Lightweight detection model, method, device and storage medium for small target defects

Abstract

The application discloses a kind of small target defect lightweight detection model, method, equipment and storage medium, detection model is constituted by main network, feature fusion neck and prediction output, main network includes 4 lightweight modules DC-SE and 4 feature extraction modules NewC3-1 connected alternately, responsible for extracting the feature information of insulator;Feature fusion neck includes 4 feature extraction modules NewC3-2, responsible for realizing the information fusion of multidimensional features extracted by main network;Prediction output is responsible for giving the detection result of insulator defect.Lightweight module DC-SE is used to weaken the interference of complex background on insulator fault, complementary extraction insulator subtle feature, and then enhance the extraction ability of target feature information of shallow network.Feature extraction module NewC3-1 and NewC3-2 are responsible for significantly reducing network parameters, and strengthening the ability of network to extract effective information of insulator.The application is helpful to realize the high-precision, high-speed inspection of transmission line, while meeting the requirements of real-time detection of insulator self-explosion defect.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of target detection, and relates to a small-target defect lightweight detection model, method, device and storage medium. BACKGROUND

[0002] A power transmission line is an important component of a power grid system, and it is necessary to study how to more safely and quickly monitor the power transmission line. As a mechanical support and electrical insulation support component in the power transmission line, an insulator plays a crucial role in safe operation of the power transmission line. The insulator is prone to self-explosion, string falling, pollution and other faults due to long-term exposure to the natural environment. Once the insulator has a defect or other faults, the use and service life of the power transmission line will be directly affected.

[0003] The power transmission line has a wide laying range and a special geographical location. Artificial inspection is not only low in efficiency but also affected by geographical environments in different inspection environments, and is prone to visual misjudgment and missed detection. The use of a drone device loaded with a deep learning algorithm for inspection can greatly solve the practical problems of dangerous, low-efficiency and high-labor-cost artificial inspection. Scholars in the field of deep learning have carried out a lot of research work in the direction of insulator detection. The mainstream research direction is divided into two kinds: one is to directly improve the network detection accuracy by optimizing the second-order network or using a larger network model. The other is to train a single-stage lightweight network to realize lightweight real-time detection by continuously optimizing the YOLO series network.

[0004] Zhao et al. improved the Faster R-CNN model by using a pyramid strategy to optimize the feature extraction ability of insulators in complex backgrounds. See W. Zhao, M. Xu, X. Cheng, and Z. Zhao, “An Insulator in Transmission Lines Recognition and Fault Detection Model Based on Improved Faster RCNN,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1-8, 2021, Art no. 5016408. Feng et al. optimized the K-means clustering algorithm for the YOLOv5x large network model to improve the detection accuracy of the network for insulator targets. See Z. Feng, L. Guo, D. Huang, and R. Li, “Electrical Insulator Defects Detection Method Based on YOLOv5,” in 2021 IEEE 10th Data Driven Control and Learning Systems Conference (DDCLS), 2021, pp. 979-984. Although the above improvements have improved the detection accuracy of the network, the large number of model parameters and complex operation process still pose a great challenge to the computing power of mobile devices, and cannot meet the actual application requirements of unmanned aerial vehicle inspection.

[0005] The representative achievements committed to lightweight include that Zhang et al. proposed a dense pyramid structure on the basis of YOLOv3 to improve the utilization rate of deep semantic information of the network and the volume of the lightweight network, details of which can be found in“X. Zhang et al. InsuDet: A Fault Detection Method for Insulators of Overhead Transmission Lines Using Convolutional Neural Networks [R]. IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1-12, 2021, no. 5018512.” Cheng et al. trained the model by using the characteristics of the YOLOv4 single-stage detection network and improved the detection accuracy of the network through different data enhancement methods, details of which can be found in“Y. Cheng: Detection of Power Line Insulator Based on Enhanced YOLO Model [C] / / 2022 IEEE Asia-Pacific Conference on Image Processing, Electronics and Computers (IPEC), 2022, pp. 626-632.” Yan et al. replaced the YOLOv5 backbone network with a shuffleNetV2 network and replaced the PANet downsampling network with a depth separable convolution to reduce the network parameters, details of which can be found in“Yan Hongwen, Wan Junjie, Pan Zhimin, Zhang Jianjun, Ma Rui. Defect Recognition of Power Distribution Components Based on Improved Yolov5-Lite Lightweight [J]. High Voltage Technology: 1-10.” The above several single-stage detection network optimization greatly reduces and compresses the network parameters and model volume, but the detection effect of the insulator self-explosion defect in the unmanned aerial vehicle equipment inspection is poor.

[0006] In order to solve the problem that the target detection algorithm in the insulator defect detection stage is complex in operation and the model volume is large, which leads to the difficulty of deploying the algorithm to the unmanned aerial vehicle embedded device, the feature extraction module and the lightweight module are designed, and the lightweight detection network DE-YOLO for insulator self-explosion defect is proposed, which can compress the model volume while ensuring the detection accuracy. SUMMARY

[0007] In order to solve the above problems, the application provides a small target defect lightweight detection model, which can weaken the interference of a complex background on insulator failure, complementarily extract insulator fine features, significantly reduce network parameters, and strengthen the ability of the network to extract effective information of the insulator, thereby solving the problems of complex operation and large size of the insulator defect detection model.

[0008] The second object of the application is to provide a small target defect lightweight detection method.

[0009] The third object of the application is to provide an electronic device.

[0010] The fourth object of the application is to provide a computer storage medium.

[0011] The technical scheme adopted by the application is a small target defect lightweight detection model, which comprises a backbone network, a feature fusion neck, and a prediction output.

[0012] The backbone network comprises 4 lightweight modules DC-SE and 4 feature extraction modules NewC3-1 connected alternately, and the backbone network is responsible for extracting feature information of the insulator.

[0013] The feature fusion neck comprises 4 feature extraction modules NewC3-2, which are responsible for realizing information fusion of multi-dimensional features extracted by the backbone network.

[0014] The feature extraction modules NewC3-1 and NewC3-2 comprise deep convolution, point-by-point convolution, and channel attention mechanism, which are aimed at reducing model parameters and enhancing the network's ability to extract effective features of the insulator by combining channel information.

[0015] The lightweight module DC-SE fuses traditional convolution and deep convolution, gradually expands the number of channels by using convolution with a number of times growth as the network deepens, and helps complementary extraction of insulator features. Meanwhile, the attention mechanism SE is introduced to strengthen the expression ability of the backbone network to detailed features, improve the feature fusion ability of the network to insulator images of different sizes, and improve the problem of loss of detailed information in the feature fusion process, thereby enhancing the detection performance of the shallow network to the insulator self-explosion defect features.

[0016] The prediction output is responsible for giving the detection result of the insulator defect.

[0017] Further, the feature extraction modules NewC3-j (j = 1, 2) are used to fully extract small target features on the insulator.

[0018] Each NewC3-j (j = 1, 2) module is composed of x New-bottleneckj (j = 1, 2) and 2 convolution modules (Conv+BN+SiLU), and the Conv+BN+SiLU is composed of 1x1 traditional convolution, BN normalization and SiLu activation function, and the channel number is the input channel number of the NewC3-j (j = 1, 2) module;

[0019] The New-bottleneckj (j = 1, 2) module is composed of 3x3 deep convolution, pointwise convolution and ECA, and is divided into New-bottleneck1 and New-bottleneck2 two structures;

[0020] The New-bottleneck1 adopts residual connection between the input and the output, and the residual connection is used to optimize the backbone network and can relieve the problem of gradient explosion;

[0021] The New-bottleneck2 adopts a serial way to transmit information, and because it acts on the fusion structure of the neck, the direct output of the New-bottleneck2 can retain the fusion feature information to a greater extent;

[0022] The backbone network contains 4 NewC3-1, and the module contains 3, 6, 9, 3 New-bottleneck1 in turn; the feature fusion neck contains 4 NewC3-2, and each contains 3 New-bottleneck2;

[0023] The number x of the New-bottleneckj (j = 1, 2) is automatically determined by the product of the configuration file.yaml and the parameter depth_multiple;

[0024] After the New-bottleneckj (j = 1, 2) is processed by 3x3 deep convolution and 1x1 pointwise convolution, the parameter quantity is about 1 / 9 of the traditional convolution, and the depth separable convolution reduces more parameter quantity while ensuring to extract more feature attributes;

[0025] The insulator image entering the New-bottleneckj module is first subjected to 3x3 deep convolution for feature extraction, and the deep separable convolution composed of deep convolution and pointwise convolution is responsible for reducing the parameter quantity of the network. Assuming that the input feature map size is HxWxC x , and the output feature map size is HxWxC y , the ratio of the operation amount after traditional convolution processing to the operation amount after deep separable convolution is:

[0026]

[0027] Among them, P sumP represents the total parameters of the depth separable convolution conv P represents the total parameters of the traditional convolution; HxWx1 is the size of the convolution kernel of the depth convolution, C x C is the number of the convolution kernel, 1x1xC x C is the size of the pointwise convolution kernel, C y HxWxC is the number of the pointwise convolution kernel; HxWxC x C is the size of the traditional convolution kernel, C y C is the number of the traditional convolution kernel.

[0028] As can be seen from formula (4), after processing by 3x3 depth convolution and 1x1 pointwise convolution, the parameter amount is about 1 / 9 of that of the traditional convolution. The depth separable convolution reduces the parameter amount while ensuring the extraction of more feature attributes.

[0029] The pointwise convolution is responsible for splicing the features generated by the depth convolution in new channels, and then transmitting the information into the attention mechanism for processing;

[0030] The attention mechanism shares the same learning parameters through fast one-dimensional convolution of the kernel, so that the model has a lighter volume while effectively capturing the cross-channel interaction, enhancing the network's ability to extract effective information from the insulator, and improving the quality of feature fusion to further improve the network detection accuracy.

[0031] Further, the lightweight module DC-SE realizes feature extraction by connecting 1x1 convolution and d depth convolutions in series, and then splicing them after BN processing and SiLU activation; then it uses the channel attention mechanism SE to solve the problem of too much interference information in the complex background of the pictures taken during power line inspection;

[0032] The DC-SE module is used in the backbone network, and in the four DC-SE modules used in sequence, the number of depth convolutions is 128, 256, 512, and 1024 respectively. They gradually expand the number of channels by using convolution with a number of times growth, and strengthen the extraction of small defect features on the insulator;

[0033] The DC-SE module extracts the channel information globally while learning the dependency between different channels adaptively with the help of SE, adaptively enhances the target feature pixels, and weakens the interference of complex background information on the insulator fault detection. It expands the receptive field of the shallow network while ensuring the accurate extraction of defects by the network.

[0034] A lightweight detection method for small target defects is performed according to the following steps:

[0035] S1, input the insulator image into the backbone network to complete the feature information extraction of the insulator, and send it to the feature fusion neck;

[0036] S2. Feature fusion: After the multi-dimensional feature information extracted by the backbone network is fused, it is sent to the prediction output.

[0037] The calculation process of the DC-SE module in the feature fusion neck is as follows:

[0038] Y0 = Conv(K i )

[0039] Y1 = Concat(Dw d (…Dw2(Dw1(Y0))),Y0)

[0040] K i+1 =SE(Y1)

[0041] Where Y0 represents the condition for input K i (i takes values ​​of 1, 2...n) The feature map extracted by Conv convolution, Y1 represents the result of concatenating Y0 with Dw convolution, and K... i+1 This represents the feature map finally extracted by the SE module;

[0042] S3. Prediction Output: The prediction is completed, and the detection results of defects on the insulator are output.

[0043] An electronic device that uses the above method to detect defects in insulators.

[0044] A computer storage medium storing at least one program instruction, which is loaded and executed by a processor to implement the above-described inductor defect detection method.

[0045] The beneficial effects of this invention are:

[0046] This invention addresses the challenges of computationally complex and large-scale target detection algorithms in the insulator defect detection stage, which hinders deployment on embedded devices like UAVs. It proposes a lightweight target detection network, DE-YOLO, suitable for mobile devices. First, a feature extraction module, NewC3, is proposed, integrating depthwise separable convolution, pointwise convolution, and ECA attention mechanisms. This module significantly reduces network parameters while enhancing the network's ability to extract effective insulator information. Then, a lightweight module, DC-SE, is designed using a strategy of exponentially increasing the number of channels and the channel attention mechanism SE. This module weakens the interference of complex backgrounds on insulator faults and complementaryly extracts subtle insulator features, thereby enhancing the shallow network's ability to extract target feature information. NewC3 and DC-SE ensure the lightweight nature of DE-YOLO while meeting the requirements for real-time detection of insulator spontaneous explosion defects. Attached Figure Description

[0047] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only represent some of the embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative work on the basis of these drawings also belong to the protection scope of the present application.

[0048] Figure 1 is a structural schematic diagram of a lightweight detection model of an embodiment of the present application.

[0049] Figure 2 is a structural schematic diagram of a New-C3(j=1, 2) module in the lightweight detection model of the embodiment of the present application.

[0050] Figure 3 is a structural schematic diagram of a lightweight module DC-SE in the lightweight detection model of the embodiment of the present application.

[0051] Figure 4 is a comparison diagram of the detection method of the embodiment of the present application and other algorithms for detecting insulators under a simple background,

[0052] wherein Figure 4 (a) is a detection effect diagram under a simple background of YOLOv5s, Figure 4 (b) is a detection effect diagram under a simple background of DE-YOLO.

[0053] Figure 5 is a comparison diagram of the detection method of the embodiment of the present application and other algorithms for detecting insulators under a complex background, wherein Figure 5 (a) is a detection effect diagram under a complex background of YOLOv5s, Figure 5 (b) is a detection effect diagram under a complex background of DE-YOLO. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work also belong to the protection scope of the present application.

[0055] Embodiment 1,

[0056] A lightweight detection model for small target defects, the structure of which is shown in Figure 1 , includes a backbone network, a feature fusion neck, and a prediction output.

[0057] The backbone network comprises 4 lightweight modules DC-SE and 4 feature extraction modules NewC3-1 connected alternately, and is responsible for extracting feature information of the insulator;

[0058] The feature fusion neck comprises 4 feature extraction modules NewC3-2, and is responsible for information fusion of multi-dimensional features extracted by the backbone network;

[0059] The feature extraction modules NewC3-1 and NewC3-2 comprise deep convolution, pointwise convolution and channel attention mechanism, aiming to reduce model parameters while enhancing the network's ability to extract effective features of the insulator by combining channel information;

[0060] The lightweight module DC-SE fuses traditional convolution and deep convolution, gradually expands the number of channels by using convolution with a number of times of growth as the network deepens, and helps complementary extraction of insulator features; at the same time, the attention mechanism SE is introduced to strengthen the expression ability of the backbone network to detail features, improve the feature fusion ability of the network to insulator images of different sizes, and improve the loss of detail information in the feature fusion process, thereby strengthening the detection performance of the shallow network to the self-explosion defect features of the insulator;

[0061] The prediction output is responsible for giving the detection result of the insulator defect.

[0062] As shown in Figure 2 ,

[0063] The feature extraction module NewC3-j (j = 1, 2) is to fully extract small target features on the insulator;

[0064] Each NewC3-j (j = 1, 2) module is composed of x New-bottleneckj (j = 1, 2) and 2 convolution modules (Conv+BN+SiLU), and the Conv+BN+SiLU is composed of 1×1 traditional convolution, BN normalization and SiLu activation function, and the channel number is the input channel number of the NewC3-j (j = 1, 2) module; the New-bottleneckj (j = 1, 2) module is composed of 3×3 deep convolution, pointwise convolution and ECA, and is divided into New-bottleneck1 and New-bottleneck2 two structures; the New-bottleneck1 adopts residual connection between the input and the output, and the residual connection is used to optimize the backbone network and can relieve the problem of gradient explosion; the New-bottleneck2 adopts a serial way for information transmission, because it acts on the fusion structure of the neck, and its direct output mode can preserve the fusion feature information to a greater extent;

[0065] The backbone network contains 4 NewC3-1, and the module contains 3, 6, 9, and 3 New-bottleneck1 in turn; the feature fusion neck contains 4 NewC3-2, each of which contains 3 New-bottleneck2;

[0066] The number x of New-bottleneckj(j=1,2) is automatically determined by the product of the configuration file.yaml and the parameter depth_multiple; after the 3×3 deep convolution and the 1×1 point convolution processing of New-bottleneckj(j=1,2), the parameter quantity is about 1 / 9 of the traditional convolution, and the depth separable convolution reduces more parameter quantity while ensuring to extract more feature attributes;

[0067] The point convolution is responsible for splicing the features generated by the deep convolution into new channels, and then transmitting the information into the attention mechanism for processing; the attention mechanism shares the same learning parameters through the fast one-dimensional convolution of the kernel, so that the model has a lighter volume while effectively capturing the cross-channel interaction, strengthening the network's ability to extract effective information of the insulator, and improving the quality of feature fusion to further improve the network detection precision.

[0068] As shown in Figure 3 ,

[0069] The lightweight module DC-SE realizes feature extraction by 1 1×1 convolution and d deep convolution in series, and then splicing with 1 1×1 convolution after BN processing and SiLU activation; then the channel attention mechanism SE is used to solve the problem of too much interference information in the background of the picture taken during the transmission line inspection;

[0070] The DC-SE module is used in the backbone network, and the number of deep convolutions in the 4 DC-SE modules used in turn is 128, 256, 512, and 1024 respectively; they gradually expand the channel number by using convolution with a number of times growth, and strengthen the extraction of small defect features on the insulator;

[0071] The DC-SE module extracts the channel information globally while learning the dependency between different channels adaptively, and adaptively enhances the target feature pixels and weakens the interference of complex background information on the insulator fault detection, which expands the receptive field of the shallow network while ensuring the accurate extraction of defects by the network.

[0072] Embodiment 2,

[0073] A lightweight detection method for small target defects is performed according to the following steps:

[0074] S1, input the insulator image into the backbone network to complete the feature information extraction of the insulator, and send it into the feature fusion neck;

[0075] S2, the multi-dimensional feature information extracted by the feature fusion neck fusion main trunk network is sent to the prediction output;

[0076] The calculation process of the DC-SE module in the feature fusion neck is as follows:

[0077] Y0 = Conv(K i )

[0078] Y1 = Concat(Dw d (...Dw2(Dw1(Y0))), Y0)

[0079] K i+1 = SE(Y1)

[0080] Wherein, Y0 represents the feature map extracted by Conv convolution on input K i (i takes 1, 2...n), Y1 represents the result of splicing Y0 after Dw convolution is performed on Y0, K i+1 represents the feature map finally extracted by the SE module;

[0081] S3, the prediction output completes the prediction, and outputs the detection result of the defect on the insulator.

[0082] In order to verify the detection effect of the DE-YOLO, a comparative experiment is carried out. The insulator image data comes from the China Power Insulator Dataset (CPLID), which contains 848 insulator images, including 600 normal insulator images and 248 defect insulator images. The original data set is small in quantity and unbalanced in sample number, so data enhancement such as random scaling, horizontal flipping, increasing brightness and increasing Gaussian noise is adopted to expand the data set to 4000, and the data set is divided into training set, verification set and test set according to 7:2:1.

[0083] The five comparative classic target detection networks are Faste R-CNN (see “Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks [J]. Advances in neural information processing systems, 2015, 28: 91-99”), SSD (see “

[21] Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector [C] / / European conference on computer vision. Springer, Cham, 2016: 21-37”), YOLO series (see “Redmon J, Farhadi A. Yolov3: An incremental improvement [J]. arXiv preprint arXiv:1804.02767, 2018.” “Bochkovskiy A, Wang C Y, Liao H YM. Yolov4: Optimal speed and accuracy of object detection [J]. arXiv preprint arXiv:2004.10934, 2020.” “J. Glenn, YOLOv5, GitHub, https: / / github.com / ultralytics / YOLOv5, 2020.”), and the comparison test results are shown in Table 1, wherein the bold represents the best index.

[0084] Table 1 Comparison test results of different detection models

[0085]

[0086] As can be seen from Table 1, the detection accuracy of the second-order target detection network Faster R-CNN is only about 1% higher than that of the single-stage network YOLO series, but the model volume and floating point operation amount are dozens of times higher than those of the YOLOv5s network, which is not feasible in the application of unmanned aerial vehicle inspection. The detection accuracy of the SSD network is the lowest, which is 81.9%, and the network model volume is large and the operation is complex. The detection accuracy of the YOLOv3 network can reach 93.6%, the model volume is 62.5M, which is dozens of times higher than that of DE-YOLO, and the floating point operation amount is as high as 156; the detection accuracy of the YOLOv4 network is 92.6%, and the model volume and floating point operation amount are large, which is not conducive to real-time detection of insulator self-explosion defects. The DE-YOLO network reduces the network depth, and the model volume is only 4.2M, which is compressed by 41.6% compared with the model volume of the YOLOv5s network, the floating point operation is reduced by 45%, and the detection accuracy is also improved to 93.2%. As can be seen from the above experimental results, the DE-YOLO network has good robustness and low model complexity, and the performance meets the satisfactory standard, and is more suitable for application in the mobile device of unmanned aerial vehicle inspection of insulators.

[0087] As Figure 4 shown,

[0088] Figure 4 The detection accuracy of the insulator self-explosion defect in (a) is only 72% in a simple environment, indicating that the extraction ability of the YOLOv5s model for small target defects is weak. Figure 4 In (b), the DE-YOLO network locates the position of the insulator with a high accuracy of 93%, and the defect detection accuracy is 75%, which realizes effective extraction of the insulator features.

[0089] As Figure 5 shown,

[0090] Figure 5 In (a), the YOLOv5s has a detection accuracy of only 77% for the self-explosion defect in the complex background, Figure 5 In (b), the DE-YOLO network not only locates the position of the insulator with an accuracy of 94%, but also detects the small target defect of the insulator self-explosion with an accuracy of 87%.

[0091] It can be seen that the DE-YOLO network proposed in the present application still has high detection accuracy even in different complex backgrounds, and can meet the application of mobile devices in terms of model lightweight.

[0092] The small target defect detection method according to the embodiments of the present application can be stored in a computer readable storage medium if realized in the form of a software function module and sold or used as an independent product. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the image reconstruction method according to the embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk and various storage medium capable of storing program codes.

[0093] The above merely describes preferred embodiments of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A light-weight detection method of small target defects, characterized by, The following steps are followed: S1, the insulator image is input into the backbone network, the feature information of the insulator is extracted, and the feature fusion neck is sent; The backbone network comprises 4 lightweight modules DC-SE and 4 feature extraction modules NewC3-1 connected alternately, and the backbone network is responsible for extracting the feature information of the insulator; the feature fusion neck comprises 4 feature extraction modules NewC3-2, which are responsible for realizing the information fusion of the multi-dimensional features extracted by the backbone network; The backbone network contains 4 NewC3-1, which contains 3, 6, 9 and 3 New-bottleneck1 in turn; the feature fusion neck contains 4 NewC3-2, each of which contains 3 New-bottleneck2; The feature extraction module NewC3-j (j = 1, 2) is used to fully extract the small target features on the insulator; Each NewC3-j (j = 1, 2) module is composed of x New-bottleneckj (j = 1, 2) and 2 convolution modules (Conv+BN+SiLU), which are composed of 1×1 traditional convolution, BN normalization and SiLu activation function, and the channel number is the input channel number of the NewC3-j (j = 1, 2) module; The New-bottleneckj (j = 1, 2) module is composed of 3×3 deep convolution, pointwise convolution and ECA, which is divided into New-bottleneck1 and New-bottleneck2 two structures; The residual connection is used between the input and output of New-bottleneck1, which is used to optimize the backbone network to relieve the problem of gradient explosion; New-bottleneck1 fuses the signals after 3×3 deep convolution, pointwise convolution and ECA attention with the low-level signals transmitted forward by residual connection, which can better focus on small target defects; New-bottleneck2 adopts a serial way of information transmission, because it acts on the fusion structure of the neck, its direct output can better preserve the fusion feature information; The number of New-bottleneckj (j = 1, 2) is determined automatically by the product of the configuration file.yaml and the parameter depth_multiple; After the 3×3 deep convolution and 1×1 pointwise convolution of New-bottleneckj (j = 1, 2), the depth separable convolution reduces more parameter quantity while ensuring to extract more feature attributes; The pointwise convolution is responsible for splicing the features generated by the depth convolution into new channels, and then transmitting the information into the attention mechanism for processing; The attention mechanism shares the same learning parameters through fast one-dimensional convolution of the kernel, so that the model has a lighter volume while effectively capturing the cross-channel interaction, strengthening the network's ability to extract effective information of the insulator, improving the quality of feature fusion and further improving the network detection precision S2, after the feature fusion neck fuses the multi-dimensional feature information extracted by the backbone network, it is sent to the prediction output; The lightweight module DC-SE fuses traditional convolution and deep convolution, gradually expands the number of channels by using convolution whose number is doubled with the deepening of the network, and helps complementary extraction of insulator features; meanwhile, the attention mechanism SE is introduced to strengthen the expression ability of the backbone network to the detail features, improve the feature fusion ability of the network to different size insulator images, and improve the loss problem of detail information in the feature fusion process, so as to strengthen the detection performance of the shallow network to the insulator self-explosion defect features; The lightweight module DC-SE realizes feature extraction by 1 1x1 convolution and multiple deep convolutions in series, and then splices 1x1 convolution after BN processing and SiLU activation; then the channel attention mechanism SE is used to solve the problem of too much complex background interference information in the pictures taken during transmission line inspection; The lightweight module DC-SE is used in the backbone network, and the number of deep convolutions in the four lightweight modules DC-SE used in turn is 128, 256, 512, and 1024, which gradually expand the number of channels by using convolution whose number is doubled, and strengthen the extraction of small defect features on the insulator; The lightweight module DC-SE extracts global channel information while learning the dependency between different channels adaptively, adaptively enhances the target feature pixels, and weakens the interference of complex background information on the insulator fault detection, expands the receptive field of the shallow network, and ensures the accurate extraction of defects by the network; The calculation process of the lightweight module DC-SE in the feature fusion neck is: Y0 = Conv(K i ) (1) Y1 = Concat(Dw d (... Dw2(Dw1(Y0))), Y0) (2) K i+1 = SE(Y1) (3) Wherein, Y0 represents the input K i The feature map extracted by the Conv convolution of (i takes 1, 2...n), Y1 represents the result of splicing Y0 after the Dw convolution is implemented on Y0, K i+1 represents the feature map finally extracted by the SE module; S3, the prediction output is completed, and the detection result of the defect on the insulator is output.

2. A computer storage medium, characterized in that, The storage medium has at least one program instruction stored therein, and the at least one program instruction is loaded and executed by the processor to realize the method of claim 1.

Images