A Deep Learning-Based Method for Small Target Defect Detection in Transmission Lines

Abstract

This invention discloses a deep learning-based method for detecting small target defects in power transmission lines. The method includes: acquiring initial image data of the power transmission line from a front-end camera; processing the initial image data to obtain image data with characteristic directions; performing image recognition and image feature extraction on the image data with characteristic directions to filter out defective image data; classifying the filtered defective image data into corresponding small target defect categories to obtain defective image data corresponding to different small target defects; establishing a deep learning-based small target defect detection model based on the defective image data corresponding to different small target defects; and using the deep learning-based small target defect detection model to detect the acquired image data to obtain the corresponding small target defect results. This invention can accurately identify and detect defects in small targets on power transmission lines without requiring manual operation at high altitudes, thus improving detection efficiency and ensuring high safety.

Description

Technical Field

[0001] This invention relates to the field of power transmission line defect detection, and specifically to a method for detecting small target defects in power transmission lines based on deep learning. Background Technology

[0002] With the rapid development of my country's power industry, the layout of transmission lines has become increasingly complex. Transmission lines are divided into overhead transmission lines and cable lines. Overhead transmission lines consist of towers, conductors, hardware, insulators, guy wires, grounding devices, etc., and are widely distributed, spanning various terrains including fields, urban areas, deserts, and lakes. The natural environment and climate of transmission lines are highly variable. Due to their long-term operation outdoors, they are subjected to extreme weather conditions such as storms, heavy rain, and direct sunlight, inevitably suffering damage from various human and non-human factors. Components such as conductors, hardware, and insulators are prone to corrosion, breakage, and broken strands. Furthermore, improper installation of components also poses hidden dangers to the safe operation of transmission lines. Therefore, regular inspection and maintenance of transmission lines are necessary.

[0003] Small target samples in transmission lines include line hardware, insulators, guy wires, grounding devices, etc. Traditionally, the inspection of small target samples mainly relies on manual inspection or RCNN algorithm. The former requires professional personnel to go to the inspection site and manually climb in a high-voltage environment. Manual inspection not only affects the work efficiency of the staff and is prone to inaccurate inspection results, but also poses great safety problems. The latter inspection method is prone to image distortion, has large memory consumption, slow detection speed, and inaccurate detection accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting small target defects in power transmission lines based on deep learning, which can solve the problems of image distortion, slow detection speed and inaccurate detection accuracy caused by existing detection methods.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] This invention provides a method for detecting small target defects in power transmission lines based on deep learning, comprising the following steps:

[0007] Step S1: Obtain initial image data of the power transmission line captured by the front-end camera;

[0008] Step S2: Process the initial image data to obtain image data with characteristic orientation;

[0009] Step S3: Perform image recognition and image feature extraction on image data with characteristic orientations to filter out defective image data;

[0010] Step S4: Divide the selected defect image data into corresponding small target defect categories to obtain defect image data corresponding to different small target defects;

[0011] Step S5: Based on the defect image data corresponding to different small target defects, establish a small target defect detection model based on deep learning;

[0012] Step S6: Use a deep learning-based small target defect detection model to detect the collected image data and obtain the corresponding small target defect results.

[0013] Furthermore, the processing of the initial image data to obtain image data with characteristic orientation specifically includes:

[0014] Step S201: Obtain the resolution of each image in the initial image data to obtain image data of the same resolution;

[0015] Step S202: Extract contours from image data of the same resolution to obtain contour curves of the image data;

[0016] Step S203: Calculate the derivative of the contour curve to obtain its curvature;

[0017] Step S204: Take the center point as the feature point of the contour curve, and obtain the direction vector of the feature point pointing to the beginning of the contour curve and the direction vector of the feature point pointing to the end of the contour curve.

[0018] Step S205: Summate the direction vectors of the feature points pointing to the beginning of the contour curve and the direction vectors of the feature points pointing to the end of the contour curve to obtain image data with feature directions.

[0019] Furthermore, the step of performing image recognition and image feature extraction on image data with characteristic directions to filter out defective image data includes:

[0020] Image recognition is performed on image data with characteristic orientations using the grayscale method. Image features are then extracted from the recognized image data, and defective image data is selected based on the extracted image features.

[0021] Furthermore, the initial image data includes defective image data and normal image data.

[0022] The beneficial effects of this invention are:

[0023] This invention processes image data of power transmission lines to obtain image data with characteristic directions, and establishes a small target defect detection model based on the processed image data, thereby realizing defect identification and detection of small targets in power transmission lines with high accuracy. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram illustrating the steps of an embodiment of a deep learning-based method for detecting small target defects in power transmission lines.

[0026] Figure 2 This is a schematic diagram illustrating the steps of another embodiment of a deep learning-based method for detecting small target defects in power transmission lines.

[0027] Figure 3 This is a schematic diagram of the RCNN model algorithm for defect detection in the prior art. Detailed Implementation

[0028] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0029] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0030] This invention provides a method for detecting small target defects in power transmission lines based on deep learning, comprising the following steps:

[0031] Step S1: Obtain initial image data of the power transmission line captured by the front-end camera;

[0032] The front-end camera can be a camera mounted on a drone, used to capture image data for the automatic inspection of power transmission lines. The front-end camera is not limited to cameras mounted on drones; it can also be any other camera capable of capturing images of power transmission lines. No specific limitations are made here.

[0033] The initial image data includes defective image data and normal image data.

[0034] Step S2: Process the initial image data to obtain image data with characteristic orientation;

[0035] Step S3: Perform image recognition and image feature extraction on image data with characteristic orientations to filter out defective image data;

[0036] Step S4: Divide the selected defect image data into corresponding small target defect categories to obtain defect image data corresponding to different small target defects;

[0037] Small target defects include fittings, insulators, guy wires, grounding devices, etc.

[0038] Step S5: Based on the defect image data corresponding to different small target defects, establish a small target defect detection model based on deep learning;

[0039] The defect image data needs to exceed 200,000 images. The defect image data is stored in an XML file, which contains the image and the annotation information of the wires, hardware, insulators, etc., including the image name, image size, target category and target rectangle size.

[0040] Step S6: Use a deep learning-based small target defect detection model to detect the collected image data and obtain the corresponding small target defect results.

[0041] Furthermore, in a preferred embodiment of this application, the processing of the initial image data to obtain image data with characteristic orientation specifically includes:

[0042] Step S201: Obtain the resolution of each image in the initial image data to obtain image data of the same resolution;

[0043] Step S202: Extract contours from image data of the same resolution to obtain contour curves of the image data;

[0044] The formula for the contour curve (X,Y) is:

[0045] in, For points on the surface that change over time, Angle that changes over time;

[0046] Step S203: Calculate the derivative of the contour curve to obtain its curvature;

[0047] Step S204: Take the center point as the feature point of the contour curve, and obtain the direction vector of the feature point pointing to the beginning of the contour curve and the direction vector of the feature point pointing to the end of the contour curve.

[0048] Step S205: Summate the direction vectors of the feature points pointing to the beginning of the contour curve and the direction vectors of the feature points pointing to the end of the contour curve to obtain image data with feature directions.

[0049] The process of performing image recognition and image feature extraction on image data with characteristic orientations to filter out defective image data includes:

[0050] Image recognition is performed on image data with characteristic orientations using the grayscale method. The formula for image recognition c(x,y) is:

[0051]

[0052] in:

[0053] (x,y) represents the position of a pixel in image data with characteristic orientation;

[0054] p represents the gray level of image data with characteristic orientation;

[0055] i represents the total number of pixels in the image data with characteristic orientation;

[0056] g(x,y) represents the grayscale coefficient of a pixel in an image with a characteristic orientation;

[0057] Then, after image recognition, image feature extraction is performed. The formula for image feature extraction p(x,y) is:

[0058]

[0059] in:

[0060] Q(x,y) represents grayscale image data;

[0061] J represents the mean gray level of the grayscale image. Based on the image features, p(x,y) is extracted to filter out defective image data.

[0062] It should be noted that when performing image recognition on image data with characteristic orientation using the grayscale method, it is necessary to ensure the image clarity of the image data with characteristic orientation, and then extract features from the image, and simultaneously extract features from images of the same transmission line from different angles.

[0063] Furthermore, in a preferred embodiment of this application, establishing a small target defect detection model based on deep learning includes:

[0064] The loss function of Faster R-CNN for small target defect detection is established as follows:

[0065]

[0066] in:

[0067] L1 is the classification loss function;

[0068] L2 is the regression loss function;

[0069] a i Let the region of the i-th anchor point be the target;

[0070] a i1 Classify the probability of small target defects in transmission lines;

[0071] x i The parameterized feature vector representing the i-th anchor point region in the small target defect sample of the transmission line;

[0072] x i1 A parameterized feature vector representing the true target region;

[0073] λ is the weight parameter;

[0074] N1 and N2 represent the normalization parameters of the loss function;

[0075] In the loss function of Faster R-CNN:

[0076] L1(a i ,a i1 )=-log[a i a i1 +(1-a i (1-a) i1 )],

[0077]

[0078] `smooth` is the robustness function of the regression process.

[0079] The probability of a target is calculated from the defect image data using the Region Candidate Network (RPN), and candidate boxes are generated.

[0080] The RPN network and the Faster RCNN detection model share convolutional layers. The parameters and weights of the non-shared convolutional networks in the RPN network are randomly initialized using a Gaussian distribution with a standard deviation of 0.01 and a mean of 0.

[0081] The RPN network is trained using training samples, and the candidate box regions calculated by the RPN network are used to replace the Faster RCNN parameters trained in the Faster RCNN detection model.

[0082] The RPN network is initialized using the network parameters of the Faster RCNN detection model, and the parameters of the non-shared convolutional network in the RPN network are trained.

[0083] The pooling layer converts the candidate bounding box regions generated by the RPN network into feature vectors of fixed dimensions.

[0084] The Softmax activation function is used to identify objects in the candidate bounding box region and to mark the candidate regions in the defective image data.

[0085] The existing RCNN model algorithm mainly consists of four parts, as follows:

[0086] Candidate region generation: Multiple candidate regions are generated from the input image data. These candidate regions can overlap and contain each other. By utilizing information such as texture, color, and edges in the image, fewer windows are selected.

[0087] Feature extraction: For each candidate region, a CNN network is used to extract feature vectors;

[0088] Classification: Linear Support Vector Machine (SVM) is used to classify the target based on the features extracted by the network;

[0089] Boundary regression: Uses a regressor to refine the position of candidate boxes, generating more stringent bounding boxes for the identified objects.

[0090] Compared to existing RCNN models, the Faster RCNN model algorithm in the deep learning-based small target defect model of transmission lines offers the following advantages:

[0091] The Faster R-CNN algorithm further improves the speed of object detection. The main improvement is the use of Region Proposal Networks (RPNs) instead of traditional region search methods to select candidate boxes, significantly increasing computational speed without sacrificing detection accuracy. The RPN network divides the feature maps extracted from the convolutional layers into different regions using sliding windows. Each sliding window is fully connected and uses ReLU as the activation function to generate a feature vector. This feature vector is connected to two fully connected layers, generating a 4k-dimensional vector for bounding box regression and a 2k-dimensional vector for classification. Then, reference boxes of different sizes and scales, also known as anchor boxes, are used in each region. In the Faster R-CNN algorithm, the RPN network shares convolutional layers with the Faster R-CNN detection module. Experimental data shows that the Faster R-CNN algorithm generally achieves a defect detection efficiency of 98%, while the R-CNN algorithm achieves approximately 75%. The defect detection efficiency of the Faster R-CNN algorithm is significantly higher than that of the R-CNN algorithm.

[0092] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0093] The above description is merely illustrative of the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc., made without creative effort within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting small target defects in transmission lines based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain initial image data of the power transmission line captured by the front-end camera; Step S2: Process the initial image data to obtain image data with characteristic orientation; Step S3: Perform image recognition and image feature extraction on image data with characteristic orientations to filter out defective image data; Step S4: Divide the selected defect image data into corresponding small target defect categories to obtain defect image data corresponding to different small target defects; Step S5: Based on the defect image data corresponding to different small target defects, establish a small target defect detection model based on deep learning; Step S6: Use a deep learning-based small target defect detection model to detect the collected image data and obtain the corresponding small target defect results; The process of performing image recognition and image feature extraction on image data with characteristic orientations to filter out defective image data includes: Image recognition is performed on image data with characteristic orientations using the grayscale method. Image features are then extracted from the recognized image data. Based on the extracted image features, defective image data is selected. The grayscale method is used for image recognition of image data with characteristic orientations. The formula for image recognition c(x,y) is: in: (x,y) represents the position of a pixel in image data with characteristic orientation; p represents the gray level of image data with characteristic orientation; i represents the total number of pixels in the image data with characteristic orientation; g(x,y) represents the grayscale coefficient of a pixel in an image with a characteristic orientation; Image feature extraction is performed after image recognition. The formula for image feature extraction p(x,y) is: in: Q(x,y) represents grayscale image data; J represents the mean gray level of the grayscale image. Based on the image features, p(x,y) is extracted to filter out defective image data.

2. The method for detecting small target defects in transmission lines based on deep learning according to claim 1, characterized in that, The process of processing the initial image data to obtain image data with characteristic orientation specifically includes: Step S201: Obtain the resolution of each image in the initial image data to obtain image data of the same resolution; Step S202: Extract contours from image data of the same resolution to obtain contour curves of the image data; Step S203: Calculate the derivative of the contour curve to obtain its curvature; Step S204: Take the center point as the feature point of the contour curve, and obtain the direction vector of the feature point pointing to the beginning of the contour curve and the direction vector of the feature point pointing to the end of the contour curve. Step S205: Summate the direction vectors of the feature points pointing to the beginning of the contour curve and the direction vectors of the feature points pointing to the end of the contour curve to obtain image data with feature directions.

3. The method for detecting small target defects in transmission lines based on deep learning according to claim 1 or 2, characterized in that, The initial image data includes defective image data and normal image data.

Images