CN117455923B

CN117455923B - Insulator defect detection method and system based on YOLO detector

Info

Publication number: CN117455923B
Application number: CN202311799187.1A
Authority: CN
Inventors: 欧周权; 谢振球; 姜雄辉
Original assignee: Tongda Electromagnetic Energy Co ltd
Current assignee: Tongda Electromagnetic Energy Co ltd
Priority date: 2023-12-26
Filing date: 2023-12-26
Publication date: 2024-03-15
Anticipated expiration: 2043-12-26
Also published as: CN117455923A

Abstract

The insulator defect detection method and system based on the YOLO detector provided by the invention are characterized in that an aerial photographing data set is acquired and processed to obtain the processed aerial photographing data set; constructing an initial defect detection model consisting of a trunk feature extraction network, a path aggregation non-attention feature pyramid network and a deformable decoupling detection head; constructing a target loss function training initial defect detection model to obtain an optimized defect detection model; and inputting the processed aerial photographing data set into an optimized defect detection model to obtain a defect detection result. The method can effectively solve the problem of poor defect detection performance caused by factors such as the characteristic of the maximum length-width ratio of the insulator, the small scale of the serial falling defect, the complexity of the background and the like by improving the characteristic extraction network, the characteristic pyramid network, the detection head and the target loss function, improves the defect detection efficiency and the defect detection precision of the insulator, and has the same beneficial effects.

Description

Insulator defect detection method and system based on YOLO detector

Technical Field

The invention relates to the technical field of image detection, in particular to a method and a system for detecting insulator defects based on a YOLO detector.

Background

With the continuous expansion of the scale of the power system in China, the power transmission network of the power system becomes more and more huge and complex. Insulators play an important role in electrical isolation and mechanical support in power transmission systems, and are an integral part of the system. However, since most insulators in the power transmission system are operated outdoors, they are extremely susceptible to the influence of extreme weather from the outside and cause defects. If the treatment and maintenance are not carried out, the electric safety accident is extremely easy to cause, and the stable operation of the electric power transmission system is seriously affected. Therefore, positioning of insulators and defect detection are one of the important links in power transmission systems.

In recent years, automatic identification of insulator defects from aerial images or videos by using Unmanned Aerial Vehicles (UAVs) has become more and more popular, and the detection method has gradually replaced manual detection image identification, so that the detection method becomes a prominent research subject. Specifically, the technologies of the existing insulator image detection technology are largely classified into a conventional method and a deep learning method. The traditional method generally adopts algorithms such as histogram equalization, canny edge extraction, hough transformation and the like. In contrast, a deep learning-based object detector may automatically extract, analyze, and predict defects of image features through a Deep Neural Network (DNN). The method greatly saves human resources, improves the detection speed and precision, and also becomes a new research hot spot in the field of insulator defect detection. The most representative of the object detectors based on deep learning is YOLO series detectors (e.g., YOLOv4, YOLOv5, etc.), which have been widely used as the mainstream method in various surface defect detection fields, including insulator defect detection. However, aerial insulator sub-images typically suffer from the following problems: multiscale (i.e., insulator strings typically belong to large regions, while string drop failure regions belong to small-scale targets), high aspect ratio (i.e., insulator itself takes a long strip shape), complex background, etc. These problems require not only the capability of extracting target features with extremely large aspect ratios, but also the capability of generalizing target detection in a multi-scale and complex background state.

Accordingly, it is an urgent need of those skilled in the art to provide a method and a system for detecting defects of an insulator based on a YOLO detector, which effectively solve the above-mentioned problems.

Disclosure of Invention

The invention aims to provide a method for detecting defects of an insulator based on a YOLO detector, which has clear logic, safety, effectiveness, reliability and simple and convenient operation, and can effectively solve the problem of poor defect detection performance caused by factors such as the characteristic of the extremely large length-width ratio of the insulator, the small scale of the serial-drop defect, the complexity of the background and the like.

Based on the above purpose, the technical scheme provided by the invention is as follows:

a method for detecting insulator defects based on a YOLO detector comprises the following steps:

acquiring a processed aerial photographing data set;

constructing an initial defect detection model consisting of a trunk feature extraction network, a path aggregation non-attention feature pyramid network and a deformable decoupling detection head;

constructing a target loss function to train the initial defect detection model so as to obtain an optimized defect detection model;

and obtaining a defect detection result according to the processed aerial photographing data set and the optimized defect detection model.

Preferably, the acquiring the processed aerial photo data set includes the following steps:

collecting a plurality of insulator images on a power transmission line through aerial photography to form an aerial photography data set;

and sequentially carrying out data enhancement, standardization processing and label making on the aerial photographing data set so as to obtain the processed aerial photographing data set.

Preferably, the backbone feature extraction network comprises: an original backbone network, a strip pooling fusion module and a deformable convolution;

introducing the striping pooling fusion module and the deformable convolution into an original backbone network to construct the backbone feature extraction network, comprising the steps of:

respectively attaching the strip pooling fusion module at the tail ends of a downsampling module and an identity block in the original backbone network;

replacing the convolution in the downsampling module with the deformable convolution.

Preferably, the path aggregation non-attention feature pyramid network comprises: the original path aggregates the feature pyramid network and the non-attention module;

introducing the non-attention module into the original path aggregation feature pyramid network to construct the path aggregation non-attention feature pyramid network, comprising the steps of:

the original path aggregation feature pyramid network is added with transverse connection of a reverse path;

attaching the non-attention module after each convolution in the original path aggregation feature pyramid network, respectively;

the downsampling module in the reverse path consists of convolution, a batch normalization layer, an activation function and downsampling operation.

Preferably, the deformable decoupling detection head comprises: a raw detection head and the deformable convolution;

obtaining a deformable decoupling detection head, comprising the steps of:

replacing the original detection head with a decoupling detection head;

replacing the first two convolution layers in the decoupling detection head with the deformable convolutions;

wherein, the single convolution branch in the original detection head comprises a classification, a regression and a predicted value of the IOU;

the decoupling detection head comprises three convolution branches, and each convolution branch corresponds to a classification, regression and a predicted value of the IOU.

Preferably, the objective loss function includes: classification loss and bounding box regression loss;

the target loss function is obtained specifically as follows:

replacing IoU loss function with CIoU loss function as the bounding box regression loss;

and replacing the cross entropy loss function by using the Focal loss function as the classification loss.

Preferably, an anchor frame generated by a K-means clustering algorithm is further included in both the initial defect detection model and the optimized defect detection model.

A YOLO detector-based insulator defect detection system, comprising:

the acquisition module is used for acquiring the processed aerial photographing data set;

the initial model module is used for constructing an initial defect detection model consisting of a trunk feature extraction network, a path aggregation non-attention feature pyramid network and a deformable decoupling detection head;

the training module is used for acquiring and constructing a target loss function to train the initial defect detection model so as to acquire an optimized defect detection model;

and the prediction module is used for acquiring a defect detection result according to the processed aerial photographing data set and the optimized defect detection model.

The insulator defect detection method based on the YOLO detector provided by the invention is characterized in that an aerial photographing data set is acquired and processed to obtain the processed aerial photographing data set; constructing an initial defect detection model consisting of a trunk feature extraction network, a path aggregation non-attention feature pyramid network and a deformable decoupling detection head; constructing a target loss function training initial defect detection model to obtain an optimized defect detection model; and inputting the processed aerial photographing data set into an optimized defect detection model to obtain a defect detection result.

Compared with the prior art, the method has the advantages that the problems of poor defect detection performance caused by factors such as the characteristic of the extremely large length-width ratio of the insulator, the small scale of the serial-drop defect, the complexity of the background and the like can be effectively solved through the improvement of the characteristic extraction network, the characteristic pyramid network, the detection head and the target loss function, and the defect detection efficiency and the defect detection precision of the insulator are improved.

The invention also provides an insulator defect detection system based on the YOLO detector, and the method has the same technical concept, solves the same technical problems, and has the same beneficial effects, and is not repeated herein.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flowchart of an insulator defect detection method based on a YOLO detector according to an embodiment of the present invention;

fig. 2 is a flowchart of constructing a backbone feature extraction network according to an embodiment of the present invention;

FIG. 3 is a flow chart of a method for constructing a path aggregation non-attention feature pyramid network according to an embodiment of the present invention;

FIG. 4 is a flowchart of acquiring a deformable decoupling detection head according to an embodiment of the present invention;

fig. 5 is a schematic structural diagram of an insulator defect detection system based on a YOLO detector according to an embodiment of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

The embodiment of the invention is written in a progressive manner.

The embodiment of the invention provides an insulator defect detection method and system based on a YOLO detector. The technical problems of poor defect detection performance caused by the factors of the extremely large length-width ratio characteristic of the insulator, the small scale of the serial-drop defect, the complexity of the background and the like in the prior art are mainly solved.

As shown in fig. 1, a YOLO detector-based insulator defect detection method includes the following steps:

s1, acquiring a processed aerial photographing data set;

s2, constructing an initial defect detection model consisting of a trunk feature extraction network, a path aggregation non-attention feature pyramid network and a deformable decoupling detection head;

s3, constructing a target loss function training initial defect detection model to obtain an optimized defect detection model;

s4, obtaining a defect detection result according to the processed aerial photographing data set and the optimized defect detection model.

In the step S1, an aerial photographing data set is collected through aerial photographing equipment and is processed;

in the embodiment, the aerial photographing device mainly comprises an unmanned aerial vehicle flight platform, a Buddhist H20 zoom camera and a computer for image storage and defect detection;

in step S2, the original CSP-dark net53 feature extraction network is improved based on the consideration of more effectively extracting the shape features of the insulator having the extremely large aspect ratio characteristic (elongated shape); based on the fact that in aerial inspection images, the image background is complex, the types of insulators are various, high-level semantic information is needed to realize accurate object detection, and an original path aggregation feature pyramid network PA-FPN is improved; improving the original detection head based on consideration of effectively alleviating conflict between the classification task and the positioning task;

in step S3, the target loss function is improved based on the consideration of the influence of the target frame regression to the detection target, wherein the influence of the background is restrained;

in step S4, the implementation logic specifically includes: inputting the processed aerial image into an improved backbone network (RegNetX-SP-400 MF) to extract depth features of different layers; inputting depth features of different layers into an aggregation path unattended feature pyramid network (PAS-FPN) to fuse multi-scale features, so that the depth features contain cross-layer information on the layers of the features of different scales; the features output by the PAS-FPN are input to a deformable decoupling detection Head (DYOLO-Head) to obtain the position, size, and object class information of the prediction frame.

Preferably, acquiring the processed aerial photo data set comprises the following steps:

In this embodiment, the transmission line insulator aerial photographing data set is manufactured in two parts, namely, image acquisition and data set manufacturing. The acquisition procedure can be described as: and shooting an insulator image on the transmission line by using a zoom camera on the unmanned aerial vehicle cradle head above the transmission line or at the edge side, and transmitting the shot image to a computer for storage. In the process of data set production, the data set is first reinforced by data reinforcing methods such as rotation, translation, overturn, mixup and the like, and then the image is subjected to standardized processing and label production for subsequent neural network learning.

As shown in fig. 2, preferably, the backbone feature extraction network includes: an original backbone network, a strip pooling fusion module and a deformable convolution;

introducing a striping pooling fusion module and a deformable convolution into an original backbone network to construct a backbone feature extraction network, comprising the steps of:

A1. respectively attaching a strip pooling fusion module to the tail ends of a downsampling module and an identity block in an original backbone network;

A2. the convolution in the downsampling module is replaced with a deformable convolution.

It should be noted that the original RegNetX-400MF is composed of 5 stages, the 1 st Stage is a stem network composed of a common convolution layer, and the last 4 stages are stacked by residual modules. Specifically, except for the 1 st Stage (stem network), the 1 st module in the last 4 stages is a downsampling module Down-sampling block, and the other modules are Identity block modules. Wherein the main branches of the two residual blocks consist of 11 convolutions, 33 group convolutions, 11 convolutions and shortcut branches. The shortcut branch of the Identity block (s=1) does not perform task processing, and the shortcut branch of the Down-sampling block (s=2) is additionally added with 11 convolutions of one step size s=2. However, during the task of insulator target detection, insulators that are elongated in appearance tend to cause significant aspect ratio problems. Most target detectors do not take this problem into account during the design process and therefore it is difficult to perform optimally on insulator detection tasks. The network structure of RegNetX400-MF is known that the network is mainly composed of a residual network composed of common 2-D convolution and packet convolution, and has good feature extraction capability and less calculation amount, but still does not provide a solution to the problem of extremely large length-width ratio.

The implementation principle in the step A1 to the step A2 is specifically as follows: the striping pooling fusion module the SPFM module first determines a rectangular pooling kernel along one spatial dimension of the input feature vector to extract longer scale correlations. Meanwhile, the narrower pooling core shape is beneficial to capturing local context information in the residual space dimension, and effectively reduces interference of background information on label prediction. However, the features that have undergone the kernel processing of stripe (narrowband) pooling in the spatial dimension are input to a 1-D convolution layer and feature expansion (expansion) and feature fusion (Element-wise) are performed to obtain the overall situation length scale information correlation. And finally, weighting a weight matrix obtained by the fusion feature through the 11 convolution layers and the Sigmoid function with the original feature map to obtain final output. The extraction capability of the backbone network for shape features is further improved by adding an SPFM module at the end of 4 stages after RegNetX-400MF and replacing the 11 convolutions in each Down-sampling block with Deformable Convolutions (DCNs).

As shown in fig. 3, the path aggregation unobtrusive feature pyramid network preferably includes: the original path aggregates the feature pyramid network and the non-attention module;

introducing a non-attention module into an original path aggregation feature pyramid network to construct the path aggregation non-attention feature pyramid network, comprising the steps of:

B1. the method comprises the steps of (1) attaching the original path aggregation feature pyramid network with transverse connection of a reverse path;

B2. respectively attaching a non-attention module to each convolution in the original path aggregation feature pyramid network;

wherein the downsampling module in the inverse path consists of convolution, batch normalization layer, activation function and downsampling operation.

The steps of performing the detection task by YOLOv4 are as follows: 1) Extracting features from the input image using a CSP-Darknet53 backbone network; 2) Feature maps with the dimensions of 52×52, 26×26 and 13×13 in a backbone network are introduced into a neg structure (a neg part mainly comprises an SPP module and a PA-FPN module); 3) The output of the Neck structure is input into the YOLO-Head module to perform the detection task. Specifically, SPP was introduced into the 13 x 13 feature map to further increase receptive fields and isolate more important contextual features; then, the obtained feature map and the feature maps of 52×52 and 26×26 dimensions are simultaneously introduced into a feature pyramid structure for feature fusion. However, as the depth of the network increases, the location information of the feature map decreases and the semantic information increases. In aerial inspection images, the image background is complex, the types of insulators are various, and accurate object detection can be realized only by high-level semantic information. In contrast, different insulator sizes and smaller insulator bundle drop areas require some low-level positional information.

The implementation principle in the steps B1 to B2 is that the path aggregation non-attention feature pyramid network (PAS-FPN) is added with the transverse connection from the bottom to the top (reverse path) on the original FPN (path aggregation feature pyramid network) of the top to the bottom (forward path), so that the path of information transmitted from the bottom layer features to the high layer features is effectively shortened, convolution operation processing is reduced, and high-layer semantic information and low-layer position information are integrated. Secondly, a SimAM non-attention mechanism is introduced into the feature fusion structure (i.e., on each path of the module) to calculate the full 3-D weights between features without increasing network parameters, and the weight calculation and feature fusion speed can be increased. Wherein the up-sampling module in the "top-down" path consists of a1 x 1 convolution, batch normalization layer BN, siLU activation functions and up-sampling operations, and the down-sampling module in the "bottom-up" path consists of a 3 x 3 convolution, BN, siLU functions and down-sampling operations.

As shown in fig. 4, preferably, the deformable decoupling detection head includes: a raw detection head and a deformable convolution;

obtaining a deformable decoupling detection head, comprising the steps of:

C1. replacing the original detection head with a decoupling detection head;

C2. replacing the first two convolution layers in the decoupling detection head with deformable convolutions;

wherein, the single convolution branch in the original detection head comprises classification, regression and prediction values of the IOU;

It should be noted that, the detection heads of the classical target detector (such as YOLOv 3-v 5) are usually designed to share a unified single-branch structure, or use parallel detection heads for classification and regression. The coupling type detection heads are simple and efficient, but the characteristic diagrams corresponding to the coupling type detection heads have similar receptive fields. However, both classification and regression tasks require different regions of the object of interest: the classification task should focus on regions rich in semantic information, while the localization task should focus on the contours of the object edges.

The implementation principle in the steps C1 to C2 is specifically as follows: the introduction of a Deformable Convolution (DCN) in the dual-branch structure enables the model to adaptively select spatial regions to correspond to the characteristics of a given task by learning offsets. Replacing the original YOLOv4 detection head (only one convolution branch comprises the predicted values of classification, regression and IoU) with a decoupling detection head (three convolution branches respectively comprising classification, regression and IoU) wherein the classification branches are used for mapping the characteristics of different types of defects; and a regression branch for mapping the contour features of the defect boundary. Considering the computational complexity of the task, the first two convolution layers in the decoupling detection head are replaced by DCNs.

Preferably, the objective loss function comprises: classification loss and bounding box regression loss;

the target loss function is obtained, specifically:

replacing IoU loss function by CIoU loss function as boundary box regression loss;

the Focal loss function is used to replace the cross entropy loss function as a classification loss.

The loss function of SMA-YOLO is composed of classification loss, bounding box regression loss, and confidence loss. Wherein a complete cross-union (CIoU) loss function is used as the bounding box regression loss. Specifically, the intersection-union (IoU) is the most widely used loss function, which calculates the ratio of the intersection of the real frame (a) and the predicted frame (B) to the union, as follows:

however, when the bounding boxes do not overlap, the gradient does not slide to optimize the non-overlapping region.

In contrast, CIoU considers the overlap region, normalized center point distance, and aspect ratio, and adds a penalty term to make the target box regression more stable, as follows:

in the method, in the process of the invention,bandb ^gt respectively representing the center points of the prediction frame and the actual frame;p ² (b, b ^gt ) Representing the Euclidean distance between the center points of two frames; c is the diagonal length of the smallest enclosed area that can contain the predicted and real frames;aas the weight coefficient of the light-emitting diode,vis the length-width ratio coefficientThe following formulas are respectively shown:

in the method, in the process of the invention,w、w ^gt the width of the prediction frame and the width of the prediction frame are respectively;h、h ^gt the height of the prediction frame and the prediction frame, respectively.

It is noted that, because the aerial insulator image is affected by weather, shooting distance, shooting angle and other factors, complex background interference is included. These parts, which are not related to the detection target, cause a great classification loss.

Focal loss effectively suppresses the background effect on the detection target, and is more suitable for dealing with this problem than Cross-entropy loss function (Cross-entropy loss).

Thus, the present embodiment uses Focal loss as a classification loss, and its formulation is as follows:

in the method, in the process of the invention,representing the true probability of a category;pa probability representing a predicted category;a=0.2 sumg=0.25 Representing the loss function adjustment factor.

Preferably, the anchor frame generated by the K-means clustering algorithm is also included in both the initial defect detection model and the optimized defect detection model.

In the actual application process, in order to detect the insulator more accurately, the embodiment applies a K-means clustering algorithm to generate Anchor frames, and finally obtains the following 9 Anchor boxes: [3, 20], [7, 5], [13, 80], [15, 16], [71, 18], [101, 33], [250, 50], [263, 80] and [263, 124]. The distance equation adopted by the K-means clustering algorithm is as follows:

in the method, in the process of the invention,representing a distance error; box represents a marked bounding box; centroid is represented as a bounding box of the cluster center.

As shown in fig. 5, an insulator defect detection system based on YOLO detector includes:

the training module is used for acquiring and constructing a target loss function training initial defect detection model so as to acquire an optimized defect detection model;

the prediction module is used for obtaining a defect detection result according to the processed aerial photographing data set and the optimized defect detection model.

In order to further verify the beneficial effects of the insulator defect detection method and system based on the YOLO detector, the invention designs a corresponding verification experiment:

in order to verify the validity of the disclosed defect detection method, an insulator dataset consisting of a power line insulator dataset CPLID and a power inspection intelligent defect detection dataset acquired by an unmanned aerial vehicle is adopted. The CPLID data set comprises 848 composite insulators 425 and Zhang Hang images shot by the unmanned aerial vehicle; the intelligent defect detection data set for electric power inspection comprises 40 aerial images of the glass insulator. In order to improve the universality of the model and solve the problem of single type of insulator defect detection, a data set containing the two types of insulators is integrated, wherein the data set comprises 607 normal insulator aerial images and 281 insulator aerial images with a serial-drop defect. Since the middle insulator has few defect objects, the embodiment adopts the image enhancement technology of changing contrast, randomly adding noise, vertically turning and horizontally turning. Finally, the reinforced aerial insulator data lump contains 3536 aerial images. The dataset was divided into training and testing sets at a ratio of 7:1, i.e. 3094 images were randomly selected as training set and 442 pictures were used for testing. In addition, according to the VOC data format, labeling the images by using a Label-img tool, wherein the labeling is divided into two types of insulators and defects (cross drop);

the experimental super parameters of this embodiment are set by the following training: 1) Setting random gradient descent (SGD) as an optimizer, wherein the weight decay rate and the momentum are 0.0005 and 0.9; 2) The initial learning rate is set to 0.001; 3) Taking cosine annealing attenuation as a weight attenuation strategy; 4) Using Mosaic (mosaics) data enhancement policies; 5) The Epoch and Batch-size are 200 and 16. In addition, a dynamic sample matching policy (SimOTA) was also used as the algorithmic sample matching policy, with a IoU threshold of non-maximum suppression (NMS) set to 0.65.

In order to quantitatively verify the effectiveness of the proposed defect detection method, 5 evaluation indexes are introduced: average Precision (AP), F-Measure, average precision (mAP), frame Per Second (FPS), model parameters (Pa). Specifically, AP is the percentage of correctly identified samples to the total number of samples, typically obtained by calculating the area under the P-R curve, and mAP is the average of all classes of detected APs. Furthermore, the P-R curve is a curve recorded according to two variables: precision (P) and Recall (R). Where Precision refers to the ratio of the number of positive samples predicted correctly to the number of samples predicted positive; recall is the ratio of the number of positive samples correctly predicted to the number of all positive samples; F-Measure is a harmonic mean of precision and recall, and can be comprehensively evaluated. The calculation formulas of these indexes are as follows:

in the method, in the process of the invention,p (r) Represents a P-R curve;mrepresenting the number of detected object categories; true Positives (TP) indicate the number of positive samples correctly identified; false Positives (FP) represent the number of false positive samples identified as negative samples; false Negatives (FN) represent the number of false positive samples that are incorrectly identified as negative samples.

(1) Ablation experimental analysis

In order to verify the effectiveness of the designed RegNetX-SP-400MF network, feature pyramid network PAS-FPN and deformable decoupling detection heads, an ablation experiment was first set up with respect to the disclosed defect detection method, the results of which are shown in the following table:

ablation experimental analysis

Remarks: (a) Representing ablation of RegNetX-SP, i.e., replacement of Backbone with CSP-dark net53; (b) Representing ablation of PAS-FPN, i.e. replacing Neck with PAFPN; (c) Indicating that the ablation is performed on the DYOLO-Head, i.e. the Head is replaced by the original coupled YOLO-Head, (d) indicating that no ablation operation is performed.

From experimental results, it is known that ablation of RegNetX-SP, PAS-FPN and DYOLO-Head all negatively affects the detection model. Among them, ablation of backbone network RegNetX-SP has the greatest impact on detector performance. Wherein the mAP of the model was reduced by about 4% and the AP corresponding to the insulator was reduced by about 6%. This shows that the designed Striping Pooling Fusion Module (SPFM) can improve the detection performance for insulators by effectively extracting insulator features with high aspect ratio characteristics. Furthermore, PAS-FPN ablation has a negative impact that is higher than DYOLO-head, i.e., mAP at the time of ablation is about 1.55% lower than that at the time of ablation. This is because there is much background interference information (e.g., cables, etc.) in aerial insulator images, while the inattentive mechanism in PAS-FPN can effectively suppress background noise. In summary, experiments show that the proposed backbone network, PAS-FPN and DYOLO-Head can effectively improve the detection performance of the detector.

(2) Backbone network comparative experimental analysis

As can be seen from the ablation experiments in the section above, the designed backbone network (RegNetX-SP-400 MF) can improve the detection performance of the model under the condition of obviously reducing the parameter number of the model. Thus, to further verify the superiority of the proposed backbone network, comparative experimental analyses were performed on some classical backbone networks, the results of which are shown in the following table. It should be noted that this comparative experiment only replaced the backbone network portion of the detector, the rest remaining unchanged.

Backbone network contrast experiment

From the experimental results, when the detector is loaded with different backbone networks, the comprehensive performance of the disclosed defect detection method is optimal, and the optimal performance (92.36%, 99.23%, 0.966 and 95.79%) is realized on the insulator AP, the serial AP, the insulator F-measure, the serial F-measure and the mAP respectively. Specifically, compared with classical residual networks such as ResNet, resNest and CSP-DarkNet53, the disclosed defect detection method has advantages in detection accuracy, detection speed and model parameter, namely, the advantages are respectively better than those of the former three in mAP, FPS and Pa by about 3.63% -5.29%, 13-22 and 40.84-64.18. In addition, mobiletv 2, although superior to the disclosed defect detection method (FPS, pa achieved 54 and 9.41, respectively) in terms of detection speed and model parameters, was significantly lower than the disclosed defect detection method in terms of detection accuracy (mAP was about 7.13% lower than the disclosed defect detection method, and AP of the insulator was about 9.8% lower than the disclosed defect detection method). Therefore, the main network RegNetX-SP-400MF can keep good balance between detection accuracy and detection speed.

(3) Multi-scale fusion network comparative experiment analysis

In order to improve the fusion efficiency of a feature pyramid network in a detector to multi-scale features, the invention provides a PAS-FPN feature pyramid network. To verify its effectiveness and superiority, some classical feature pyramid networks were compared with them, and the experimental results are shown in the following table:

characteristic pyramid network contrast experiment

Compared with other characteristic pyramid characteristic fusion networks, the PAS-FPN provided by the invention has the advantages that the optimal performance is realized on various indexes, namely 92.36%, 99.23%, 0.918%, 0.966 and 95.79%, respectively. The PA-FPN and the Bi-FPN have proper performances, and are superior to BFN and FPN in mAP detection accuracy index by about 1.5% -1.94%. This is because the former two have a "top-down" and "bottom-up" feature fusion process, and the latter two lack cross-layer feature extraction due to the lack of a "top-down" feature fusion process. In addition, note that PAS-FPN's mAP is about 0.5% -1% better than PA-FPN and Bi-FPN, indicating that optimizing the feature pyramid feature fusion network through the SimAM attention mechanism is effective. In conclusion, the method provided by the invention can effectively improve the detection performance of the model and has certain advantages compared with a comparison method.

(4) Contrast experiment of different detection models

To verify the superiority of the proposed detection algorithm, some classical two-part detectors (Faster-RCNN, libra-RCNN), anchor-free detector (FCOS, cornerNet), single-step detector (Yolov 4, yolox), etc. were compared with the disclosed defect detection methods, and the experimental results are shown in the following table:

comparative experiments with different models

Remarks: the backbone networks for the Faster-RCNN, libra-RCNN, FCOS and CornerNet detectors are ResNet50.

As can be seen from the experimental results of the table, the mentioned SMA-YOLO achieved the best performance on mAP, FPS and Pa indexes, 95.79%, 45 and 16.81 respectively. Furthermore, from the results, it is known that the single-step detectors YOLOX-m and YOLOv4 have sub-optimal comprehensive detection performance. Wherein YOLOX-m achieves an optimum in AP and F-measure indices of the insulator of 92.67% and 0.921%, respectively, but still below the disclosed defect detection methods by about 0.1%, 5 and 8.47 in mAP, FPS and Pa indices. Second, mAP of Faster-RCNN and Libra-RCNN, which are two-step detectors, was found to be about 0.2% -2.53% better than that of Anchor-free detector (FCOS, cornerNet). This may be because the Anchor-free detector lacks a preset Anchor box, and returns to the bounding box only through midpoints or corner points, so that the model positioning performance is limited. In summary, compared with a comparative model, the disclosed defect detection method has the advantages of smaller model parameters and faster FPS, and also has the advantages of detection accuracy index (mAP), which verifies the advantages of the disclosed defect detection method.

In the embodiments provided in the present application, it should be understood that the disclosed method and system may be implemented in other manners. The system embodiment described above is merely illustrative, for example, the division of modules is merely a logical function division, and there may be other division manners in actual implementation, such as: multiple modules or components may be combined, or may be integrated into another system, or some features may be omitted, or not performed. In addition, the various components shown or discussed may be coupled or directly coupled or communicatively coupled to each other via some interface, whether indirectly coupled or communicatively coupled to devices or modules, whether electrically, mechanically, or otherwise.

In addition, each functional module in each embodiment of the present invention may be integrated in one processor, or each module may be separately used as one device, or two or more modules may be integrated in one device; the functional modules in the embodiments of the present invention may be implemented in hardware, or may be implemented in a form of hardware plus a software functional unit.

Those of ordinary skill in the art will appreciate that: all or part of the steps for implementing the above method embodiments may be implemented by program instructions and associated hardware, where the program instructions may be stored in a computer readable storage medium, and where the program instructions, when executed, perform steps comprising the above method embodiments; and the aforementioned storage medium includes: a mobile storage device, a Read Only Memory (ROM), a magnetic disk or an optical disk, or the like, which can store program codes.

It should be appreciated that the terms "system," "apparatus," "unit," and/or "module," if used herein, are merely one method for distinguishing between different components, elements, parts, portions, or assemblies at different levels. However, if other words can achieve the same purpose, the word can be replaced by other expressions.

As used in this application and in the claims, the terms "a," "an," "the," and/or "the" are not specific to the singular, but may include the plural, unless the context clearly dictates otherwise. In general, the terms "comprises" and "comprising" merely indicate that the steps and elements are explicitly identified, and they do not constitute an exclusive list, as other steps or elements may be included in a method or apparatus. The inclusion of an element defined by the phrase "comprising one … …" does not exclude the presence of additional identical elements in a process, method, article, or apparatus that comprises an element.

If a flowchart is used in the present application, the flowchart is used to describe the operations performed by the system according to embodiments of the present application. It should be appreciated that the preceding or following operations are not necessarily performed in order precisely. Rather, the steps may be processed in reverse order or simultaneously. Also, other operations may be added to or removed from these processes.

The insulator defect detection method and system based on the YOLO detector provided by the invention are described in detail. The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. The insulator defect detection method based on the YOLO detector is characterized by comprising the following steps of:

acquiring a processed aerial photographing data set;

obtaining a defect detection result according to the processed aerial photographing data set and the optimized defect detection model;

the feature extraction network includes: an original backbone network, a strip pooling fusion module and a deformable convolution;

the path aggregation non-attention feature pyramid network comprises: the original path aggregates the feature pyramid network and the non-attention module;

the deformable decoupling detection head includes: a raw detection head and the deformable convolution;

replacing the convolutions in the downsampling module with the deformable convolutions;

attaching a transverse connection of a reverse path to the original path aggregation feature pyramid network;

the downsampling module in the reverse path consists of convolution, a batch normalization layer, an activation function and downsampling operation;

the method for obtaining the deformable decoupling detection head comprises the following steps:

replacing the original detection head with a decoupling detection head;

2. The YOLO detector-based insulator defect detection method of claim 1, wherein the acquiring the processed aerial dataset comprises the steps of:

3. The YOLO detector-based insulator defect detection method of claim 1, wherein the target loss function comprises: classification loss and bounding box regression loss;

the target loss function is obtained specifically as follows:

4. The YOLO detector-based insulator defect detection method of claim 1, further comprising anchor boxes generated by a K-means clustering algorithm in both the initial defect detection model and the optimized defect detection model.

5. A YOLO detector-based insulator defect detection system, comprising:

the prediction module is used for obtaining a defect detection result according to the processed aerial photographing data set and the optimized defect detection model;

replacing the original detection head with a decoupling detection head;