Target identification model, transformer area on-site target identification method and system, and medium

By introducing PConvC2fEMA and VoVGSCSP modules in the YOLOv8n model, the problem of high parameters and complexity of the YOLOv8n model in the on-site intelligent self-management system in the station area is solved, and faster target recognition speed and higher recognition accuracy are achieved. In particular, Reduced error detection in hard helmet wear detection.

CN119048733APending Publication Date: 2024-11-29SOUTHERN POWER GRID DIGITAL GRID RESEARCH INSTITUTE CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411057806.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

In the existing smart self-management system in the station area, the parameters and complexity of the YOLOv8n basic model are high, resulting in large amounts of calculations of the detection model, which is not conducive to terminal deployment, especially in the detection of hard helmet wear detection. question.

Method used

The target recognition model is constructed based on the YOLOv8n algorithm, the PConvC2fEMA module is used to improve the backbone network, the EMA attention mechanism and VoVGSCSP module are introduced to improve the neck network, the model parameters and calculation amount are reduced, and the attention and nonlinear expression ability are enhanced for target representation.

Benefits of technology

The target recognition model is lightweight, the recognition speed and accuracy are improved, the error detection and missed detection are reduced, and the accuracy of safety helmet wear detection is enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048733A_ABST
    Figure CN119048733A_ABST
Patent Text Reader

Abstract

The invention discloses a target recognition model, a transformer area on-site target recognition method and system and a medium, the target recognition model is constructed based on a YOLOv8n algorithm, and a backbone network of the YOLOv8n model is improved by specifically adopting a PConvC2fEMA module, so that the reduction of model parameter quantity and calculation quantity is realized, the attention on target representation is enhanced, and the model recognition precision is improved; the neck network of the YOLOv8n model is improved by adopting the VoVGSCSP module, so that the nonlinear expression capability of the neck network can be improved; according to the method, the YOLOv8n model is subjected to lightweight improvement to obtain the target recognition model, and compared with the YOLOv8n model, the target recognition model is smaller in scale and calculation amount, higher in target recognition speed and still higher in recognition precision.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to a target recognition model, a method, a system and a medium for local target recognition in a station area. Background Art

[0002] With the progress of society and the development of science and technology, the demand for electricity is growing, and the scale and complexity of the distribution network are also increasing. In this context, the power industry is facing important issues such as how to realize the intelligence, self-management, self-control and self-evaluation of the distribution network. In view of the complex on-site environment and personnel of the substation area, and the lack of self-management capabilities, this project uses the on-site intelligent self-management technology of the substation area to solve the shortcomings of existing technical means. By integrating video acquisition, on-site intelligence and other technologies on the edge side, the on-site intelligent self-management of the substation area is realized. The video data collected on the scene in real time by the edge end is used to identify and analyze the real-time video images through video analysis technology. If there are non-compliance with safety regulations or dangerous factors in the video, an alarm will be issued to achieve an early warning effect.

[0003] At present, the on-site intelligent self-management of the substation area intelligently identifies the on-site video data by deploying the edge computing terminal on-site. In order to better adapt to the edge terminal equipment, a lightweight intelligent recognition algorithm is needed to improve the detection speed and ensure a certain accuracy. In view of the on-site safety wear management and control issues in the substation area, the detection of the staff's helmet wearing is particularly important. Among the current target detection algorithms, YOLOv8 is a newer algorithm in the YOLO series of real-time object detectors, with cutting-edge performance in terms of accuracy and speed. However, the YOLOv8n basic model has a high number of parameters and complexity, and the detection model still has the problem of large computational complexity, which is not conducive to terminal deployment tasks. Summary of the invention

[0004] The present invention provides a target recognition model, a method, a system and a medium for local target recognition in a radio station area, which can effectively improve the recognition speed of the local target object in the radio station area.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a target recognition model, which is constructed based on the YOLOv8n algorithm and includes:

[0006] a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a fifth CBS module, a sixth CBS module, a seventh CBS module, a first PConvC2fEMA module, a second PConvC2fEMA module, a third PConvC2fEMA module, a fourth PConvC2fEMA module, an SPPF module, a first Upsample module, a second Upsample module, a first Concat module, a second Concat module, a third Concat module, a fourth Concat module, a first VoVGSCSP module, a second VoVGSCSP module, a third VoVGSCSP module, a fourth VoVGSCSP module, a first output module, a second output module, and a third output module; wherein:

[0007] The first CBS module serves as an input end of the target recognition model, and is used to input an image to be recognized, and its output end is connected to the input end of the second CBS module;

[0008] The output end of the second CBS module is connected to the input end of the first PConvC2fEMA module;

[0009] The output end of the first PConvC2fEMA module is connected to the input end of the third CBS module;

[0010] The output end of the third CBS module is connected to the input end of the second PConvC2fEMA module;

[0011] The first output terminal of the second PConvC2fEMA module is connected to the input terminal of the fourth CBS module; the second output terminal of the second PConvC2fEMA module is connected to the first input terminal of the second Concat module;

[0012] The output end of the fourth CBS module is connected to the input end of the third PConvC2fEMA module;

[0013] The first output end of the third PConvC2fEMA module is connected to the input end of the fifth CBS module; the second output end of the third PConvC2fEMA module is connected to the first input end of the first Concat module;

[0014] The output end of the fifth CBS module is connected to the input end of the fourth PConvC2fEMA module;

[0015] The output end of the fourth PConvC2fEMA module is connected to the input end of the SPPF module;

[0016] The output end of the SPPF module is connected to the input end of the first Upsample module and the first input end of the fourth Concat module;

[0017] The output end of the first Upsample module is connected to the second input end of the first Concat module;

[0018] The output end of the first Concat module is connected to the input end of the first VoVGSCSP module;

[0019] The first output terminal of the first VoVGSCSP module is connected to the input terminal of the second Upsample module; the second output terminal of the first VoVGSCSP module is connected to the first input terminal of the third Concat module;

[0020] The output end of the second Upsample module is connected to the second input end of the second Concat module;

[0021] The output end of the second Concat module is connected to the input end of the second VoVGSCSP module;

[0022] The first output terminal of the second VoVGSCSP module is connected to the input terminal of the sixth CBS module;

[0023] The output end of the sixth CBS module is connected to the second input end of the third Concat module;

[0024] The output end of the third Concat module is connected to the input end of the third VoVGSCSP module;

[0025] The output end of the third VoVGSCSP module is connected to the input end of the seventh CBS module;

[0026] The output end of the seventh CBS module is connected to the second input end of the fourth Concat module;

[0027] The first output terminal of the fourth Concat module is connected to the input terminal of the fourth VoVGSCSP module;

[0028] The second output end of the fourth Concat module is connected to the input end of the first output module;

[0029] The output end of the third Concat module is connected to the input end of the second output module;

[0030] The output end of the second VoVGSCSP module is connected to the input end of the third output module;

[0031] The first output module, the second output module and the third output module serve as output ends of the target recognition model and are used to output recognition results.

[0032] The beneficial effects of the embodiments of the present invention are:

[0033] Compared with the prior art that uses the YOLOv8n basic model with high parameter amount and complexity for target recognition, the embodiment of the present invention constructs a target recognition model based on the YOLOv8n algorithm, and specifically uses the PConvC2fEMA module to improve the backbone network of the YOLOv8n model to achieve a reduction in the model parameter amount and the amount of calculation, and enhances the focus on target representation to improve the accuracy of model recognition; and the present invention uses the VoVGSCSP module to improve the neck network of the YOLOv8n model, which can improve the nonlinear expression ability of the neck network; the present invention performs lightweight improvement on the YOLOv8n model to obtain a target recognition model. Compared with the YOLOv8n model, the target recognition model has a smaller scale and a smaller amount of calculation, can have a faster target recognition speed, and still have a higher recognition accuracy.

[0034] As a preferred solution, the first PConvC2fEMA module, the second PConvC2fEMA module, the third PConvC2fEMA module and the fourth PConvC2fEMA module have the same composition; the first PConvC2fEMA module includes: a first input unit, a first CBS unit, a second CBS unit, a split unit, a first PConvBlock unit, a second PConvBlock unit, a third PConvBlock unit, a first Concat unit, an EMA attention mechanism unit and a first output unit; wherein:

[0035] The input end of the first input unit is used as the input end of the first PConvC2fEMA module;

[0036] The output end of the first input unit is connected to the input end of the first CBS unit;

[0037] The output end of the first CBS unit is connected to the input end of the split unit;

[0038] The first output end of the split unit is connected to the input end of the first PConvBlock unit and the first input end of the first Concat unit; the second output end of the split unit is connected to the first input end of the first Concat unit;

[0039] The first output terminal of the first PConvBlock unit is connected to the second input terminal of the first Concat unit;

[0040] The second output terminal of the first PConvBlock unit is connected to the input terminal of the second PConvBlock unit;

[0041] The first output end of the second PConvBlock unit is connected to the input end of the third PConvBlock unit; the second output end of the second PConvBlock unit is connected to the third input end of the first Concat unit;

[0042] The output end of the third PConvBlock unit is connected to the third input end of the first Concat unit;

[0043] The output end of the first Concat unit is connected to the input end of the second CBS unit;

[0044] The output end of the second CBS unit is connected to the input end of the EMA attention mechanism unit;

[0045] The output end of the EMA attention mechanism unit is connected to the input end of the first output unit;

[0046] The output end of the first output unit serves as the output end of the first PConvC2fEMA module.

[0047] The EMA attention mechanism is introduced in the PConvC2fEMA module of this preferred solution. The EMA attention mechanism generates spatial attention weight values ​​through cross-space learning, integrates multi-scale spatial information and position information, can capture the pixel-level pairwise relationship of the feature map and highlight the global context information of the pixel without increasing the number of parameters and calculations. It can fully integrate the input feature map information, so that the number of parameters of the PConvC2fEMA module is small, and the number of parameters of the backbone network in the model is further reduced to achieve the goal of lightweight model. Moreover, after the introduction of the EMA attention mechanism, the interference of complex background can be weakened, so that the target recognition model pays more attention to the task of whether to wear a helmet to detect, avoid false detection, missed detection, etc., and improve the recognition precision and accuracy of the target recognition model.

[0048] As a preferred solution, the first PConvBlock unit, the second PConvBlock unit and the third PConvBlock unit have the same composition; the first PConvBlock unit includes: an input subunit, a PConv subunit, a BN subunit, a SiLU subunit, a first Conv subunit, a second Conv subunit, and an output subunit; wherein:

[0049] The input end of the input subunit serves as the input end of the first PConvBlock unit;

[0050] The output end of the input subunit is connected to the input end of the PConv subunit;

[0051] The output end of the PConv subunit is connected to the input end of the first Conv subunit;

[0052] The output end of the first Conv subunit is connected to the input end of the BN subunit;

[0053] The output end of the BN subunit is connected to the input end of the SiLU subunit;

[0054] The output end of the SiLU subunit is connected to the input end of the second Conv subunit;

[0055] The output end of the second Conv subunit and the output end of the input subunit are matrix-summed and then connected to the input end of the output subunit;

[0056] The output end of the output subunit serves as the output end of the first PConvBlock unit.

[0057] In this preferred embodiment, the PConvBlock unit includes a PConv subunit and two Conv subunits, wherein the middle layer uses a normalization layer and introduces a SiLU subunit based on the SiLU activation function. The PConvBlock unit uses the inverted residual idea in MobileNetV2 to expand the number of channels in the middle layer, thereby significantly reducing the number of network parameters and accelerating the reasoning speed by reducing the number of convolution operations. The PConvBlock unit in this preferred embodiment has low complexity and small number of parameters, which can reduce the number of parameters of the PConvC2fEMA module and further reduce the number of parameters and computational complexity of the target recognition model.

[0058] As a preferred solution, the first VoVGSCSP module, the second VoVGSCSP module, the third VoVGSCSP module and the fourth VoVGSCSP module have the same composition; the first VoVGSCSP module includes: a second input unit, a first Conv unit, a second Conv unit, a third Conv unit, a GSBottleneck unit, a second Concat unit and a second output unit; wherein:

[0059] The input end of the second input unit serves as the input end of the first VoVGSCSP module;

[0060] The output end of the second input unit is connected to the input end of the first Conv unit and the input end of the second Conv unit;

[0061] The output end of the second Conv unit is connected to the first input end of the second Concat unit;

[0062] The output end of the first Conv unit is connected to the input end of the GSBottleneck unit;

[0063] The output end of the GSBottleneck unit is connected to the second input end of the second Concat unit;

[0064] The output end of the second Concat unit is connected to the input end of the third Conv unit;

[0065] The output end of the third Conv unit is connected to the input end of the second output unit;

[0066] The output end of the second output unit serves as the output end of the first VoVGSCSP module.

[0067] In this preferred solution, a VoVGSCSP module is constructed based on several Conv units, GSBottleneck units, Concat units and input-output units. Features are extracted through two branches, the first Conv unit and the second Conv unit. The feature maps of the branches are then aggregated. Convolution is used to extract information from the aggregated multi-channel feature maps to obtain richer feature information. Combined with the cross-stage operation of the residual class, the nonlinear expression ability of the neck network in the target recognition model is further improved, the full integration of the feature map information extracted by the backbone network is achieved, and the recognition and detection capabilities of small targets are enhanced.

[0068] Accordingly, in order to solve the above technical problems, an embodiment of the present invention further provides a method for identifying local targets in a substation area, comprising:

[0069] Constructing a target recognition model as described in any of the above items;

[0070] Collecting on-site image data of the station area, and obtaining a helmet data set based on the image data;

[0071] Train and test the object recognition model based on the helmet dataset to obtain the optimal recognition model;

[0072] The optimal recognition model is used to perform on-site safety helmet recognition in the station area.

[0073] The above-mentioned on-site target recognition method in the substation first constructs the target recognition model obtained based on the YOLOv8n model, trains and tests the target recognition model according to the obtained safety helmet data set, so that it can learn the characteristic data of the safety helmets in the substation, obtain the optimal recognition model, and use the optimal recognition model to perform on-site safety helmet recognition in the substation. The present invention performs on-site safety helmet recognition in the substation based on the target recognition model, which can improve the speed of on-site safety helmet recognition in the substation and ensure better recognition accuracy.

[0074] As a preferred solution, the collecting of local image data of the station area includes:

[0075] Collect on-site image data of the station area from several camera angles.

[0076] This preferred solution collects on-site image data of the station area from different camera angles to make the data in the subsequent data set more diversified. When the target recognition model is subsequently trained and tested, the generalization ability of the optimal recognition model obtained can be enhanced to reduce the risk of overfitting.

[0077] As a preferred solution, obtaining a helmet dataset according to the image data includes:

[0078] Preprocessing the image data to obtain an initial data set;

[0079] The initial data set is annotated based on a preset annotation type to obtain a helmet data set.

[0080] This preferred solution preprocesses the acquired image data to ensure the accuracy and reliability of the data so that the acquired data can be reliably used for subsequent model training and testing; and this preferred solution labels the initial data set based on a preset labeling type to obtain a safety helmet data set so that the subsequent target recognition model can learn the safety helmet features during model training, thereby enabling the target recognition model to accurately identify unknown data.

[0081] As a preferred solution, the training and testing of the target recognition model based on the helmet data set to obtain the optimal recognition model includes:

[0082] The target recognition model is trained and tested based on the helmet data set until the loss of the target recognition model on the validation set meets the preset conditions, thereby obtaining the optimal recognition model.

[0083] Accordingly, in order to solve the above technical problems, an embodiment of the present invention further provides a system for local target recognition in a substation area, comprising: a model building module, a data acquisition module, a model training module and a target recognition module;

[0084] Wherein, the model building module is used to build any of the target recognition models described above;

[0085] The data acquisition module is used to collect image data on the spot of the platform area, and obtain a helmet data set based on the image data;

[0086] The model training module is used to train and test the target recognition model based on the helmet data set to obtain the optimal recognition model;

[0087] The target recognition module is used to use the optimal recognition model to identify the on-site safety helmet in the platform area. The embodiment of the present invention also provides a storage medium, which stores a computer program, and when the computer program is executed by a processor, the steps of the on-site safety helmet recognition method in the platform area are implemented as described in any one of the above. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 : A structural schematic diagram of an embodiment of a target recognition model provided by the present invention;

[0089] Figure 2 : A schematic diagram of a flow chart of an embodiment of the PConvC2fEMA module provided by the present invention;

[0090] Figure 3 : A structural diagram of an embodiment of a PConvBlock unit provided by the present invention;

[0091] Figure 4 : A schematic diagram of the EMA attention mechanism provided by the present invention;

[0092] Figure 5 : A structural diagram of an embodiment of the VoVGSCSP module provided by the present invention;

[0093] Figure 6 : A flow chart of an embodiment of a method for on-site target identification in a substation area provided by the present invention. DETAILED DESCRIPTION

[0094] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0095] The existing YOLOv8 model is improved on the basis of the YOLOv5 model. The YOLOv8 model consists of three parts: Backbone (backbone network), Neck (neck network) and Head (prediction head).

[0096] In the backbone network, the YOLOv8 model still uses the CSP architecture, dividing the feature map into two parts. The first part uses convolution operations, and the second part is directly connected to the data of the first part. The C2f module of the YOLOv5 model is combined with the ELAN of the YOLOv7 model to design C3f, so that the model can obtain rich gradient flow information.

[0097] In the neck network, the network layers are deepened to obtain richer feature information, and the feature pyramid network (FPN) and path aggregation network (PAN) are combined to realize multi-scale feature fusion (FPN-PAN), which works better in dense detection tasks. However, deeper networks often bring more convolution operations, which will lead to serious loss of object location information and feature information.

[0098] In the head structure, YOLOX's decoupling head is used. The classification head and regression head no longer share parameters, and only the classification and regression branches are retained. At the same time, the Anchor-Free detection method is adopted to determine the central area of ​​the detected object and estimate the distance to the bounding box. In the calculation of the loss function, the cross entropy loss function (DistributionFocalLoss) is used.

[0099] Among the current target detection algorithms for on-site safety wearable management and control issues in substations, YOLOv8 is a newer algorithm in the YOLO series of real-time object detectors. It has cutting-edge performance in terms of accuracy and speed. However, the YOLOv8n basic model has a high number of parameters and complexity, and the detection model still has the problem of high computational complexity, which is not conducive to terminal deployment tasks.

[0100] Embodiment 1

[0101] In order to solve the above technical problems, the present invention provides a target recognition model, a method, a system and a medium for local target recognition in the station area, which can effectively improve the recognition speed of target objects on site in the station area based on a lightweight target recognition model.

[0102] Please refer to Figure 1 , Figure 1 A schematic diagram of the structure of an embodiment of a target recognition model provided by the present invention.

[0103] like Figure 1 As shown, the target recognition model is constructed based on the YOLOv8n algorithm and includes:

[0104] a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a fifth CBS module, a sixth CBS module, a seventh CBS module, a first PConvC2fEMA module, a second PConvC2fEMA module, a third PConvC2fEMA module, a fourth PConvC2fEMA module, an SPPF module, a first Upsample module, a second Upsample module, a first Concat module, a second Concat module, a third Concat module, a fourth Concat module, a first VoVGSCSP module, a second VoVGSCSP module, a third VoVGSCSP module, a fourth VoVGSCSP module, a first output module, a second output module, and a third output module; wherein:

[0105] The first CBS module serves as an input end of the target recognition model, and is used to input an image to be recognized, and its output end is connected to the input end of the second CBS module;

[0106] The output end of the second CBS module is connected to the input end of the first PConvC2fEMA module;

[0107] The output end of the first PConvC2fEMA module is connected to the input end of the third CBS module;

[0108] The output end of the third CBS module is connected to the input end of the second PConvC2fEMA module;

[0109] The first output terminal of the second PConvC2fEMA module is connected to the input terminal of the fourth CBS module; the second output terminal of the second PConvC2fEMA module is connected to the first input terminal of the second Concat module;

[0110] The output end of the fourth CBS module is connected to the input end of the third PConvC2fEMA module;

[0111] The first output end of the third PConvC2fEMA module is connected to the input end of the fifth CBS module; the second output end of the third PConvC2fEMA module is connected to the first input end of the first Concat module;

[0112] The output end of the fifth CBS module is connected to the input end of the fourth PConvC2fEMA module;

[0113] The output end of the fourth PConvC2fEMA module is connected to the input end of the SPPF module;

[0114] The output end of the SPPF module is connected to the input end of the first Upsample module and the first input end of the fourth Concat module;

[0115] The output end of the first Upsample module is connected to the second input end of the first Concat module;

[0116] The output end of the first Concat module is connected to the input end of the first VoVGSCSP module;

[0117] The first output terminal of the first VoVGSCSP module is connected to the input terminal of the second Upsample module; the second output terminal of the first VoVGSCSP module is connected to the first input terminal of the third Concat module;

[0118] The output end of the second Upsample module is connected to the second input end of the second Concat module;

[0119] The output end of the second Concat module is connected to the input end of the second VoVGSCSP module;

[0120] The first output terminal of the second VoVGSCSP module is connected to the input terminal of the sixth CBS module;

[0121] The output end of the sixth CBS module is connected to the second input end of the third Concat module;

[0122] The output end of the third Concat module is connected to the input end of the third VoVGSCSP module;

[0123] The output end of the third VoVGSCSP module is connected to the input end of the seventh CBS module;

[0124] The output end of the seventh CBS module is connected to the second input end of the fourth Concat module;

[0125] The first output terminal of the fourth Concat module is connected to the input terminal of the fourth VoVGSCSP module;

[0126] The second output end of the fourth Concat module is connected to the input end of the first output module;

[0127] The output end of the third Concat module is connected to the input end of the second output module;

[0128] The output end of the second VoVGSCSP module is connected to the input end of the third output module;

[0129] The first output module, the second output module and the third output module serve as output ends of the target recognition model and are used to output recognition results.

[0130] Compared with the prior art that uses the YOLOv8n basic model with high parameter quantity and complexity for target recognition, the embodiment of the present invention constructs a target recognition model based on the YOLOv8n algorithm, specifically uses the PConvC2fEMA module to improve the backbone network of the YOLOv8n model to achieve the reduction of model parameter quantity and calculation amount, and enhances the focus on target representation to improve the accuracy of model detection; and uses the VoVGSCSP module to improve the neck network of the YOLOv8n model, which can improve the nonlinear expression ability of the neck network; the present invention performs lightweight improvement on the YOLOv8n model to obtain a target recognition model, and compared with the YOLOv8n model, the target recognition model has a smaller scale and a smaller calculation amount, can have a faster target recognition speed, and still have a higher recognition accuracy.

[0131] Please refer to the following Figure 2 , Figure 2 A schematic diagram of a flow chart of an embodiment of the PConvC2fEMA module provided by the present invention.

[0132] Furthermore, if Figure 2 As shown, the first PConvC2fEMA module, the second PConvC2fEMA module, the third PConvC2fEMA module and the fourth PConvC2fEMA module have the same composition; the first PConvC2fEMA module includes: a first input unit, a first CBS unit, a second CBS unit, a split unit, a first PConvBlock unit, a second PConvBlock unit, a third PConvBlock unit, a first Concat unit, an EMA attention mechanism unit and a first output unit; wherein:

[0133] The input end of the first input unit is used as the input end of the first PConvC2fEMA module;

[0134] The output end of the first input unit is connected to the input end of the first CBS unit;

[0135] The output end of the first CBS unit is connected to the input end of the split unit;

[0136] The first output end of the split unit is connected to the input end of the first PConvBlock unit and the first input end of the first Concat unit; the second output end of the split unit is connected to the first input end of the first Concat unit;

[0137] The first output terminal of the first PConvBlock unit is connected to the second input terminal of the first Concat unit;

[0138] The second output terminal of the first PConvBlock unit is connected to the input terminal of the second PConvBlock unit;

[0139] The first output end of the second PConvBlock unit is connected to the input end of the third PConvBlock unit; the second output end of the second PConvBlock unit is connected to the third input end of the first Concat unit;

[0140] The output end of the third PConvBlock unit is connected to the third input end of the first Concat unit;

[0141] The output end of the first Concat unit is connected to the input end of the second CBS unit;

[0142] The output end of the second CBS unit is connected to the input end of the EMA attention mechanism unit;

[0143] The output end of the EMA attention mechanism unit is connected to the input end of the first output unit;

[0144] The output end of the first output unit serves as the output end of the first PConvC2fEMA module.

[0145] In the real on-site scene of the substation, the environmental background is relatively complex, which makes the detection of helmet wearing easily interfered by the complex background, resulting in false detection, missed detection and other situations in the detection task. The EMA attention mechanism is introduced in the PConvC2fEMA module of this embodiment. The EMA attention mechanism generates spatial attention weight values ​​through cross-space learning, integrates multi-scale spatial information and position information, can capture the pixel-level pairwise relationship of the feature map and highlight the global context information of the pixel without increasing the number of parameters and calculations. The input feature map information is fully integrated to make the number of parameters of the PConvC2fEMA module smaller, further reduce the number of parameters of the backbone network in the model, and achieve the goal of lightweight model. Moreover, after the introduction of the EMA attention mechanism, the interference of the complex background can be weakened, so that the target recognition model pays more attention to the task of whether to wear a helmet to avoid false detection, missed detection and other situations, and improves the recognition precision and accuracy of the target recognition model.

[0146] Please refer to the following Figure 3 , Figure 3 A schematic structural diagram of an embodiment of a PConvBlock unit provided by the present invention.

[0147] Furthermore, if Figure 3 As shown, the first PConvBlock unit, the second PConvBlock unit and the third PConvBlock unit have the same composition; the first PConvBlock unit includes: an input subunit, a PConv subunit, a BN subunit, a SiLU subunit, a first Conv subunit, a second Conv subunit, and an output subunit; wherein:

[0148] The input end of the input subunit serves as the input end of the first PConvBlock unit;

[0149] The output end of the input subunit is connected to the input end of the PConv subunit;

[0150] The output end of the PConv subunit is connected to the input end of the first Conv subunit;

[0151] The output end of the first Conv subunit is connected to the input end of the BN subunit;

[0152] The output end of the BN subunit is connected to the input end of the SiLU subunit;

[0153] The output end of the SiLU subunit is connected to the input end of the second Conv subunit;

[0154] The output end of the second Conv subunit and the output end of the input subunit are matrix-summed and then connected to the input end of the output subunit;

[0155] The output end of the output subunit serves as the output end of the first PConvBlock unit.

[0156] In this embodiment, the convolution kernel size of the first Conv subunit and the second Conv subunit is 1*1.

[0157] The Bottleneck in the C2f module in the YOLOv8n model is composed of two consecutive CBS modules. The CBS module consists of three parts: ordinary convolution, batch normalization, and SiLU activation function. The CBS module has the problems of large number of parameters and high computational complexity. Therefore, this embodiment proposes a low-complexity and lightweight PConvBlock unit.

[0158] In this embodiment, the PConvBlock unit includes a PConv subunit and two Conv subunits, wherein the middle layer uses a normalization layer, and introduces a SiLU subunit based on the SiLU activation function, and replaces the ReLU activation function with the SiLU activation function. The PConvBlock unit uses the inverted residual idea in MobileNetV2, expands the number of channels in the middle layer, and achieves a significant reduction in the number of network parameters by reducing the number of convolution operations, thereby accelerating the reasoning speed; the PConvBlock unit in this embodiment has low complexity and small number of parameters, which can reduce the number of parameters of the PConvC2fEMA module. Replacing all C2f modules of the Backbone network in YOLOv8n with PConvC2fEMA modules can reduce the number of parameters of the backbone network, further reduce the number of parameters and computational complexity of the target recognition model, and achieve the goal of lightweight target recognition model.

[0159] In this embodiment, PConv has higher FLOPS than other types of convolution units such as DWConv, and PConv can better utilize the computing power on the device.

[0160] In this embodiment, one way to construct the target recognition model is to construct a PConvC2f module based on the PConvBlock unit, replace the Bottleneck in the C2f module of the original YOLOv8n model with the PConvBlock module; add an EMA attention mechanism unit based on the EMA attention mechanism to the end of the PConvC2f module to construct a PConvC2fEMA module.

[0161] In this embodiment, the EMA attention mechanism can generate a high-level feature map of pixel-level attention without reducing the channel dimension in the convolution operation. The EMA attention mechanism follows the design strategy of the CA (Coordinate Attention) attention mechanism (spatial attention mechanism), embeds the position information into the channel attention, and decomposes the channel attention into a one-dimensional feature encoding process that aggregates features along two different directions. One-dimensional global average pooling is performed in the horizontal and vertical directions respectively. Different from the channel attention, the input is converted into a single feature vector using two-dimensional global pooling; the EMA attention mechanism decomposes the channel attention into feature encodings of two one-dimensional vectors, captures long-distance dependencies in the vertical direction, and retains precise position information in the horizontal direction. At the same time, a parallel branch with a convolution kernel size of 3x3 is added to aggregate multi-scale spatial structure information. They can be applied to the input feature map complementarily, effectively establish long-term and short-term dependencies, and enhance the attention to the representation of objects of interest.

[0162] Please refer to Figure 4 , Figure 4Schematic diagram of the EMA attention mechanism provided by the present invention.

[0163] Specifically, the EMA attention mechanism is mainly divided into the following three parts:

[0164] 1. Feature grouping, given the size Input∈R C×H×W ,EMA attention mechanism divides the direction of the Input across the channel dimension into G sub-features, learns the regions of interest in different sub-features, and strengthens the feature representation.

[0165] 2. Parallel subnetwork,EMA attention mechanism uses three parallel routes to extract group feature maps. In the information encoding stage, the first branch uses a pooling kernel of size (1,W) in the vertical coordinate direction to encode each channel. The specific calculation formula is as follows:

[0166]

[0167] in, It represents the output of the cth channel with a width of w. It only processes the vertical direction and does not change the horizontal direction.

[0168] The second branch encodes each channel horizontally using a pooling kernel of size (H, 1), which is calculated as follows:

[0169]

[0170] in, Indicates the output of the cth channel with height h. Only the horizontal direction is processed, and the vertical direction remains unchanged.

[0171] The feature maps in the vertical and horizontal directions are fused, and a convolution with a convolution kernel size of 1×1 is used to communicate cross-channel information. The specific calculation formula is as follows:

[0172] f=P([Z h , Z w ])

[0173] In the above formula, f represents the feature map after fusion of vertical and horizontal directions, where f∈R C / / G×(H+W) , G represents the division ratio, and P represents the convolution with a convolution kernel size of 1×1, which is used to transform the number of channels.

[0174] The fused feature map f is further decomposed into two independent tensors along the channel dimension, with the vertical direction being fw∈R C / / G×H×1 , the horizontal direction is f w ∈R C / / G×1×w , and then use the sigmoid activation function. The calculation formula is as follows:

[0175] g h =σ(f h )

[0176] g w =σ(f w )

[0177] Among them, σ represents the sigmoid activation function, g w and g h The value of is used as the intermediate weight of the EMA attention mechanism; the third branch is a normal convolution with a convolution kernel size of 3×3.

[0178] 3. Cross-space learning stage

[0179] Use 2D global average pooling to act on the processed horizontal and vertical feature map information, and transform the output dimension of the smallest branch to match it before the joint activation of the feature channels; the operation of 2D global pooling is shown in the formula. The input size is H×W×C and the output size is 1×1×C. The process is as follows:

[0180]

[0181] Among them, Z c Describe the output related to the cth channel, H and W represent the height and width of the image respectively, x c (i, j) represents the input of the cth channel with height i and width j; after 2D global average pooling, the Softmax nonlinear activation function is used to adapt to the linear transformation, and a spatial attention map is obtained by multiplying the output of the Softmax function of parallel processing, which integrates the spatial information of different scales in the same stage; similarly, the ordinary convolution with a convolution kernel size of 3*3 is extracted using 2D global average pooling and Softmax nonlinear activation function to generate another spatial attention map, which accurately records the spatial position information, aggregates the information of the generated spatial attention weight value, and finally uses the Sigmoid nonlinear activation function.

[0182] Please refer to the following Figure 5 , Figure 5 A schematic structural diagram of an embodiment of the VoVGSCSP module provided by the present invention.

[0183] Furthermore, if Figure 4As shown, the first VoVGSCSP module, the second VoVGSCSP module, the third VoVGSCSP module and the fourth VoVGSCSP module have the same composition; the first VoVGSCSP module includes: a second input unit, a first Conv unit, a second Conv unit, a third Conv unit, a GSBottleneck unit, a second Concat unit and a second output unit; wherein:

[0184] The input end of the second input unit serves as the input end of the first VoVGSCSP module;

[0185] The output end of the second input unit is connected to the input end of the first Conv unit and the input end of the second Conv unit;

[0186] The output end of the second Conv unit is connected to the first input end of the second Concat unit;

[0187] The output end of the first Conv unit is connected to the input end of the GSBottleneck unit;

[0188] The output end of the GSBottleneck unit is connected to the second input end of the second Concat unit;

[0189] The output end of the second Concat unit is connected to the input end of the third Conv unit;

[0190] The output end of the third Conv unit is connected to the input end of the second output unit;

[0191] The output end of the second output unit serves as the output end of the first VoVGSCSP module.

[0192] In this embodiment, the aggregation method cross-stage local network module, namely the VoVGSCSP module, is composed of a GSBottleneck module, which is composed of a GSConv basic module. GSConv retains the connection between channels to the maximum extent with low time complexity and has less redundant and repeated feature map information; the GSConv module is composed of a standard convolution (Conv), a depthwise separable convolution (DWConv) and a feature map shuffle, and the DWConv and shuffle operations enhance the nonlinear expression ability; the GSBottleneck module combines the residual idea on the basis of GSConv; the design of the VoVGSCSP module is based on the GSBottleneck module, uses two branches to extract features, aggregates the feature maps of the branches, uses convolution to extract information from the aggregated multi-channel feature maps, obtains richer feature information, combines the cross-stage operation of the residual class, and improves the nonlinear expression ability of the neck network of the target recognition model.

[0193] In this embodiment, in order to solve the problem that small distant targets are difficult to detect in the helmet detection task, in the neck network of the target recognition model, the VoVGSCSP module replaces the C2f module in the neck module in YOLOV8n to fully integrate the feature map information extracted by the backbone network; wherein the VoVGSCSP module is composed of a number of Conv units, GSBottleneck units, Concat units and input-output units, and extracts features through two branches of the first Conv unit and the second Conv unit, and then aggregates the feature maps of the branches, uses convolution to extract information from the aggregated multi-channel feature maps, and obtains richer feature information, and combines the cross-stage operation of the residual class to further improve the nonlinear expression ability of the neck network in the target recognition model, realize the full integration of the feature map information extracted by the backbone network, improve the recognition of occluded targets and small targets, and enhance the recognition and detection capabilities of small targets.

[0194] Accordingly, in order to solve the above technical problems, the embodiment of the present invention also provides a method for identifying local targets in a station area, please refer to Figure 6 , Figure 6 The present invention provides a flow chart of an embodiment of a method for local target identification in a substation area.

[0195] like Figure 6 As shown, the method for identifying local targets in the substation area includes steps 101 to 104, which are specifically:

[0196] Step 101: constructing a target recognition model as described in any one of the above items;

[0197] Step 102: Collecting image data of the station area on site, and obtaining a helmet data set according to the image data;

[0198] Step 103: training and testing the target recognition model based on the helmet data set to obtain an optimal recognition model;

[0199] Step 104: Use the optimal recognition model to identify safety helmets on site in the workstation area.

[0200] The on-site target recognition method in the substation proposed in the present invention first constructs the target recognition model obtained based on the YOLOv8n model mentioned above, and trains and tests the target recognition model according to the obtained safety helmet data set to enable it to learn the characteristic data of the safety helmets in the substation, obtain the optimal recognition model, and use the optimal recognition model to perform on-site safety helmet recognition in the substation. The present invention performs on-site safety helmet recognition in the substation based on the target recognition model, which can improve the speed of on-site safety helmet recognition in the substation and ensure better recognition accuracy.

[0201] Furthermore, in step 102, collecting the local image data of the station area includes:

[0202] Collect on-site image data of the station area from several camera angles.

[0203] This embodiment collects on-site image data of the station area from different camera angles to make the data in the subsequent data set more diversified. When the target recognition model is subsequently trained and tested, the generalization ability of the obtained optimal recognition model can be enhanced to reduce the risk of overfitting.

[0204] Further, in step 102, obtaining a helmet dataset according to the image data includes:

[0205] Preprocessing the image data to obtain an initial data set;

[0206] The initial data set is annotated based on a preset annotation type to obtain a helmet data set.

[0207] This embodiment preprocesses the acquired image data to ensure the accuracy and reliability of the data so that the acquired data can be reliably used for subsequent model training and testing; and this embodiment labels the initial data set based on a preset labeling type to obtain a safety helmet data set, so that the subsequent target recognition model can learn the safety helmet features during model training, thereby enabling the target recognition model to accurately identify unknown data.

[0208] In this embodiment, the video of personnel wearing helmets in actual work is collected through cameras deployed in the on-site environment of the station, and images are collected on-site from different camera angles as much as possible. The collected video images are quality checked and preprocessed, and poor quality images such as black screen, blurred, overexposed and jittery images in the video images are deleted. In this way, the collected data set includes a variety of different on-site environments, which improves the diversity of training data and further enhances the generalization of the training model.

[0209] In this embodiment, the preprocessed initial data set can be annotated using an annotation tool; the annotated samples are divided into two categories, one is people who normally wear safety helmets, and the other is people who do not wear safety helmets, where those wearing safety helmets include safety helmets of different colors; in an implementation case of the present invention, the annotated data sample statistics total 12,267 pictures and label files, and the ratio of samples wearing and not wearing safety helmets is approximately 6:4. The annotation labels use "hat" to indicate wearing a safety helmet and "person" to indicate not wearing a safety helmet. Finally, the safety helmet data set is divided into a training set, a validation set, and a test set according to a preset ratio; the preset ratio can be but is not limited to 7:2:1.

[0210] The method for local target recognition in the area of ​​the present embodiment, before step 103, training and testing the target recognition model based on the helmet data set to obtain the optimal recognition model, further comprises:

[0211] Set the environment configuration and hyperparameters for training the object recognition model.

[0212] In this embodiment, the environment configuration used is shown in Table 1:

[0213] Table 1

[0214] Environment Name Specific configuration operating system Ubuntu 20.04 Graphics NVIDIA GeForce RTX 4090 Deep Learning Environment Pytorch 1.12.0 Software learning environment Anaconda, CUDA11.1, Python3.7

[0215] In this embodiment, when training the target recognition model, the hyper parameters set are as shown in Table 2:

[0216] Table 2

[0217] Parameter name Configuration Optimizer Adam Image resolution 640*640 Batchsize 32 Set the initial learning rate 0.001 Final learning rate 0.00003

[0218] Further, in step 103, the training and testing of the target recognition model based on the helmet data set to obtain the optimal recognition model includes:

[0219] The target recognition model is trained and tested based on the helmet data set until the loss on the validation set meets the preset conditions, thereby obtaining the optimal recognition model.

[0220] In this embodiment, the preset condition is that the loss of the landmark recognition model on the validation set does not reach a new minimum value for 10 consecutive epochs.

[0221] In this embodiment, the strategy for training the target recognition model may be but is not limited to the early stopping method, which is a strategy used to prevent overfitting of machine learning models. During the model training process, the performance of the model on the training set will generally improve with the increase of training time, but the performance on the validation set may begin to decline after a certain point because the model begins to learn the noise and specific features in the training data, thereby reducing its generalization ability. The goal of the early stopping method is to stop training before the generalization ability of the model decreases to obtain the best model performance. When the preset conditions are triggered, it is believed that the model may have entered the overfitting stage, at which time early stopping will be triggered to complete the model training.

[0222] In this embodiment, the trained model needs to be evaluated using evaluation indicators to verify the accuracy of the model. The present invention uses intersection over union, precision, recall, AP and mean average precision (mAP) to evaluate the trained model; AP is the area under the PR curve of a certain category.

[0223] The calculation formula of intersection-over-union ratio is:

[0224] IoU = TP / (TP+FN+FP)

[0225] The calculation formula for accuracy is:

[0226] Precision=TP / (TP+FP)=TP / all detections

[0227] The calculation formula for recall is:

[0228] Recall=TP / (TP+FN)=TP / all ground trusts

[0229] Among them, TP is the area predicted correctly by the algorithm detection box; FP is the area predicted incorrectly by the algorithm detection box; FN is the area where the actual annotation box is correct but the algorithm detection box is not predicted; all detections is the area predicted by the algorithm detection detection; all ground trusts is the area of ​​the actual annotation box;

[0230] The calculation formula of AP is:

[0231]

[0232] The average precision calculation formula is:

[0233]

[0234] Among them, mAP is the mean of AP values ​​under all categories.

[0235] Accordingly, in order to solve the above technical problems, an embodiment of the present invention further provides a system for local target recognition in a substation area, comprising: a model building module, a data acquisition module, a model training module and a target recognition module;

[0236] Wherein, the model building module is used to build any of the target recognition models described above;

[0237] The data acquisition module is used to collect image data on the spot of the platform area, and obtain a helmet data set based on the image data;

[0238] The model training module is used to train and test the target recognition model based on the helmet data set to obtain the optimal recognition model;

[0239] The target recognition module is used to use the optimal recognition model to perform on-site safety helmet recognition in the workstation area.

[0240] An embodiment of the present invention further provides a storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the method for identifying an on-site safety helmet in a workstation area as described in any one of the above items are implemented.

[0241] Compared with the prior art, the embodiment of the present invention aims at the problem of on-site safety wear management in the station area. In order to better adapt to the scenario of edge devices, a lightweight PConvC2f module is designed to improve the backbone network of YOLOv8n, thereby reducing the number of network parameters and the amount of calculation; an EMA attention mechanism is introduced into the PConvC2f module to integrate spatial dependency and position information, establish long-term and short-term dependencies, and enhance the focus on target representation to improve the accuracy of model detection; a VoVGSCSP module is used to improve the neck network to improve the recognition of occluded targets and small targets; compared with the YOLOv8n algorithm, the present invention reduces the model size and reduces the amount of calculation.

[0242] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A target recognition model, characterized in that: It is constructed based on the YOLOv8n algorithm, including: a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a fifth CBS module, a sixth CBS module, a seventh CBS module, a first PConvC2fEMA module, a second PConvC2fEMA module, a third PConvC2fEMA module, a fourth PConvC2fEMA module, an SPPF module, a first Upsample module, a second Upsample module, a first Concat module, a second Concat module, a third Concat module, a fourth Concat module, a first VoVGSCSP module, a second VoVGSCSP module, a third VoVGSCSP module, a fourth VoVGSCSP module, a first output module, a second output module and a third output module; wherein: The first CBS module serves as an input end of the target recognition model, and is used to input an image to be recognized, and its output end is connected to the input end of the second CBS module; The output end of the second CBS module is connected to the input end of the first PConvC2fEMA module; The output end of the first PConvC2fEMA module is connected to the input end of the third CBS module; The output end of the third CBS module is connected to the input end of the second PConvC2fEMA module; The first output terminal of the second PConvC2fEMA module is connected to the input terminal of the fourth CBS module; the second output terminal of the second PConvC2fEMA module is connected to the first input terminal of the second Concat module; The output end of the fourth CBS module is connected to the input end of the third PConvC2fEMA module; The first output end of the third PConvC2fEMA module is connected to the input end of the fifth CBS module; the second output end of the third PConvC2fEMA module is connected to the first input end of the first Concat module; The output end of the fifth CBS module is connected to the input end of the fourth PConvC2fEMA module; The output end of the fourth PConvC2fEMA module is connected to the input end of the SPPF module; The output end of the SPPF module is connected to the input end of the first Upsample module and the first input end of the fourth Concat module; The output end of the first Upsample module is connected to the second input end of the first Concat module; The output end of the first Concat module is connected to the input end of the first VoVGSCSP module; The first output terminal of the first VoVGSCSP module is connected to the input terminal of the second Upsample module; the second output terminal of the first VoVGSCSP module is connected to the first input terminal of the third Concat module; The output end of the second Upsample module is connected to the second input end of the second Concat module; The output end of the second Concat module is connected to the input end of the second VoVGSCSP module; The first output terminal of the second VoVGSCSP module is connected to the input terminal of the sixth CBS module; The output end of the sixth CBS module is connected to the second input end of the third Concat module; The output end of the third Concat module is connected to the input end of the third VoVGSCSP module; The output end of the third VoVGSCSP module is connected to the input end of the seventh CBS module; The output end of the seventh CBS module is connected to the second input end of the fourth Concat module; The first output terminal of the fourth Concat module is connected to the input terminal of the fourth VoVGSCSP module; The second output end of the fourth Concat module is connected to the input end of the first output module; The output end of the third Concat module is connected to the input end of the second output module; The output end of the second VoVGSCSP module is connected to the input end of the third output module; The first output module, the second output module and the third output module serve as output ends of the target recognition model and are used to output recognition results.

2. A target recognition model as claimed in claim 1, characterized in that: The first PConvC2fEMA module, the second PConvC2fEMA module, the third PConvC2fEMA module and the fourth PConvC2fEMA module have the same composition; the first PConvC2fEMA module includes: a first input unit, a first CBS unit, a second CBS unit, a split unit, a first PConvBlock unit, a second PConvBlock unit, a third PConvBlock unit, a first Concat unit, an EMA attention mechanism unit and a first output unit; wherein: The input end of the first input unit is used as the input end of the first PConvC2fEMA module; The output end of the first input unit is connected to the input end of the first CBS unit; The output end of the first CBS unit is connected to the input end of the split unit; The first output end of the split unit is connected to the input end of the first PConvBlock unit and the first input end of the first Concat unit; the second output end of the split unit is connected to the first input end of the first Concat unit; The first output terminal of the first PConvBlock unit is connected to the second input terminal of the first Concat unit; The second output terminal of the first PConvBlock unit is connected to the input terminal of the second PConvBlock unit; The first output end of the second PConvBlock unit is connected to the input end of the third PConvBlock unit; the second output end of the second PConvBlock unit is connected to the third input end of the first Concat unit; The output end of the third PConvBlock unit is connected to the third input end of the first Concat unit; The output end of the first Concat unit is connected to the input end of the second CBS unit; The output end of the second CBS unit is connected to the input end of the EMA attention mechanism unit; The output end of the EMA attention mechanism unit is connected to the input end of the first output unit; The output end of the first output unit serves as the output end of the first PConvC2fEMA module.

3. A target recognition model as claimed in claim 2, characterized in that: The first PConvBlock unit, the second PConvBlock unit and the third PConvBlock unit have the same composition; the first PConvBlock unit includes: an input subunit, a PConv subunit, a BN subunit, a SiLU subunit, a first Conv subunit, a second Conv subunit, and an output subunit; wherein: The input end of the input subunit serves as the input end of the first PConvBlock unit; The output end of the input subunit is connected to the input end of the PConv subunit; The output end of the PConv subunit is connected to the input end of the first Conv subunit; The output end of the first Conv subunit is connected to the input end of the BN subunit; The output end of the BN subunit is connected to the input end of the SiLU subunit; The output end of the SiLU subunit is connected to the input end of the second Conv subunit; The output end of the second Conv subunit and the output end of the input subunit are matrix-summed and then connected to the input end of the output subunit; The output end of the output subunit serves as the output end of the first PConvBlock unit.

4. A target recognition model as claimed in claim 1, characterized in that: The first VoVGSCSP module, the second VoVGSCSP module, the third VoVGSCSP module and the fourth VoVGSCSP module have the same composition; the first VoVGSCSP module includes: a second input unit, a first Conv unit, a second Conv unit, a third Conv unit, a GSBottleneck unit, a second Concat unit and a second output unit; wherein: The input end of the second input unit serves as the input end of the first VoVGSCSP module; The output end of the second input unit is connected to the input end of the first Conv unit and the input end of the second Conv unit; The output end of the second Conv unit is connected to the first input end of the second Concat unit; The output end of the first Conv unit is connected to the input end of the GSBottleneck unit; The output end of the GSBottleneck unit is connected to the second input end of the second Concat unit; The output end of the second Concat unit is connected to the input end of the third Conv unit; The output end of the third Conv unit is connected to the input end of the second output unit; The output end of the second output unit serves as the output end of the first VoVGSCSP module.

5. A method for identifying local targets in a substation area, characterized in that: include: Constructing a target recognition model as described in any one of claims 1 to 4; Collecting on-site image data of the station area, and obtaining a helmet data set based on the image data; Train and test the object recognition model based on the helmet dataset to obtain the optimal recognition model; The optimal recognition model is used to perform on-site safety helmet recognition in the station area.

6. A method for identifying on-site targets in a station area as claimed in claim 5, characterized in that: The collecting of local image data of the station area includes: Collect on-site image data of the station area from several camera angles.

7. A method for identifying on-site targets in a station area as claimed in claim 5, characterized in that: The step of obtaining a helmet dataset according to the image data comprises: Preprocessing the image data to obtain an initial data set; The initial data set is annotated based on a preset annotation type to obtain a helmet data set.

8. A method for identifying on-site targets in a station area as claimed in claim 5, characterized in that: The method of training and testing the target recognition model based on the helmet data set to obtain the optimal recognition model includes: The target recognition model is trained and tested based on the helmet data set until the loss of the target recognition model on the validation set meets the preset conditions, thereby obtaining the optimal recognition model.

9. A local target recognition system in a substation area, characterized in that: include: Model building module, data acquisition module, model training module and target recognition module; Wherein, the model building module is used to build the target recognition model as claimed in any one of claims 1 to 4; The data acquisition module is used to collect image data on the spot of the platform area, and obtain a helmet data set based on the image data; The model training module is used to train and test the target recognition model based on the helmet data set to obtain the optimal recognition model; The target recognition module is used to use the optimal recognition model to perform on-site safety helmet recognition in the workstation area.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying on-site safety helmets in the workstation area as described in any one of claims 5 to 8 are implemented.

Citation Information

Cited By

  • Small target detection method and system based on improved YOLOv8

    CN119888365A