Lightweight improvement-based L-YOLOv8s insulator defect detection method

By adding MobileNetv3-ECA-SPPF backbone network and Slim-Neck module to the YOLOv8s model, and using WIoUv3 border loss function, the L-YOLOv8s model was solved, and the large size of the target detection algorithm in drone inspection was achieved. High-precision and fast insulator defect detection, suitable for real-time detection of insulators on transmission lines.

CN119006427AInactive Publication Date: 2024-11-22CHINA THREE GORGES UNIV

Patent Information

Application Number
CN202411152989.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-11-22
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing target detection algorithm is too large during drone inspection to meet the needs of lightweight processing, resulting in the inability to effectively detect real-time detection of insulator defects in transmission lines.

Method used

By adding MobileNetv3-ECA-SPPF backbone network, Slim-Neck module and WIoUv3 border loss function to the YOLOv8s model, lightweight optimization is carried out to build the L-YOLOv8s model, reducing model parameters and calculation complexity, and improving detection speed and accuracy.

Benefits of technology

It realizes high-precision and fast insulator defect detection, can meet the lightweight requirements of drone inspection, and is suitable for real-time detection of insulators on aerial photos of transmission lines.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006427A_ABST
    Figure CN119006427A_ABST
Patent Text Reader

Abstract

The invention provides an L-YOLOv8s insulator defect detection method based on light weight improvement. The method comprises the following steps: acquiring a defect image of the insulator of the power transmission line through an unmanned aerial vehicle; marking the acquired image by using Labelimg to obtain an insulator defect data set, and dividing the insulator defect data set; the method comprises the following steps: adding a MobileNetV3, an ECANet attention mechanism, a Slim-Neck structure and a WioUv3 loss function into a YOLOv8s model, so as to obtain an L-YOLOv8s model; the sorted data set is put into an L-YOLOv8s model to be trained, and a trained model is obtained; and finally, putting a to-be-detected image into the model for positioning and detecting. The method is low in parameter quantity and high in detection speed, can be embedded into an inspection unmanned aerial vehicle, can meet the requirements of real-time detection of power transmission line insulators, and improves the accuracy and speed of insulator defect detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of power transmission line inspection, and in particular to a lightweight and improved L-YOLOv8s insulator defect detection method. Background Art

[0002] Insulators are an indispensable part of the power system and play an important role in electrical isolation and mechanical support. Different types of insulators will produce different types of defects under load or under the condition of external force intervention. Therefore, regular inspections of transmission lines and elimination of insulator defect faults are of great practical significance for improving the safety and stability of the power system.

[0003] At present, the use of drones embedded with target detection algorithms to inspect transmission lines has become a mainstream method. However, most target detection algorithms are too large to meet the conditions for drones to carry them. The algorithms need to be lightweight to meet the requirements of drone inspections. In recent years, with the continuous development of deep learning in the field of computer vision, compared with traditional methods, the use of target detection technology in the field of computer vision to identify and detect defects in transmission line insulators has higher accuracy and real-time performance. At present, the target detection algorithms based on deep learning are mainly two-stage algorithms represented by Faster R-CNN, Mask R-CNN, etc. and one-stage algorithms represented by SSD and YOLO. As the most advanced single-stage target detection algorithm in the YOLO series model, YOLOv8 is capable of performing transmission line insulator defect inspection tasks and meets the conditions for real-time detection of transmission line aerial insulators. Summary of the invention

[0004] The purpose of the present invention is to provide an insulator defect detection method based on lightweight improvement of L-YOLOv8s, which improves detection accuracy, speeds up detection, and reduces model parameters. By deploying it on other mobile devices such as drones, application tests are carried out to further study its adaptability and practicality, so as to meet the conditions for real-time detection of aerial insulators of transmission lines.

[0005] To achieve the above object, the present invention is achieved by: The invention provides a method for detecting insulator defects based on lightweight improved L-YOLOv8s. The method uses YOLOv8s as the benchmark model, uses the MobileNetv3-ECA-SPPF module to perform lightweight optimization on the backbone network, adds the Slim-Neck module to the feature fusion part, and finally uses WIoUv3 to optimize the border loss function in the head network. The experimental results show that the model proposed by the invention has high detection accuracy, fast detection speed, and small model parameters, and can meet the conditions for real-time detection of aerial photography insulators of transmission lines.

[0006] A method for detecting insulator defects based on lightweight improved L-YOLOv8s, characterized in that it comprises the following steps: S1. Collect images of transmission line insulators through drones and sort them manually; S2. Labelimg is used to annotate the screened and sorted transmission line insulator images to obtain an insulator defect dataset, which is then divided into a training set, a test set, and a validation set according to a certain ratio; S3, combine MobileNetV3, ECANet attention mechanism, Slim-Neck module, and WioUv3 border loss function into the YOLOv8s model to obtain the L-YOLOv8s model; S4. Put the sorted data set into the L-YOLOv8s model for training to obtain a trained model; S5. Use the trained model to locate and identify the image to be detected.

[0007] In S1, the specific steps are as follows: The use of drones to collect images of transmission line insulator defects requires that some insulator images have certain occlusions, small targets, and dark light problems, which can increase the generalization performance of the model of the present invention. Finally, images with a resolution greater than 640*640 are manually screened as training images of the present invention.

[0008] In S2, the specific steps are as follows: The Labelimg tool is used to annotate the insulator defect images obtained in S1. The VOC data format is selected as the format, where glass insulators are annotated as Glass insulators, composite insulators are annotated as Composite insulators, and porcelain insulators are annotated as Porcelain insulators. The defect types are divided into damage and flashover, where broken insulators are annotated as Broken insulators and flashover insulators are annotated as Flashover damage insulaor. Each label name and its relative position are saved in an .xml file, and then converted into a .txt file for training. The training set, test set, and validation set are then divided into 8:1:1 ratios.

[0009] In S3, the specific steps are as follows: Add the MobileNetV3 module and the ECANet attention mechanism to the YOLOv8s backbone network; add the Slim-Neck module to the feature fusion part; add the WIoUv3 border loss function to the head network to build the L-YOLOv8s model. The specific operations are as follows: 1) Adding MobileNetV3 module and ECANet attention mechanism to the backbone network: YOLOv8s requires a long training time due to its deep backbone network and numerous parameters. To solve the above problems, the present invention proposes a MobileNetv3-ECA-SPPF backbone network for lightweight optimization of the YOLOv8s backbone network, which reduces the number of model parameters and model complexity without losing accuracy, so that it can be applied to mobile devices.

[0010] The MobileNetv3-ECA-SPPF backbone network is completed by the joint action of MobileNetv3, ECA, and SPPF: the main component structure Bneck in the MoblieNetv3 network mainly includes the inverse residual structure of the deep separable convolution and the linear bottleneck. The MoblieNetv3 network includes two different structures, Large and Small. The present invention uses a small structure. In order to adapt to the size of the insulator image in this article, the size of the first layer input is changed to 640×640×3. In the Bneck unit of MobileNetv3-Small, first use 1×1 convolution to increase the dimension of the input feature map, and then input the output feature map into the deep separable convolution with a convolution kernel size of 3×3, and then input the output feature map into the SENet module to obtain the weighted features of the SENet module, and finally use 1×1 convolution to reduce the dimension of the output features, and finally obtain the feature map output by MobileNetv3-Small, which can reduce the number of model parameters and reduce the computational cost. Embedding the ECANet channel attention mechanism in the 12th layer of the MobileNetv3 network can learn the correlation between channels and adaptively adjust the channel weights to improve the performance of the network. ECANet implements channel attention weighting in two steps: one is to extract channel features, and the other is to calculate channel weights. It proposes a local cross-channel interaction strategy without dimensionality reduction and a method for adaptively selecting the size of one-dimensional convolution kernels, thereby achieving performance improvement. To further improve the performance of the model, the original SPPF structure of YOLOv8s is added to this module to achieve local and global feature fusion.

[0011] 2) Add Slim-Neck module to the feature fusion part: Considering the problems of poor image quality of insulators, poor feature extraction effect and slow speed caused by poor light and angle problems during the inspection process, the present invention integrates the Slim-Neck structure into the feature fusion part of YOLOv8s. The Slim-Neck structure is composed of GSConv and VoVGSCSP. Among them, GSConv is composed of three parts: standard convolution, depth-separable convolution and channel shuffling. The main idea of ​​​​the convolution is to infiltrate the information generated by the standard convolution into the information generated by the depth-separable convolution. Through the channel shuffling operation, it can retain the information interaction between channels as much as possible with a lower time loss, enhance the model performance and reduce the number of model parameters. According to the structural definition, it is added to the feature fusion part before the two upsampling operations. In the VoV-GSCSP module, GSConv is used to replace the original traditional convolution. By connecting two GSConvs in series, feature maps of different scales can be spliced ​​to improve the feature diversity of the model. The cross-level part is used to increase the depth and nonlinearity of the network model. VoV-GSCSP replaces the original C2f modules in the 17th, 21st, 24th, and 27th layers.

[0012] 3) Adding WIoUv3 border loss function to the head network: The original YOLOv8 model uses the CIoU border loss function, which can focus on the overlapping area of ​​the border and the center point distance and aspect ratio, making the positioning more accurate. However, once the width and height aspect ratios of the predicted box and the real box are in linear proportion, the CIoU loss function will not work. In order to solve this problem, the present invention introduces a WIoUv3 border loss function with a dynamic non-monotonic mechanism to replace the original CIoU border loss function. The WIoUv3 algorithm uses outliers instead of IoU to evaluate the quality of the anchor frame, and provides a gradient gain allocation strategy, which reduces the competitiveness of high-quality anchor frames and also reduces the harmful gradients generated by low-quality samples. This allows WIoUv3 to focus on anchor frames of general quality. At the same time, because the aspect ratio is not calculated, it is faster. A small outlier means that the anchor frame is of high quality, and a small gradient gain is assigned to it so that the bounding box regression focuses on the anchor frame of ordinary quality. Assigning smaller gradient gains to anchor boxes with larger outliers will effectively prevent low-quality examples from generating large harmful gradients, thereby improving the overall performance of the network model.

[0013] Among them, W g , H gIndicates the width and height of the minimum bounding box, is the normalized distance between the center points of the two bounding boxes. Produces gradients that hinder convergence, Indicates that W g and H g Separated from the computational graph, it effectively eliminates the factors that hinder convergence, and β is the outlier degree.

[0014] The L-YOLOv8s model includes a backbone network, a feature fusion module, and a detection head module. The backbone network includes a CBH module, a MobileNetV3 module, an ECA attention mechanism module, and an SPPF module. The CBH module consists of a standard convolution, normalization, and an H-Swish activation function. MobileNetV3 consists of a 3×3 standard convolution and many Bneck structures. The Bneck structure is used for feature extraction and is an important part of MobileNetv3. It integrates deep separable convolution and residual structures, and introduces a lightweight attention module SENet to improve the ability of feature extraction. This module can reduce the number of model parameters, reduce the computational cost, learn the correlation between channels, and adaptively adjust the weights of channels to improve the performance of the network and achieve local and global feature fusion; the feature fusion module includes a CBS module and a Slim-Neck module. The CBS module is composed of standard convolution, normalization, and SiLU activation functions. The Slim-Neck module includes GSConv and VoV-GSCSP modules, which can retain the information interaction between channels as much as possible with lower time loss, enhance model performance and reduce the number of model parameters, and increase the depth and nonlinearity of the network model. The detection head module adopts a decoupling head and also includes the WIoUv3 border loss function, which can effectively prevent low-quality examples from generating large harmful gradients, thereby improving the overall performance of the network model.

[0015] In S4, the specific steps are as follows: Put the labeled and divided data set into the L-YOLOv8s model for training to get the best.pt file.

[0016] In S5, the specific steps are as follows: Put the image to be detected into the model using the best.pt weight file for positioning and recognition.

[0017] The beneficial effects of the technical solution provided by the present invention are as follows: In view of the problems that may be caused by power transmission line inspection, the present invention proposes a method for detecting insulator defects based on lightweight improved L-YOLOv8s. First, the present invention proposes a lightweight backbone network MobileNetv3-ECA-SPPF, which can reduce the parameter redundancy of the model and improve the detection speed; by using the Slim-Neck module to improve the feature fusion part of YOLOv8s, the amount of calculation and the complexity of the network are reduced, while ensuring the accuracy of the model; for the recognition accuracy of small targets such as damage and flash pollution, the WIoUv3 border loss function replaces the original CIoU border loss function, which is more conducive to the model's recognition efficiency of small targets. Experimental results show that L-YOLOv8s can accurately and quickly identify different types of insulators and their defects. Compared with several traditional models, L-YOLOv8s has practical significance in the identification of insulators and their defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the present invention; Figure 2 Annotate example graphs for the dataset of the present invention; Figure 3 It is a model structure diagram of L-YOLOv8s of the present invention; Figure 4 This is the MobileNetv3-ECA-SPPF backbone network diagram of the present invention; Figure 5 This is a Slim-Neck structure diagram of the present invention; Figure 6 GSConv structure diagram of the present invention; Figure 7 This is a structural diagram of the VoV-GSCSP of the present invention; Figure 8 A comparison chart of the results of the present invention and other traditional detection models; Fig. 9 A comparison chart of the results of the present invention and other traditional detection models; DETAILED DESCRIPTION In order to enable those skilled in the art to more clearly understand the purpose, technical solutions and technical effects of the present invention, the present invention will be fully explained below through specific implementation cases. The present invention is fully described in detail and is subject to conventional techniques in the art.

[0019] Figure 1 This is a specific implementation flow chart of the present invention, and the detailed process is shown in the following steps.

[0020] S1. Collect images of transmission line insulators through drones and sort them manually; The use of drones to collect images of transmission line insulator defects requires that some insulator images have certain occlusions, small targets, and dark light problems, which can increase the generalization performance of the model of the present invention. Finally, images with a resolution greater than 640*640 are manually screened as training images of the present invention.

[0021] S2. Labelimg is used to annotate the screened and sorted transmission line insulator images to obtain an insulator defect dataset, which is then divided into a training set, a test set, and a validation set according to a certain ratio; Use Labelimg tool to label the insulator defect image obtained in S1, such as Figure 2 As shown in the figure, the VOC data format is selected, where glass insulators are marked as Glass insulators, composite insulators are marked as Composite insulators, and porcelain insulators are marked as Porcelain insulators. The defect types are divided into damage and flashover, where broken insulators are marked as Broken insulators and flashover insulators are marked as Flashover damage insulaor. Each label name and its relative position are saved in the .xml file, and then converted into a .txt file for training. Then divide the training set, test set, and validation set in a ratio of 8:1:1.

[0022] S3, combine MobileNetV3, ECANet attention mechanism, Slim-Neck module, and WioUv3 border loss function into the YOLOv8s model to obtain the L-YOLOv8s model, as shown in Figure 3 As shown; 1) Adding MobileNetV3 module and ECANet attention mechanism to the backbone network: YOLOv8s requires a long training time due to its deep backbone network and numerous parameters. To solve the above problems, the present invention proposes a MobileNetv3-ECA-SPPF backbone network for lightweight optimization of the YOLOv8s backbone network, which reduces the number of model parameters and model complexity without losing accuracy, so that it can be applied to mobile devices.

[0023] The MobileNetv3-ECA-SPPF backbone network is completed by the joint action of MobileNetv3, ECA, and SPPF. Figure 4As shown: Bneck, the main component structure in the MoblieNetv3 network, mainly includes the inverse residual structure of the depthwise separable convolution and the linear bottleneck. The MoblieNetv3 network includes two different structures, Large and Small. The present invention uses a small structure, as shown in the following table S, where SE indicates whether the attention module is used, NL indicates which activation function is used, and s indicates the step size. In order to adapt to the size of the insulator image in this article, the size of the first layer input is changed to 640×640×3. In the Bneck unit of MobileNetv3-Small, a 1×1 convolution is first used to increase the dimension of the input feature map, and then the output feature map is input into a depthwise separable convolution with a convolution kernel size of 3×3. In the depthwise separable convolution, it is assumed that the size of the input feature map is D F ×D F ×M, the convolution kernel size is D K ×F K ×1, the number of convolution kernels is N, and the output feature map size is D F ×D F ×N. The specific steps are: input the feature map, generate 1×1×M feature maps through deep convolution operation, which is used to reduce the model parameters, and then combine the above M feature maps through point-by-point convolution operation with the convolution kernel size as N new feature maps, and then input the output feature map into the SENet module to obtain the weighted features of the SENet module, and finally use 1×1 convolution to reduce the dimension of the output features, and finally obtain the feature map output by MobileNetv3-Small, which can reduce the model parameters and reduce the computational cost. Embedding the ECANet channel attention mechanism in the 12th layer of the MobileNetv3 network can learn the correlation between channels and adaptively adjust the channel weights to improve the performance of the network. ECANet implements channel attention weighting through two steps: one is to extract channel features, and the other is to calculate channel weights. It proposes a local cross-channel interaction strategy without dimensionality reduction and a method for adaptively selecting the size of the one-dimensional convolution kernel, thereby achieving performance improvement. In order to further improve the performance of the model, the original SPPF structure of YOLOv8s is added to this module to achieve local and global feature fusion.

[0024] Add the Slim-Neck module to the feature fusion part, such as Figure 5As shown in the figure: Considering the problems that poor light and angle problems may lead to poor insulator image quality, poor feature extraction effect and slow speed during the inspection process, this paper integrates the Slim-Neck structure into the feature fusion part of YOLOv8s. The Slim-Neck structure is composed of GSConv and VoVGSCSP. Among them, GSConv is composed of three parts: standard convolution, depth-separable convolution and channel mixing, as shown in the figure. Figure 6 As shown in the figure, the main idea of ​​this convolution is to infiltrate the information generated by the standard convolution into the information generated by the depthwise separable convolution. Through the channel shuffling operation, the information interaction between channels can be retained as much as possible at a lower time loss, thereby enhancing the model performance and reducing the number of model parameters. According to the structural definition, it is added before the two upsampling operations of the feature fusion part. Figure 7 As shown in the figure, in the VoV-GSCSP module, GSConv is used to replace the original traditional convolution. By connecting two GSConvs in series, feature maps of different scales can be spliced ​​to improve the feature diversity of the model. The cross-level part is used to increase the depth and nonlinear ability of the network model. VoV-GSCSP replaces the original C2f modules in the 17th, 21st, 24th, and 27th layers.

[0025] 3) Adding WIoUv3 border loss function to the head network: The original YOLOv8 model uses the CIoU border loss function, which can focus on the overlapping area of ​​the border and the center point distance and aspect ratio, making the positioning more accurate. However, once the width and height aspect ratios of the predicted box and the real box are in linear proportion, the CIoU loss function will not work. In order to solve this problem, the present invention introduces a WIoUv3 border loss function with a dynamic non-monotonic mechanism to replace the original CIoU border loss function. The WIoUv3 algorithm uses outliers instead of IoU to evaluate the quality of the anchor frame, and provides a gradient gain allocation strategy, which reduces the competitiveness of high-quality anchor frames and also reduces the harmful gradients generated by low-quality samples. This allows WIoUv3 to focus on anchor frames of general quality. At the same time, because the aspect ratio is not calculated, it is faster. A small outlier means that the anchor frame is of high quality, and a small gradient gain is assigned to it so that the bounding box regression focuses on the anchor frame of ordinary quality. Assigning smaller gradient gains to anchor boxes with larger outliers will effectively prevent low-quality examples from generating large harmful gradients, thereby improving the overall performance of the network model.

[0026] Among them, W g , H g Indicates the width and height of the minimum bounding box, is the normalized distance between the center points of the two bounding boxes. Produces gradients that hinder convergence, Indicates that W g and H g Separated from the computational graph, it effectively eliminates the factors that hinder convergence, and β is the outlier degree.

[0027] The L-YOLOv8s model includes a backbone network, a feature fusion module, and a detection head module. The backbone network includes a CBH module, a MobileNetV3 module, an ECA attention mechanism module, and an SPPF module. The CBH module consists of a standard convolution, normalization, and an H-Swish activation function. MobileNetV3 consists of a 3×3 standard convolution and many Bneck structures. The Bneck structure is used for feature extraction and is an important part of MobileNetv3. It integrates deep separable convolution and residual structures, and introduces a lightweight attention module SENet to improve the ability of feature extraction. This module can reduce the number of model parameters, reduce the computational cost, learn the correlation between channels, and adaptively adjust the weights of channels to improve the performance of the network and achieve local and global feature fusion; the feature fusion module includes a CBS module and a Slim-Neck module. The CBS module is composed of standard convolution, normalization, and SiLU activation functions. The Slim-Neck module includes GSConv and VoV-GSCSP modules, which can retain the information interaction between channels as much as possible with lower time loss, enhance model performance and reduce the number of model parameters, and increase the depth and nonlinearity of the network model. The detection head module adopts a decoupling head and also includes the WIoUv3 border loss function, which can effectively prevent low-quality examples from generating large harmful gradients, thereby improving the overall performance of the network model.

[0028] S4. Put the sorted data set into the L-YOLOv8s model for training. Get the trained model; Put the labeled and divided data set into the L-YOLOv8s model for training to get the best.pt file. After the training is completed, use the validation set for verification.

[0029] S5. Use the trained model to locate and identify the image to be detected.

[0030] The image to be detected is placed in the model using the best.pt weight file for positioning and recognition.

[0031] The original YOLOv8s, YOLOv7, YOLOv5s, YOLOv4, SSD, and Faster-RCNN traditional models were used for comparative experiments with the proposed model, all using the same hyperparameters and training and test sets. The comparison chart is as follows: Figure 8 and 9 As shown in the figure, compared with other traditional models, the L-YOLOv8s proposed in the present invention is better in detection speed and detection accuracy, has lower model complexity, and has stronger detection capabilities for small targets and targets in complex backgrounds. The experiment shows that the model proposed in the present invention has excellent generalization ability and can meet the conditions for real-time detection of aerial photography insulators of transmission lines.

[0032] Although CN116895030A can reduce the complexity of the model, the reduction is small and the volume is still large; CN116883801A only improves the accuracy of the model, but does not change the size of the model parameters. The volume is too large and cannot meet the lightweight requirements of the drone inspection model. Compared with the detection methods proposed by CN116895030A and CN116883801A, the present invention can significantly reduce the complexity of the model while improving the detection accuracy, thereby enabling it to improve the detection speed and accuracy of the transmission line drone inspection.

[0033] The hardware and software configurations used in this experiment are as follows: the operating platform is pycharm 11.0.15, and pytorch1.13.1 is used as the deep learning model building framework. The CPU model is 15vCPU Intel(R)Xeon(R)Platinum 8358PCPU@2.60GHz, the graphics card (GPU) is RTX 3090, the video memory is 24GB, and the cuda version is 12.1.

[0034] In summary, the present invention uses the improved YOLOv8s model to detect and identify defects in transmission line insulators, which improves the recognition accuracy and speed of insulator defects and has practical significance for real-time detection of aerial photography insulators in transmission lines.

[0035] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to the equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A lightweight and improved L-YOLOv8s insulator defect detection method, characterized in that: The following steps are involved: S1. Collect images of transmission line insulators through drones and sort them manually; S2. Labelimg is used to annotate the screened and sorted transmission line insulator images to obtain an insulator defect dataset, which is then divided into a training set, a test set, and a validation set according to a certain ratio; S3, combine MobileNetV3, ECANet attention mechanism, Slim-Neck module, and WioUv3 border loss function into the YOLOv8s model to obtain the L-YOLOv8s model; S4. Put the sorted data set into the L-YOLOv8s model for training to obtain a trained model; S5. Use the trained model to locate and identify the image to be detected.

2. According to claim 1, a lightweight improved L-YOLOv8s insulator defect detection method is characterized in that: Step S2 uses Labelimg to annotate different types of insulator images, where glass insulators are annotated as Glass insulators, composite insulators are annotated as Composite insulators, and porcelain insulators are annotated as Porcelain insulators. The defect types are divided into damage and flashover, where damaged insulators are annotated as Brokeninsulators and flashover insulators are annotated as Flashover damage insulaor. The training set, test set, and validation set are then divided according to a ratio of 8:1:

1.

3. The insulator defect detection method based on lightweight improved L-YOLOv8s according to claim 1 is characterized in that: In step S3, the MobileNetV3 module and the ECANet attention mechanism are added to the YOLOv8s backbone network; the Slim-Neck module is added to the feature fusion part; the WIoUv3 loss function is added to the head network to construct the L-YOLOv8s model.

4. The insulator defect detection method based on lightweight improved L-YOLOv8s according to claim 3 is characterized in that: Use MobileNetV3 to replace the original CSPDarkNet53 module in YOLOv8s, embed the ECANet attention mechanism between the last layer of MobileNetV3 and the original SPPF module of YOLOv8s, and obtain the MobileNetv3-ECA-SPPF backbone network, in which the ECANet attention mechanism is embedded in the 12th layer of the backbone network.

5. The insulator defect detection method based on lightweight improved L-YOLOv8s according to claim 3 is characterized in that: Specifically, GSConv in the Slim-Neck structure is added before the two upsampling operations in the feature fusion part, and VoV-GSCSP replaces the C2f modules originally in the 17th, 21st, 24th, and 27th layers.

6. The method for detecting insulator defects based on lightweight improved L-YOLOv8s according to claim 3 is characterized in that: The WIoUv3 bounding box loss function is used to replace the CIoU loss function in the YOLOv8s detection head module.

7. The insulator defect detection method based on lightweight improved L-YOLOv8s according to claim 6 is characterized in that: The WIoUv3 border loss function uses a dynamic non-monotonic mechanism, and its formula is: Among them, W g , H g Indicates the width and height of the minimum bounding box, is the normalized distance between the center points of two bounding boxes, Indicates that W g and H g Separated from the computational graph, β is the outlier degree.

8. The insulator defect detection method based on lightweight improved L-YOLOv8s according to claim 3 is characterized in that: The L-YOLOv8s model includes a backbone network, a feature fusion module and a detection head module; the backbone network includes a CBH module, a MobileNetV3 module, an ECA attention mechanism module and an SPPF module; the CBH module includes standard convolution, normalization and H-Swish activation function; MobileNetV3 includes standard convolution and multiple Bneck structures, among which the Bneck structure is used for feature extraction, integrating deep separable convolution and residual structure, and introducing a lightweight attention module SENet to improve the feature extraction capability; the feature fusion module includes a CBS module and a Slim-Neck module; the CBS module includes standard convolution, normalization and SiLU activation function; the Slim-Neck module includes GSConv and VoV-GSCSP modules; the detection head module uses a decoupling head and also includes a WIoUv3 border loss function.

9. The insulator defect detection method based on lightweight improved L-YOLOv8s according to claim 1 is characterized in that: Step S4 is to put the labeled and divided data set into the L-YOLOv8s model for training to obtain the best.pt file.

10. The insulator defect detection method based on lightweight improved L-YOLOv8s according to claim 1 is characterized in that: Step S5 is to place the image to be detected into the model using the best.pt weight file for positioning and recognition.

Citation Information

Patent Citations

  • Wind turbine surface defect detection method

    CN115719337A

  • Lightweight target detection method and system based on YOLOv8 improvement

    CN116740538A

  • Three-stage cascade detection method for defects of glass fabric

    CN117974572A

  • Detection method, device and equipment of substation equipment and storage medium

    CN118014979A

  • Wood board small target surface defect identification and detection method based on machine vision

    CN118396968A

Cited By

  • Insulator defect detection method and system based on RT-DETR improvement

    CN119323572A

  • Insulator Defect Detection Method and System Improved Based on RT-DETR

    CN119323572B

  • Multi-source data integration system applied to intelligent recognition and detection of underwater defects of hydropower station dam

    CN119880929A

  • Breaker switch intelligent detection method and system based on deep learning, medium and equipment

    CN120932036A