CN114299354A - Insulator detection positioning method based on rotating frame identification network - Google Patents
Insulator detection positioning method based on rotating frame identification network Download PDFInfo
- Publication number
- CN114299354A CN114299354A CN202111431311.XA CN202111431311A CN114299354A CN 114299354 A CN114299354 A CN 114299354A CN 202111431311 A CN202111431311 A CN 202111431311A CN 114299354 A CN114299354 A CN 114299354A
- Authority
- CN
- China
- Prior art keywords
- module
- insulator
- fpn
- network
- pan
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to an insulator detection positioning method based on a rotating frame identification network, which comprises the following steps: s1: constructing a data set of an insulator image of the power transmission line, marking an insulator, and expanding a database; s2: an insulator detection model is constructed, the insulator detection model comprises a trunk network, a feature processing component and a prediction module which are connected in sequence, and an input image of the insulator detection module is sent into the prediction module after feature extraction is carried out on the input image through the trunk network and the feature processing component in sequence; s3: training a multi-angle candidate frame generation network based on an insulator image data set; s4: and sending the insulator image to be detected into a multi-angle candidate frame generation network to obtain a prediction frame of the insulator in the insulator image to be detected. Compared with the prior art, the invention has the advantages of high detection and extraction precision and the like.
Description
Technical Field
The invention relates to the field of insulator detection, in particular to an insulator detection positioning method based on a rotating frame identification network.
Background
The continuous expansion of the scale of the power grid in the modern society provides great challenges for the maintenance and inspection of the power industry, and intelligent inspection becomes a new trend of the inspection development of the power transmission line. The insulator on the power transmission line has the characteristics of light weight, pollution flashover resistance and the like, can play the roles of electrical insulation and mechanical fixation on the line, and is widely applied to power grids in China. Therefore, the detection of the insulator also becomes an important link of the power transmission line inspection. The power transmission line inspection method is generally divided into manual inspection and unmanned aerial vehicle inspection, inspection workload is greatly increased along with the expansion of the scale of a power grid, the defects of low efficiency, danger and the like are exposed in the traditional manual inspection method, and powerful help is provided for the inspection of the power transmission line through the development of the unmanned aerial vehicle aerial photography technology. Unmanned aerial vehicle patrols and examines and has advantages such as with low costs, efficient, flexibility are strong, can follow different positions, angle and apart from collecting the insulation subdata information on the transmission line.
In recent years, with continuous breakthrough of a deep learning theory, the convolutional neural network greatly improves the accuracy of image recognition, deep learning technology gradually starts to permeate into various fields, and an insulator detection method based on deep learning is rapidly developed. At present, the following problems mainly exist in the detection of the insulator of the power transmission line: 1) background information of the power transmission line is very complex, and factors such as vegetation in the background and staggered towers increase a lot of interference for insulator detection, so that the difficulty of insulator detection is greatly increased. 2) The insulator has a large length and a large width, and an insulator prediction frame identified by a common target detection network contains a lot of useless information except the insulator, so that the subsequent analysis of the state and the defect of the insulator is not facilitated.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide an insulator detection and positioning method based on a rotating frame identification network.
The purpose of the invention can be realized by the following technical scheme:
an insulator detection positioning method based on a rotating frame identification network comprises the following steps:
s1: constructing a data set of an insulator image of the power transmission line, marking an insulator, and expanding a database;
s2: an insulator detection model is constructed, the insulator detection model comprises a trunk network, a feature processing component and a prediction module which are connected in sequence, and an input image of the insulator detection module is sent into the prediction module after feature extraction is carried out on the input image through the trunk network and the feature processing component in sequence;
s3: training a multi-angle candidate frame generation network based on an insulator image data set;
s4: and sending the insulator image to be detected into a multi-angle candidate frame generation network to obtain a prediction frame of the insulator in the insulator image to be detected.
Preferably, the backbone network includes a first extraction module, a second extraction module, a third extraction module, and a fourth extraction module, which are connected in sequence.
Preferably, the feature processing component comprises an FPN module and a PAN module which are connected in sequence.
Preferably, the FPN module includes first FPN module, second FPN module and third FPN module, the output characteristic diagram that the fourth drawed the module is sent into first FPN module and is handled, the output characteristic diagram that first FPN module, third drawed the module is sent into second FPN module and is handled, the output characteristic diagram that the second drawed module, second FPN module is sent into the third FPN module and is handled.
Preferably, the PAN modules include a first PAN module, a second PAN module, and a third PAN module, the output characteristic diagram of the third FPN module is sent to the third PAN module for processing, the output characteristic diagram of the second FPN module and the third PAN module is sent to the second PAN module for processing, the output characteristic diagram of the first FPN module and the second PAN module is sent to the first PAN module for processing, and the output characteristic diagram of the first PAN module, the second PAN module, and the third PAN module is sent to the prediction module for processing.
Preferably, the prediction module performs multi-angle insulator detection by using a CSL annular smooth label.
Preferably, the expression of the CSL is:
where g (x) is a window function, x is an input value, r is a radius of the window function, and θ represents an angle of the current bounding box.
Preferably, the loss function of the prediction module is:
where N denotes the number of preselected boxes, objnIs a binary value, obj n1 denotes foreground, obj n0 denotes background, no regression denotes background, v'njIndicates the predicted offset value, vnjRepresenting the true value of the object, thetan、θ'nRespectively representing the true and predicted values of the angle, tnLabels representing objects, pnIs the probability distribution, λ, calculated by the Sigmoid function1、λ2、λ3Represents the compromise values in different losses, x, y, w, h, thetaregRespectively, the coordinate, width, height and angle function of the midpoint of the prediction frame, LregTo return loss, LCSLFor loss of focus, LcslIs a classification loss.
Preferably, the prediction module processes the anchor point frame by using a K-means clustering algorithm, and the formula of the K-means clustering algorithm is as follows:
where k denotes the number of clusters to be divided, NiDenotes the ith cluster set, xjRepresents the jth sample value, ciDenotes the center of the ith cluster, and E denotes the sum of the squares of the euclidean distances of all data sample objects to the cluster center to which it belongs.
Preferably, the method adopted by the database expansion in step S1 includes flipping, rotating, adding noise, and changing color gamut.
Compared with the prior art, the invention has the following advantages:
1) according to the invention, the original network is improved by increasing the angle parameters, the rotating frame which is more fit with the appearance of the insulator is constructed according to the appearance characteristics of the insulator, the extraction capability of the model on the insulator characteristics is greatly improved, the characteristics of the insulator target can be effectively extracted under the complex background, and the detection of the model on the insulator shot at a far position is also advantageous due to the input scale of the high-resolution picture.
2) The invention effectively solves the problem of sudden increase of angle boundary difference caused by the continuity of angle parameters by adding a CSL (circular smooth label) model in the network and outputting the angle loss in a classification loss mode. Network training effect and the rotating frame detection precision for the insulator are greatly improved.
3) The invention uses a data enhancement means to expand the data set, prevents under-fitting condition during model training, and solves the problem of insufficient data in target detection.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic view of the model structure of the present invention;
FIG. 3 is an illustration of a four parameter defined prediction;
FIG. 4 is an illustration of a five parameter defined prediction;
FIG. 5 is an original prediction box;
FIG. 6 is a prediction block of the prediction module output after the introduction of CSL;
FIG. 7 is an analysis diagram of the angular overshoot problem;
FIG. 8 is a diagram of a circular smooth label structure;
FIG. 9 is a block diagram of a feature processing component.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. Note that the following description of the embodiments is merely a substantial example, and the present invention is not intended to be limited to the application or the use thereof, and is not limited to the following embodiments.
Examples
An insulator detection positioning method based on a rotating frame identification network comprises the following steps:
s1: and constructing a data set of the insulator image of the power transmission line, marking the insulator, and expanding the database. The method comprises the steps that an unmanned aerial vehicle shoots power transmission line insulator images under different backgrounds, images with insulators need to be collected in a data set, images without insulator targets need to be collected as training negative samples, and the data set is divided into a training set and a testing set. And (3) labeling the data set by using RoLabelImg software, and storing the coordinates, width and height of the central point of the target in the image and the rotation angle in a YOLO format.
In this embodiment, the data set is augmented by a data enhancement means, which mainly includes means such as flipping, rotating, adding noise, and changing color gamut, and the augmented data set is augmented by a method of 7: and 3, dividing the ratio into a training set and a test set, and inputting the training set and the test set into the model for training.
S2: and constructing an insulator detection model, wherein the insulator detection model comprises a trunk network, a feature processing component and a prediction module which are sequentially connected, and an input image of the insulator detection module is sent into the prediction module after feature extraction is sequentially carried out on the input image by the trunk network and the feature processing component. In this embodiment, the feature processing component includes an FPN module and a PAN module, which are connected in sequence.
Specifically, as shown in fig. 2, the backbone network includes a first extraction module, a second extraction module, a third extraction module, and a fourth extraction module, which are connected in sequence;
the FPN module comprises a first FPN module, a second FPN module and a third FPN module, the output characteristic diagram of the fourth extraction module is sent to the first FPN module for processing, the output characteristic diagrams of the first FPN module and the third extraction module are sent to the second FPN module for processing, and the output characteristic diagrams of the second extraction module and the second FPN module are sent to the third FPN module for processing;
the PAN module comprises a first PAN module, a second PAN module and a third PAN module, the output characteristic diagram of the third FPN module is sent to the third PAN module for processing, the output characteristic diagram of the second FPN module and the third PAN module is sent to the second PAN module for processing, the output characteristic diagram of the first FPN module and the output characteristic diagram of the second PAN module are sent to the first PAN module for processing, and the output characteristic diagram of the first PAN module, the output characteristic diagram of the second PAN module and the output characteristic diagram of the third PAN module are sent to the prediction module for processing.
When the method is applied, the parameter definition of the target detection frame is improved, a long edge definition method is added, as shown in the five-parameter definition in figure 3, and four parameters of the original network prediction rectangular frame are defined: xmin, ymin, xmax, ymax, modifying the parameter definition form, as shown in fig. 4, where the modified prediction rectangle frame parameters are: the centerx, centery, width and height are composed of the coordinate of the center point of the rectangular frame, the length and width of the rectangular frame and the included angle between the long side of the rectangular frame and the x axis. Compared with the original parameter definition method, the angle parameter is added on the basis of the horizontal definition of the prediction frame, the extraction capability of the network on the specific target characteristics is greatly improved, and the network group can obtain better effect in the training and prediction stages.
Fig. 5 shows the type of the prediction frame before improvement, and the type of the prediction frame after improvement is as shown in fig. 6, on the basis of the original horizontal frame, a rotation angle option is provided, the labeling mode of the rotation frame is more suitable for the lifting point of the appearance of the insulator string, and during labeling, the edge of the labeling frame can be more attached to the edge of the insulator. The accurate marking mode can provide more accurate characteristic information for the network, and is beneficial to restraining the training direction of the network and reducing the convergence time of the network.
In the embodiment, the insulator detection model is improved, the angle parameters are introduced, and in the angle regression process of network training, because of the angle continuity, the problem of sudden increase of the difference value of the angle at the boundary can occur, so that the loss value is increased suddenly, and the learning difficulty is increased. The specific problem analysis is shown in fig. 7. To solve this problem, the prediction module introduces a Circular Smooth Label (CSL), as shown in fig. 8, to classify continuous angles and convert the regression problem into a classification problem, that is, convert the continuous angle problem into a discrete angle problem, so as to avoid abrupt change of loss value caused by abrupt change of angle. The serialization is converted into the discretization, certain precision loss is bound to exist, the feasibility of the method is explored, and the influence of the precision loss caused by the discretization on the final detection result needs to be calculated. The calculation formula is as follows.
Maximum loss of precision:
average loss of precision:
through calculation, when the angle ω is 2, the maximum loss of precision is 1, the average loss is 0.5, the length-width ratio of the insulator string is large, and the length-width ratio of the marking frame can reach 1 on average during marking: 9. if the angle difference between the two insulator marking frames is 1 and 0.5, the intersection ratio of the prediction frame and the real frame is reduced by about 0.01 and 0.05 during calculation, and the influence on the insulator detection effect is small. Therefore, the method can effectively solve the problem of angle sudden increase caused by continuous angle boundaries, and simultaneously greatly improves the robustness of network training.
The expression of CSL (circular smooth label) is as follows:
where g (x) is a window function, x is an input value, r is a radius of the window function, and θ represents an angle of the current bounding box.
3) The insulator string has a large length and a large width, and the size of an initial anchor point frame is not suitable for detecting the insulator. The anchor box K-means clustering algorithm in the YOLOv3 network is utilized, and the formula is as follows:
where k denotes the number of clusters to be divided, NiDenotes the ith cluster set, xjRepresents the jth sample value, ciDenotes the center of the ith cluster, and E denotes the sum of the squares of the euclidean distances of all data sample objects to the cluster center to which it belongs. And when the clustering center point is not changed or the value of E is converged, outputting a clustering result.
Because the network adds new angle information, the algorithm needs to be modified according to the labeling parameters of the rotating frame to obtain new anchor point frame parameters: [(184,21),(190,27),(251,30)],[(31,285),(300,38),(415,63)]
[ (611,77), (655,90), (758,99) ], new parameters are input into the network instead of the original parameters, improving the training effect of the network on the insulator.
The overall model structure formed in this embodiment is as shown in fig. 2, the input size of the network is 1024 × 1024, the original image is processed through the Focus structure after being input into the network, the picture is sliced, the 1024 × 3 image is changed into a 512 × 12 feature map, and the feature map is finally changed into a 512 × 32 feature map through a convolution operation. Compared with the ordinary downsampling operation, the calculation amount of the Focus is larger, but the downsampling operation does not cause information loss. Then, the insulator characteristics are extracted through the backbone network, and the extracted characteristics are input into the FPN + PAN structure, where the structure diagram is shown in fig. 9. FPN layer conveys strong semantic feature from the top downwards, and the PAN structure conveys strong location feature from the bottom upwards, and two structures combine together, play different backbone layers and carry out the effect of polymerization to different detection layer characteristics. And finally obtaining three feature maps with different sizes for prediction after the feature maps are fused through the features of the FPN + PAN structure.
S3: and training the multi-angle candidate frame generation network based on the insulator image data set.
Comparing the result of the network prediction with the real value of the labeling frame, and performing regression training of each output parameter, wherein the regression loss types comprise: confidence loss, object classification loss, and localization loss. The addition of the angle information has small correlation between classification loss and confidence loss, so that only the positioning loss needs to be modified, and the modified positioning loss function is as follows:
where N denotes the number of preselected boxes, objnIs a binary value, obj n1 denotes foreground, obj n0 denotes background, no regression denotes background, v'njIndicates the predicted offset value, vnjRepresenting the true value of the object, thetan、θ'nRespectively representing the true and predicted values of the angle, tnLabels representing objects, pnIs the probability distribution, λ, calculated by the Sigmoid function1、λ2、λ3Represents the compromise values in different losses, x, y, w, h, thetaregRespectively, the coordinate, width, height and angle function of the midpoint of the prediction frame, LregTo return loss, LCSLFor loss of focus, LcslIs a classification loss. And calculating the sum of the losses, and outputting the weight after model training when the total loss value is converged. And detecting the target by using the weight. And finally, screening out the frames with the repetition degree of more than 0.5 with the real frames of the targets through NMS (non-maximum suppression) to obtain the position and the type information of the detected targets.
In this embodiment, using precision (precision), Recall (Recall), and average accuracy (mAP) to evaluate model performance, the calculation of precision and Recall may be expressed as:
wherein TP represents the number of prediction blocks that are actually positive samples and are predicted as positive samples; FP represents the number of prediction boxes that are actually negative samples but are predicted to be positive samples; FN represents the number of prediction blocks that are actually positive samples but are predicted as negative samples.
The integral of a curve of the precision p changing along with the recall rate r from 0 to 1 is regarded as the Accuracy (AP), the average value of the APs of all the classes is the average accuracy (mAP), and the calculation is expressed as:
and finishing training based on the content to obtain a trained detection model.
S4: and sending the insulator image to be detected into a multi-angle candidate frame generation network to obtain a prediction frame of the insulator in the insulator image to be detected.
The above embodiments are merely examples and do not limit the scope of the present invention. These embodiments may be implemented in other various manners, and various omissions, substitutions, and changes may be made without departing from the technical spirit of the present invention.
Claims (10)
1. An insulator detection positioning method based on a rotating frame identification network is characterized by comprising the following steps:
s1: constructing a data set of an insulator image of the power transmission line, marking an insulator, and expanding a database;
s2: an insulator detection model is constructed, the insulator detection model comprises a trunk network, a feature processing component and a prediction module which are connected in sequence, and an input image of the insulator detection module is sent into the prediction module after feature extraction is carried out on the input image through the trunk network and the feature processing component in sequence;
s3: training a multi-angle candidate frame generation network based on an insulator image data set;
s4: and sending the insulator image to be detected into a multi-angle candidate frame generation network to obtain a prediction frame of the insulator in the insulator image to be detected.
2. The insulator detection and positioning method based on the rotating frame identification network according to claim 1, wherein the backbone network comprises a first extraction module, a second extraction module, a third extraction module and a fourth extraction module which are sequentially connected.
3. The insulator detection and positioning method based on the rotating frame recognition network as claimed in claim 2, wherein the feature processing component comprises an FPN module and a PAN module which are connected in sequence.
4. The insulator detection and positioning method based on the rotating frame recognition network as claimed in claim 3, wherein the FPN modules comprise a first FPN module, a second FPN module and a third FPN module, the output feature map of the fourth extraction module is sent to the first FPN module for processing, the output feature maps of the first FPN module and the third extraction module are sent to the second FPN module for processing, and the output feature maps of the second extraction module and the second FPN module are sent to the third FPN module for processing.
5. The method as claimed in claim 4, wherein the PAN modules include a first PAN module, a second PAN module, and a third PAN module, the output feature map of the third FPN module is sent to the third PAN module for processing, the output feature maps of the second FPN module and the third PAN module are sent to the second PAN module for processing, the output feature maps of the first FPN module and the second PAN module are sent to the first PAN module for processing, and the output feature maps of the first PAN module, the second PAN module, and the third PAN module are sent to the prediction module for processing.
6. The insulator detection and positioning method based on the rotating frame recognition network of claim 1, wherein the prediction module uses a CSL annular smooth label for multi-angle insulator detection.
7. The insulator detection and positioning method based on the rotating frame recognition network according to claim 6, wherein the expression of the CSL is as follows:
where g (x) is a window function, x is an input value, r is a radius of the window function, and θ represents an angle of the current bounding box.
8. The insulator detection and positioning method based on the rotating frame recognition network as claimed in claim 6, wherein the loss function of the prediction module is:
where N denotes the number of preselected boxes, objnIs a binary value, objn1 denotes foreground, objn0 denotes background, no regression denotes background, v'njIndicates the predicted offset value, vnjRepresenting the true value of the object, thetan、θ'nRespectively representing the true and predicted values of the angle, tnLabels representing objects, pnIs the probability distribution, λ, calculated by the Sigmoid function1、λ2、λ3Represents the compromise values in different losses, x, y, w, h, thetaregAre respectively predictionCoordinate of the frame midpoint, width, height and angle function, LregTo return loss, LCSLFor loss of focus, LcslIs a classification loss.
9. The insulator detection and positioning method based on the rotating frame recognition network as claimed in claim 8, wherein the prediction module processes the anchor frame by using a K-means clustering algorithm, and the formula of the K-means clustering algorithm is as follows:
where k denotes the number of clusters to be divided, NiDenotes the ith cluster set, xjRepresents the jth sample value, ciDenotes the center of the ith cluster, and E denotes the sum of the squares of the euclidean distances of all data sample objects to the cluster center to which it belongs.
10. The insulator detection and positioning method based on the rotating frame recognition network as claimed in claim 1, wherein the database expansion in step S1 adopts methods including flipping, rotating, adding noise, and color gamut changing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111431311.XA CN114299354A (en) | 2021-11-29 | 2021-11-29 | Insulator detection positioning method based on rotating frame identification network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111431311.XA CN114299354A (en) | 2021-11-29 | 2021-11-29 | Insulator detection positioning method based on rotating frame identification network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114299354A true CN114299354A (en) | 2022-04-08 |
Family
ID=80965500
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111431311.XA Pending CN114299354A (en) | 2021-11-29 | 2021-11-29 | Insulator detection positioning method based on rotating frame identification network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114299354A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117274748A (en) * | 2023-11-16 | 2023-12-22 | 国网四川省电力公司电力科学研究院 | Lifelong learning power model training and detecting method based on outlier rejection |
-
2021
- 2021-11-29 CN CN202111431311.XA patent/CN114299354A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117274748A (en) * | 2023-11-16 | 2023-12-22 | 国网四川省电力公司电力科学研究院 | Lifelong learning power model training and detecting method based on outlier rejection |
| CN117274748B (en) * | 2023-11-16 | 2024-02-06 | 国网四川省电力公司电力科学研究院 | Lifelong learning power model training and detecting method based on outlier rejection |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112199993B (en) | Method for identifying transformer substation insulator infrared image detection model in any direction based on artificial intelligence | |
| CN114462555B (en) | Multi-scale feature fusion power distribution network equipment identification method based on raspberry group | |
| CN112184692B (en) | Multi-target detection method for power transmission line | |
| CN111914720B (en) | Method and device for identifying insulator burst of power transmission line | |
| CN112507896B (en) | Method for detecting cherry fruits by adopting improved YOLO-V4 model | |
| CN112818905B (en) | Finite pixel vehicle target detection method based on attention and spatio-temporal information | |
| CN111428815B (en) | Mechanical arm grabbing detection method based on Anchor angle mechanism | |
| CN112750125B (en) | Glass insulator piece positioning method based on end-to-end key point detection | |
| CN115272204A (en) | Bearing surface scratch detection method based on machine vision | |
| CN112464704B (en) | Remote sensing image recognition method based on feature fusion and rotating target detector | |
| CN114862768A (en) | Improved YOLOv5-LITE lightweight-based power distribution assembly defect identification method | |
| CN116385958A (en) | Edge intelligent detection method for power grid inspection and monitoring | |
| CN112884715A (en) | Composite insulator grading ring inclination fault detection method based on deep learning | |
| CN112633088A (en) | Power station capacity estimation method based on photovoltaic component identification in aerial image | |
| CN115984238A (en) | Power grid insulator defect detection method and system based on deep neural network | |
| CN109919215B (en) | Target detection method for improving characteristic pyramid network based on clustering algorithm | |
| Liao et al. | Lr-cnn: Local-aware region cnn for vehicle detection in aerial imagery | |
| CN115713546A (en) | Lightweight target tracking algorithm for mobile terminal equipment | |
| CN114299354A (en) | Insulator detection positioning method based on rotating frame identification network | |
| CN113989296A (en) | Unmanned aerial vehicle wheat field remote sensing image segmentation method based on improved U-net network | |
| CN112507929A (en) | Vehicle body spot welding slag accurate detection method based on improved YOLOv3 network | |
| CN112329798B (en) | Image scene classification method based on optimized visual word bag model | |
| CN111353509B (en) | Key point extractor generation method of visual SLAM system | |
| CN113837088A (en) | Point cloud scene re-identification method based on semantic scanning context | |
| CN114550197A (en) | Terminal strip image detection information matching method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |