CN-122023267-A - Lightweight insulator defect detection model, insulator defect detection system, method, equipment and medium

CN122023267ACN 122023267 ACN122023267 ACN 122023267ACN-122023267-A

Abstract

The invention provides a lightweight insulator defect detection model, an insulator defect detection system, an insulator defect detection method, an insulator defect detection device and a lightweight insulator defect detection medium, which comprise the steps of constructing and training a lightweight insulator defect detection model, and carrying out defect identification on an insulator image to be detected based on the trained lightweight insulator defect detection model; the light insulator defect detection model sequentially comprises a light backbone network for extracting initial characteristics, a double-channel attention module for enhancing characteristic perception, an adaptive characteristic quantization module for adaptive compression, and a characteristic fusion and local enhancement network for fusing multi-scale characteristics. By structurally introducing the design, the model remarkably reduces the calculation complexity and the parameter while maintaining higher detection precision, so that the efficient and accurate real-time insulator defect detection can be realized, and the practical application requirement of intelligent power inspection is met.

Inventors

Gong Xiangkui
LU XIN
SUN JING
FAN WENBO
XIE FANG
SUN HAOYANG
Cheng Linjian
JIANG PENGFEI
ZHOU LIMEI
SHANG YUWEI
FANG HENGFU
Bai Shuaitao
YU ZHIPENG
WANG GUANYING
YANG LE
WANG SHUAIPENG

Assignees

中国电力科学研究院有限公司
国家电网有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (20)

1. A lightweight insulator defect detection model, characterized by being constructed based on YOLOv's 7 frame, comprising: The light-weight backbone network is complexed with a depth separable convolution and channel selection mechanism and is used for extracting initial characteristics of the insulator image to obtain an original characteristic diagram; The dual-channel attention module is arranged between the backbone network and the feature pyramid and is used for processing the feature map output by the lightweight backbone network; the self-adaptive feature quantization module is arranged in the middle layer of the network and is used for determining quantization levels according to the spatial importance of different areas in the processed feature map and positioning one or more candidate insulator areas; And the feature fusion and local enhancement network is used for carrying out feature extraction and defect identification based on the insulator candidate region based on improvement of the YOLOv feature pyramid network.
2. The model of claim 1, wherein the dual channel attention module comprises: An attention map calculation sub-module for calculating a global attention map and a local attention map of the feature map in parallel; And the defect feature extraction sub-module is used for fusing the global attention map and the local attention map with the original feature map so as to extract feature information related to defect detection.
3. The model of claim 1, wherein the adaptive feature quantization module comprises: the weighting processing sub-module is used for dividing the processed feature map into areas based on the two-dimensional feature response map and weighting each area according to the importance degree; the quantization level determination submodule is used for respectively adopting different quantization levels for different areas according to the weighted result; The area is divided into a background area and a potential defect area, wherein the background area adopts a first quantization level, and the potential defect area adopts a second quantization level.
4. A model as claimed in claim 3, wherein the weighted processing submodule comprises: the system comprises a two-dimensional characteristic response diagram generating unit, a processing unit and a processing unit, wherein the two-dimensional characteristic response diagram generating unit is used for carrying out statistical compression on the channel dimension on the processed characteristic diagram to generate a two-dimensional characteristic response diagram reflecting the importance of the space position; a weight determining unit for mapping the two-dimensional characteristic response graph into a normalized space weight matrix through an activation function Wherein Represented by two-dimensional characteristic response diagram at coordinates Feature importance of the location; An area dividing unit for dividing the importance threshold value according to the system setting Dividing the region, wherein the weight of a certain position Determining the area as background area when the weight of a certain position And judging the area as a potential defect area.
5. The model of claim 1, wherein the feature fusion and local enhancement network structure is specifically configured to: Based on the potential defect area, the high-resolution local features of the shallow network and the semantic global features of the deep network are fused through additional bypass connection and fine-grained feature fusion.
6. The model of claim 5, wherein the feature fusion uses the following calculation formula Wherein, the In order to obtain the feature map after the fusion, Is a characteristic map of a local area, and ; Is a global feature map, and ; The height of the local feature map represents the spatial resolution of the shallow features; is the width of the local feature diagram, and Determining the spatial dimensions of the local features; the number of channels of the local feature map represents the depth of the shallow features; Is the height of the global feature map, and its size is smaller than ; Is the width of the global feature map, and the size is smaller than ; The number of channels of the global feature map generally contains richer semantic information; the element values in the characteristic diagram are represented as real number sets and belong to real number domains; representing an upsampling operation; representing a channel splicing operation; Representing a fusion convolution operation.
7. An insulator defect detection system, comprising: the acquisition module is used for acquiring an insulator image to be detected; The defect identification module is used for inputting the insulator image into a lightweight insulator defect detection model, adopting a hierarchical reasoning strategy, and utilizing a lightweight candidate region generation network to rapidly analyze the image and locate one or more insulator candidate regions; wherein the lightweight insulator defect detection model is the lightweight insulator defect detection model of claim 1.
8. The system of claim 7, wherein the insulator defect detection system further comprises a model training module to: acquiring an insulator defect data set; constructing a comprehensive loss function by weighted summation of cross entropy classification loss and regression loss based on cross-correlation ratio for model training; And taking the insulator defect data set as a training set, and taking the minimization of the comprehensive loss function as a target, and performing end-to-end training on the lightweight defect detection model.
9. The system of claim 8, wherein the integrated loss function comprises the expression: Wherein: is a comprehensive loss function; Is cross entropy loss; based on the cross-over ratio Is used to determine the regression loss of (1), And Is the weight coefficient of the weight of the object, And is also provided with 。
10. The system of claim 7, wherein the insulator image is an aerial or patrol image of a power transmission line acquired by an unmanned aerial vehicle or a mobile patrol device, and the defect comprises one or more of insulator self-explosion, breakage, contamination, and flashover trace.
11. An insulator defect detection method, comprising: acquiring an insulator image to be detected; inputting the insulator image into a pre-trained lightweight insulator defect detection model, and adopting a hierarchical reasoning strategy to rapidly analyze the image by utilizing a lightweight candidate region generation network and locate one or more insulator candidate regions; wherein, the the lightweight insulator defect detection model is as defined in claim 1.
12. The method of claim 11, wherein the lightweight defect detection model is constructed and trained by: Carrying out initial feature extraction on the insulator image by utilizing a lightweight backbone network fused with a depth separable convolution and a channel selection mechanism to obtain an original feature map; Introducing a dual-channel attention module between a backbone network and a feature pyramid to process a feature map output by the lightweight backbone network; embedding a self-adaptive characteristic quantization module in the middle layer of the network, which is used for determining quantization levels according to the spatial importance of different areas in the processed characteristic diagram and positioning one or more candidate insulator areas; Based on the insulator candidate region, introducing a local enhancement unit to improve a YOLOv characteristic pyramid network, so as to obtain a light insulator defect detection model for characteristic extraction and defect identification; And training the lightweight insulator defect detection model by using an insulator defect data set.
13. The method of claim 12, wherein the initial feature extraction of the insulator image using a lightweight backbone network that incorporates depth separable convolution and channel selection mechanisms, comprises: weighting the feature map by adaptively learning channel weights; and guiding a calculation process of the depth separable convolution by using the weighted result to realize the initial feature extraction of the insulator image and obtain an original feature map.
14. The method of claim 12, wherein the introducing a dual channel attention module between the backbone network and the feature pyramid processes the feature map output by the lightweight backbone network, comprising: Computing a global attention map and a local attention map of the feature map in parallel; The global attention map and the global attention map are fused with the original feature map to extract feature information related to defect detection.
15. The method according to claim 12, wherein the embedding the adaptive feature quantization module in the middle layer of the network is configured to determine quantization levels according to spatial importance of different regions in the processed feature map, including: Dividing the processed feature map into regions based on the two-dimensional feature response map, and weighting each region according to the importance degree; different quantization levels are respectively adopted for different areas according to the weighting result; The area is divided into a background area and a potential defect area, wherein the background area adopts a first quantization level, and the potential defect area adopts a second quantization level.
16. The method of claim 15, wherein the partitioning the region for the processed feature map based on the two-dimensional feature response map comprises: Carrying out statistical compression on the channel dimension of the processed feature map to generate a two-dimensional feature response map reflecting the importance of the space position, wherein the statistical compression on the channel dimension adopts the operation of maximum pooling or average pooling; Mapping the two-dimensional characteristic response map into a normalized spatial weight matrix by an activation function Wherein Represented by two-dimensional characteristic response diagram at coordinates Feature importance of the location; Importance threshold set according to system Dividing the region, wherein the weight of a certain position Determining the area as background area when the weight of a certain position And judging the area as a potential defect area.
17. The method of claim 16, wherein the introducing the local enhancement unit improves on the feature pyramid network of YOLOv, comprising: And based on the potential defect area, fusing the high-resolution local features of the shallow network with the semantic global features of the deep network through additional bypass connection and fine-granularity feature fusion.
18. The method of claim 17, wherein the feature fusion uses the following formula Wherein, the In order to obtain the feature map after the fusion, Is a characteristic map of a local area, and ; Is a global feature map, and ; The height of the local feature map represents the spatial resolution of the shallow features; is the width of the local feature diagram, and Determining the spatial dimensions of the local features; the number of channels of the local feature map represents the depth of the shallow features; Is the height of the global feature map, and its size is smaller than ; Is the width of the global feature map, and the size is smaller than ; The number of channels of the global feature map generally contains richer semantic information; the element values in the characteristic diagram are represented as real number sets and belong to real number domains; representing an upsampling operation; representing a channel splicing operation; Representing a fusion convolution operation.
19. The method of claim 12, wherein training the lightweight insulator defect detection model with an insulator defect dataset comprises: acquiring an insulator defect data set; constructing a comprehensive loss function by weighted summation of cross entropy classification loss and regression loss based on cross-correlation ratio for model training; And taking the insulator defect data set as a training set, and taking the minimization of the comprehensive loss function as a target, and performing end-to-end training on the lightweight defect detection model.
20. The method of claim 19, wherein the integrated loss function comprises the expression: Wherein: is a comprehensive loss function; Is cross entropy loss; based on the cross-over ratio Is used to determine the regression loss of (1), And Is the weight coefficient of the weight of the object, And is also provided with 。

Description

Lightweight insulator defect detection model, insulator defect detection system, method, equipment and medium Technical Field The invention relates to the field of computer vision and target detection, in particular to a lightweight insulator defect detection model, an insulator defect detection system, a method, equipment and a medium. Background The insulator in the high-voltage transmission line is used as key electrical equipment and plays a vital role in an electrical level and a mechanical level. However, due to long-term exposure to outdoor harsh environments (e.g., wind and rain, temperature changes, lightning strikes, dirt dust, etc.), insulators are extremely prone to aging, cracking, breakage, flashover, and the like. Such defects, if not detected in time, may cause large-scale blackouts and cause significant economic losses. Traditional insulator detection relies on manual operation, and is inefficiency and rate of accuracy low. With the development of the deep convolutional neural network, various defects in the insulator can be detected by processing pictures shot by the unmanned aerial vehicle. However, unmanned aerial vehicle image processing often faces problems of complex background, tiny defect positions, shielding and the like, so that the accuracy of an actual detection result is low. In addition, the real-time performance is another challenge because the algorithm needs to be deployed in the edge device. Therefore, the exploration of an insulator defect detection method capable of solving the problems has important research significance and application value. Researchers have proposed a series of insulator defect detection methods based on target detection. Hao et al effectively inhibit the interference of complex background on detection results by optimizing YOLOv backbone network and Feature Pyramid (FPN) structure, han et al introduce a high-efficiency channel attention mechanism in YOLOv5 to improve the recognition capability of overlapping insulator targets, chang et al strengthen the feature extraction effect of shielding and small-size insulators by integrating a global attention mechanism into a YOLOv model, and a Insu-Yolo model developed by Chen et al is based on a YOLOv frame to achieve the detection rate of mAP and 87 FPS of 95.9%. Although the methods obviously improve the precision by enhancing the feature extraction or improving the detection mechanism, the parameter quantity and the calculation complexity are often too large, and the real-time performance of the actual scene is restricted. Although some lightweight models attempt to reduce the computational burden to optimize the real-time performance, it is still difficult to achieve a synergistic improvement of the detection accuracy and the operation efficiency because the accuracy loss of some lightweight models is serious and the requirements of high-accuracy models cannot be satisfied. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a lightweight insulator defect detection model, which is constructed based on YOLOv frames and comprises the following components: The light-weight backbone network is complexed with a depth separable convolution and channel selection mechanism and is used for extracting initial characteristics of the insulator image to obtain an original characteristic diagram; The dual-channel attention module is arranged between the backbone network and the feature pyramid and is used for processing the feature map output by the lightweight backbone network; the self-adaptive feature quantization module is arranged in the middle layer of the network and is used for determining quantization levels according to the spatial importance of different areas in the processed feature map and positioning one or more candidate insulator areas; And the feature fusion and local enhancement network is used for carrying out feature extraction and defect identification based on the insulator candidate region based on improvement of the YOLOv feature pyramid network. Preferably, the dual-channel attention module includes: An attention map calculation sub-module for calculating a global attention map and a local attention map of the feature map in parallel; And the defect feature extraction sub-module is used for fusing the global attention map and the local attention map with the original feature map so as to extract feature information related to defect detection. Preferably, the adaptive feature quantization module includes: the weighting processing sub-module is used for dividing the processed feature map into areas based on the two-dimensional feature response map and weighting each area according to the importance degree; the quantization level determination submodule is used for respectively adopting different quantization levels for different areas according to the weighted result; The area is divided into a background area and a potential defect area, wherein t