a category prediction submodule.
7. The abnormal target identification method according to claim 1, characterized in that: After obtaining the abnormal target recognition result of the image to be detected, the method further includes: inputting the abnormal target recognition result into a database in the grok-1 model so as to query the abnormal target recognition result based on the query interface of the grok-1 model.
8. An abnormal target identification device, characterized in that: include: A model acquisition module is used to acquire a pre-trained abnormal target recognition model; the abnormal target recognition model includes an initial convolution module, a first feature extraction module, a second feature extraction module, a third feature extraction module and a fourth feature extraction module connected in series in a forward propagation direction, a first feature fusion module, a second feature fusion module, a third feature fusion module, a first detection head, a second detection head, a third detection head and a fourth detection head; The initial feature map acquisition module is used to input the image to be detected into the initial convolution module and output the initial feature map; A feature extraction module, configured to input the initial feature map into a first feature extraction module, a second feature extraction module, a third feature extraction module, and a fourth feature extraction module connected in series in a forward propagation direction, and obtain a first feature map, a second feature map, a third feature map, and a fourth feature map based on outputs of the feature extraction modules; A feature fusion module, used for inputting the first feature map and the second feature map into a first feature fusion module, and outputting a first fused feature map; inputting the second feature map and the third feature map into a second feature fusion module, and outputting a second fused feature map; inputting the third feature map and the fourth feature map into a third feature fusion module, and outputting a third fused feature map; A detection module, used for inputting the first feature map into a first detection head and outputting a first target recognition result; Inputting the first fusion feature map into a second detection head, and outputting a second target recognition result; Inputting the second fused feature map into a third detection head, and outputting a third target recognition result; Inputting the third fusion feature map into a fourth detection head, and outputting a fourth target recognition result; The recognition result acquisition module is used to obtain the abnormal target recognition result of the image to be detected based on the first target recognition result, the second target recognition result, the third target recognition result and the fourth target recognition result.
9. An abnormal target identification device, characterized in that: include: An image acquisition device, used for acquiring images to be detected; A host computer, which is in communication with the image acquisition device, is used to obtain the image to be detected, and execute the steps of the abnormal target recognition method according to any one of claims 1 to 7 to obtain an abnormal target recognition result of the image to be detected; The display device is connected to the host computer for displaying the abnormal target recognition result of the image to be detected.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the abnormal target identification method according to any one of claims 1 to 7 are implemented.