Power transmission line icing thickness detection method and system based on deep learning

Abstract

The invention discloses a power transmission line icing thickness detection method and system based on deep learning, and relates to the technical field of image processing and power transmission line online detection, and the method comprises the steps: collecting an icing image of a power transmission line through a standardized dual-light optical device, and carrying out the image preprocessing, and generating an original data set; and performing image segmentation on the original data set, training the deep neural network model by using the lightweight YOLACT network, and performing evaluation. Edge detection is carried out through a multi-direction Sobel operator template, an Otsu method is adopted for binarization processing and nonlinear operator filtering denoising, a boundary chain code array is extracted based on a boundary tracking algorithm of a topological structure, and the MASK of an original power transmission line is determined. And acquiring an image of the power transmission line by using a dual-light camera, performing camera calibration, performing registration and fusion on the calibrated image, and outputting a calculation result of the icing thickness of the power transmission line by using the trained neural network. According to the invention, the detection precision and efficiency are improved, and the thickness estimation deviation caused by image errors is reduced.

Description

Technical Field

[0001] This invention relates to the fields of image processing and online inspection technology of power transmission lines, specifically to a method and system for detecting the icing thickness of power transmission lines based on deep learning. Background Technology

[0002] The earliest research on icing conditions of transmission lines, both domestically and internationally, was conducted in the mid-20th century. Initially, the monitoring of icing conditions on transmission lines was not widely promoted because the efficiency of manual inspections was very low and the cycle of repeated inspections was very long due to the influence of geographical environment and weather conditions.

[0003] Currently, the common methods for detecting icing thickness are direct manual measurement and indirect image comparison. The former suffers from low manual efficiency and personnel safety risks, while the latter suffers from low accuracy and cumbersome operation. Therefore, we propose a deep learning algorithm-based method for detecting icing thickness on power transmission lines. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, to solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for detecting the icing thickness of transmission lines based on deep learning, comprising: acquiring icing images of transmission lines using standardized dual-light optical equipment, and performing image preprocessing on the icing images to generate an original dataset;

[0006] Image segmentation is performed on the original dataset, and a deep neural network model is trained using a lightweight YOLACT network. The trained deep neural network model is then evaluated.

[0007] Edge detection is performed by designing a multi-directional Sobel operator template, and binarization is performed by Otsu's method and filtering and denoising are performed by nonlinear operators. The boundary chain code array is extracted based on the topology-based boundary tracing algorithm to determine the original transmission line mask.

[0008] Images of the transmission line are acquired using a dual-light camera and the camera is calibrated. The calibrated images are then registered and fused, and the calculated ice thickness of the transmission line is output using a trained neural network.

[0009] As a preferred embodiment of the deep learning-based transmission line icing thickness detection method of the present invention, the image preprocessing includes size transformation, brightness transformation, color enhancement, contrast enhancement and image denoising.

[0010] The image segmentation includes feature extraction, mask generation, object detection, and instance segmentation;

[0011] High-precision instance segmentation of transmission line icing images is achieved by using a single-stage instance segmentation algorithm and a polar coordinate representation encoding mask method.

[0012] The single-stage instance segmentation algorithm includes,

[0013] Feature extraction is performed using a backbone network, and the extracted features are processed in two branches: a classification branch and a regression branch.

[0014] The classification branch includes category classification and polar coordinate centrality prediction, which are used to determine which category a pixel belongs to and the confidence level of the pixel's distance from the center, respectively.

[0015] Using a point in the feature map as the center point of polar coordinates, generate N rays and calculate the ray distance from each target edge to the center point;

[0016] The target edge is reconstructed using ray distance to determine the optimal mask shape;

[0017] In each FPN layer, 1000 best samples are retained and sorted and filtered based on the product of classification confidence score and polar coordinate centrality score;

[0018] Based on the known ray distance, connect the coordinates of the edge points sequentially, starting from 0°, to form the outline of the target;

[0019] After inputting remote sensing images, image features are extracted in the feature backbone network;

[0020] Image features are input into two branches. The prototype mask generation network branch generates a set of prototype masks of the same size as the original image through a fully convolutional network.

[0021] The mask coefficient branch is used to generate mask coefficients for each instance and predict the category and bounding box location of the candidate boxes;

[0022] The prototype mask is linearly combined with the corresponding mask coefficients to generate the instance mask;

[0023] The predicted bounding box is used to clip the mask to obtain the instance segmentation result.

[0024] As a preferred embodiment of the deep learning-based transmission line icing thickness detection method of the present invention, the step of training the deep neural network model includes:

[0025] The input images are preprocessed and resized uniformly.

[0026] By passing the data through the ResNet backbone network and FPN layer, multi-scale features are obtained, and the multi-scale features are input into the parallel detection module for coefficient combination prediction.

[0027] The p3 feature map output from the feature pyramid is input into the prototype network to generate a pre-segmentation model and predict the prototype mask of the image.

[0028] A fast nonmaximum suppression method is used to remove interfering detection boxes, and a prototype mask is combined to generate predicted boxes and masks.

[0029] Calculate the loss functions for the detection box, classification, and mask respectively, and update the weights using gradient backpropagation.

[0030] Repeat the above steps until the model satisfies the stability condition;

[0031] The evaluation of the trained deep neural network model includes,

[0032] The deep neural network model is evaluated using training evaluation metrics;

[0033] The training evaluation metrics include,

[0034] Each PNG image label corresponds to a mask of the target region in the original image, and all PNG image labels are binary images;

[0035] The white portion represents the foreground of the corresponding source image, and the black portion represents the background.

[0036] Convert the XML file generated by LabelImg annotation into a PNG image tag using a self-written program;

[0037] Both object detection and instance segmentation algorithms use mAP (memory AP) as a training evaluation metric.

[0038] The mAP value is calculated as shown in the following formula:

[0039]

[0040] Where C is the number of categories of the target, AP i Let be the average precision of the i-th category.

[0041] As a preferred embodiment of the deep learning-based transmission line icing thickness detection method of the present invention, the design of the multi-directional Sobel operator template includes,

[0042] Two templates with different orientations are added to the traditional Sobel operator: 0°, 90°, 45°, and 135°.

[0043] Edge information is extracted using templates in four directions. The template matrices for the four directions are shown below:

[0044]

[0045] The Sobel templates in the four directions mentioned above are applied to the input infrared image for convolution operations.

[0046] Calculate the weighted difference of the neighborhood of each pixel in the image, determine the extreme value reached by the weighted difference, and set the extreme value as the threshold.

[0047] When the weighted difference of pixels is greater than the threshold, the pixel is retained as an edge point;

[0048] When the weighted difference of pixels is less than the threshold, the pixel is removed to determine the target edge;

[0049] The binarization process using Otsu's method includes...

[0050] The infrared image is converted to grayscale and divided into L grayscale levels. The probability of each grayscale level is defined and calculated using the following formula:

[0051]

[0052] Where, n i p is the number of pixels with gray level i, N is the total number of pixels in the image, and p i Let i be the probability of gray level i.

[0053] The grayscale image is divided into two categories based on its grayscale value. When 1 ≤ grayscale value ≤ k, the grayscale image is classified into one category of grayscale image, denoted as C0.

[0054] When k+1≤gray value≤L, the gray image is divided into two classes of gray images, denoted as C1;

[0055] The probability of occurrence for each category is calculated using the following formula:

[0056]

[0057] Where ω0 is the probability of a grayscale image C0 of class 1, ω1 is the probability of a grayscale image C1 of class 2, and ω(k) is the cumulative probability of grayscale values ​​from grayscale level 1 to k.

[0058] The formula for calculating the class mean is as follows:

[0059]

[0060] Where μ0 is the C0 mean of the first-class grayscale image, μ1 is the C1 mean of the second-class grayscale image, and μ(k) is the cumulative average value. T The overall mean of the image;

[0061] The formula for calculating class variance is as follows:

[0062]

[0063] in, Let C0 be the variance. Let C1 be the variance;

[0064] Three criteria for evaluating the quality of binarization are defined, and the calculation formulas are as follows:

[0065]

[0066]

[0067] in, For within-class variance, For inter-class variance, Let λ be the population variance, λ be the ratio of between-class variance to within-class variance, κ be the ratio of population variance to within-class variance, and η be the ratio of between-class variance to population variance.

[0068] The optimal binarization threshold is found by iterating through the k values, and the calculation formula is as follows:

[0069]

[0070] in, The optimal binarization threshold is k, where k is the number of iterations.

[0071] Achieve optimal infrared image binarization processing.

[0072] As a preferred embodiment of the deep learning-based transmission line icing thickness detection method of the present invention, the boundary tracing algorithm based on topology structure includes the determination of the starting point, boundary discrimination criteria, and search termination criteria.

[0073] The MASK for determining the original transmission line includes,

[0074] The boundary tracing algorithm can extract the boundary of the power transmission line in the infrared image in the form of a chain code array to obtain the edge data of the power transmission line;

[0075] Guided by the acquired edge data, the correct boundary of the transmission line area, namely the boundary chain code array, was selected using the edge data;

[0076] Find the minimum bounding rectangle of the boundary chain code array to obtain the optimal original transmission line mask.

[0077] As a preferred embodiment of the deep learning-based transmission line icing thickness detection method of the present invention, the camera calibration includes camera intrinsic parameters and camera extrinsic parameters;

[0078] Camera parameters include focal length, principal point coordinates, pixel size, skew coefficient, and distortion;

[0079] The steps to obtain camera intrinsic parameters include:

[0080] The coordinates of the coordinate points are obtained through measurement or software generation.

[0081] Take multiple sets of pictures with the camera pointed at the calibration board, including the entire plane of the calibration board. Fix the camera on a tripod and take 30 pictures of the chessboard calibration board from different distances and angles.

[0082] Feature points on the calibration board are extracted from the captured images, and feature extraction algorithms are used to obtain the feature points.

[0083] Camera parameters are calculated by minimizing reprojection error and using an optimization algorithm;

[0084] The steps for obtaining camera extrinsic parameters based on vanishing points include:

[0085] The distance between the camera and the power line is obtained by using a laser rangefinder, which gives the specific location of the pitch and yaw angles;

[0086] When a three-dimensional spatial image is projected onto a two-dimensional plane, perspective distortion occurs, meaning that two lines that were originally parallel in the three-dimensional space intersect at a single point in the two-dimensional plane, which is the vanishing point.

[0087] When two parallel lines in three-dimensional space are parallel to the image plane, the vanishing point is located at infinity on the image plane; otherwise, the vanishing point is located at a finite distance on the image plane. That is, when there is a yaw angle, the x-coordinate of the vanishing point will be shifted.

[0088] When a pitch angle exists, the ordinate of the vanishing point will shift;

[0089] If the dashed rectangular area is the standard imaging plane, that is, the camera has no pitch and yaw angles, then the pixel coordinates of the vanishing point are V(u0,v0);

[0090] If the black rectangular area is the imaging plane when there are yaw and pitch angles, then the vanishing point coordinates are V(u1,v1);

[0091] When the camera has a yaw angle, the formula for calculating the yaw angle, based on geometric relationships, is as follows:

[0092]

[0093] Where m is the size of the imaging plane, f is the camera focal length, γ is the camera yaw angle, and l is the distance between the ordinate of the vanishing point and the ordinate of the center point of the image plane;

[0094] When the camera has a pitch angle, the formula for calculating the pitch angle, based on geometric relationships, is as follows:

[0095]

[0096] Where θ is the camera pitch angle and n is the size of the imaging plane.

[0097] As a preferred embodiment of the deep learning-based transmission line icing thickness detection method of the present invention, the image registration and fusion include:

[0098] Optical RGB images and infrared images are used for mask registration and fusion of the transmission line area;

[0099] Registration error σ x The calculation formula is as follows:

[0100]

[0101] Where f is the focal length, I pix d is the pixel size. c D is the baseline length. optimal D is the target distance. target For a specific target distance;

[0102] The calculation of standard icing thickness for power transmission lines includes,

[0103] Calculated according to the standard formula for calculating the thickness of icing on power transmission lines;

[0104] The formula for calculating the standard icing thickness B0 of transmission lines is as follows:

[0105]

[0106] Where ρ is the estimated density of icing, r is the original transmission line radius, and K s R is the icing shape coefficient, and R is the icing radius.

[0107] Another objective of this invention is to provide a deep learning-based system for detecting the icing thickness of transmission lines. To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning-based system for detecting the icing thickness of transmission lines, comprising: a data acquisition module, an image processing module, an edge detection module, a camera calibration module, and an image fusion module;

[0108] The data acquisition module is used for image acquisition and image preprocessing;

[0109] The image processing module is used to process image segmentation, instance segmentation, and feature extraction.

[0110] The edge detection module is used to perform edge detection, and uses the Otsu method for binarization, nonlinear operators for filtering and noise reduction, and extracts boundaries.

[0111] The camera calibration module is used to obtain and calibrate the camera's intrinsic and extrinsic parameters.

[0112] The image fusion module is used to perform image registration and fusion on the transmission line images after camera calibration, and to output the calculation results of the ice thickness using a trained neural network.

[0113] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of the deep learning-based transmission line icing thickness detection method described above.

[0114] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the deep learning-based transmission line icing thickness detection method described above.

[0115] The beneficial effects of this invention are as follows: This invention acquires icing images of transmission lines using standardized dual-light optical equipment and performs preprocessing to generate a high-quality raw dataset, providing reliable basic data for subsequent deep learning model training; through image segmentation and lightweight YOLACT network training, accurate identification and segmentation of icing areas in the images are achieved, reducing computational complexity and improving model training and inference speed; a multi-directional Sobel operator template is designed for multi-directional edge detection, combined with Otsu's binarization method and nonlinear operator filtering for noise reduction, ensuring the accuracy of edge detection and image clarity, generating a mask of the original transmission line; through camera calibration and image registration and fusion using dual-light cameras, accurate image alignment and fusion are achieved, and the icing thickness is calculated using a deep learning model, providing high-precision thickness estimation; these steps comprehensively improve detection accuracy and efficiency, providing reliable data support for the safety monitoring of transmission lines and reducing thickness estimation deviations caused by image errors. Attached Figure Description

[0116] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0117] Figure 1 This is a schematic diagram of the terminal structure in the first embodiment of the present invention;

[0118] Figure 2 This is a flowchart illustrating the process in the first embodiment of the present invention;

[0119] Figure 3This is a schematic diagram of the structure of the computer device in the third embodiment of the present invention;

[0120] Figure 4 This is a schematic diagram illustrating the calculation principle of the equivalent ice thickness in the first embodiment of the present invention.

[0121] Figure 5 This is a diagram illustrating the Square boundary tracking algorithm and its limitations in the first embodiment of the present invention. Detailed Implementation

[0122] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0123] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0124] Example 1

[0125] Reference Figures 1-5 This invention provides a method for detecting the icing thickness of transmission lines based on deep learning, as one embodiment of the present invention.

[0126] First, the deep learning-based method for detecting icing thickness on transmission lines provided in this application can be applied to, for example... Figure 1 In the terminal shown. For example... Figure 1 As shown, the terminal may include one or two ( Figure 1 Only one processor and a memory for storing data are shown in the diagram. The processor may be, but is not limited to, a processing system such as a microprocessor (MCU) or a programmable logic device (FPGA). The terminal may also include transmission devices for communication functions and input / output devices. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the terminal described above. For example, the terminal may also include components that are larger than... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0127] The memory can be used to store computer programs, such as the computer program corresponding to the deep learning-based transmission line icing thickness detection method in this embodiment. The processor executes various functional applications and data processing by running the computer program stored in the memory, thereby implementing the above-described method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage systems, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0128] Transmission devices are used to receive or send data over a network. This network includes wireless networks provided by the terminal's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0129] like Figure 2 As shown, this embodiment of the invention provides a deep learning-based method for detecting the icing thickness of transmission lines, which is applied to... Figure 1 Taking the terminal in the example, the explanation includes the following steps:

[0130] S1: Acquire images of icing on transmission lines using standardized dual-optical equipment, and perform image preprocessing on the icing images to generate the original dataset.

[0131] S1.1: Image preprocessing includes size transformation, brightness transformation, color enhancement, contrast enhancement, and image denoising.

[0132] S2: Perform image segmentation on the original dataset, train a deep neural network model using the lightweight YOLACT network, and evaluate the trained deep neural network model.

[0133] S2.1: Image segmentation includes feature extraction, mask generation, object detection, and instance segmentation;

[0134] High-precision instance segmentation of transmission line icing images is achieved by using a single-stage instance segmentation algorithm and a polar coordinate representation encoding mask method.

[0135] Single-stage instance segmentation algorithms include,

[0136] Feature extraction is performed using a backbone network, and the extracted features are processed in two branches: a classification branch and a regression branch.

[0137] The classification branch includes category classification and polar coordinate centrality prediction, which are used to determine which category a pixel belongs to and the confidence level of the pixel's distance from the center, respectively.

[0138] Using a point in the feature map as the center point of polar coordinates, generate N rays (each ray is spaced 10 degrees apart, for a total of 36 rays);

[0139] Calculate the ray distance from the edge of each target to the center point;

[0140] Use these distances to reconstruct the target edges and determine the optimal mask shape;

[0141] Each FPN layer retains 1000 best samples, which are sorted and filtered based on the product of classification confidence score and polar coordinate centrality score;

[0142] Finally, based on the known ray distances, the coordinates of the edge points are connected sequentially, starting from 0°, to form the outline of the target.

[0143] S2.2: The lightweight YOLACT network includes,

[0144] YOLACT network is a real-time, end-to-end fully convolutional network model for instance segmentation. After inputting remote sensing images, it extracts image features in the feature backbone network.

[0145] Image features are input into two branches. The prototype mask generation network branch generates a set of prototype masks of the same size as the original image through a fully convolutional network.

[0146] The mask coefficient branch is used to generate mask coefficients for each instance, predicting the category and bounding box location of the candidate boxes;

[0147] The prototype mask is linearly combined with the corresponding mask coefficients to generate the instance mask;

[0148] The predicted bounding box is used to perform a final cropping of the mask to obtain the final instance segmentation result.

[0149] S2.3: Training a deep neural network model includes,

[0150] Model training parameters include image size, backbone network, batch size, learning rate, and number of iterations. PNG image labels and the original image are used as input. Examples of model training parameters are shown in the table below.

[0151] Parameter name Parameter value Image size 640*640 Main network ResNet50+FPN Batch size 3 Learning rate 0.0001 Number of iterations 360000

[0152] The model training steps are as follows:

[0153] The input images are preprocessed and resized uniformly.

[0154] The data is processed through the ResNet backbone and FPN to obtain multi-scale features, and the multi-scale features are then input into a parallel detection module for coefficient combination prediction.

[0155] The p3 feature map output from the feature pyramid is input into the prototype network to generate a pre-segmentation model and predict the prototype mask of the image.

[0156] The fast nonmaximum suppression method is used to remove interfering detection boxes, and then the prototype mask in the third step is combined to generate predicted boxes and masks;

[0157] Calculate the loss functions for the detection box, classification, and mask respectively, and then update the weights using gradient backpropagation.

[0158] Repeat the above steps until the model meets the stability condition or other requirements;

[0159] This model ensures high accuracy in traverse images with straight-line similarity features in complex backgrounds. It can achieve accurate traverse image segmentation in complex backgrounds such as woodlands, fog, soil, and local freezing, and can also segment traverse images considering different resolutions, traverse bending, and irregular icing.

[0160] S2.4: Evaluating the trained deep neural network model includes,

[0161] The deep neural network model is evaluated using training evaluation metrics;

[0162] Training evaluation metrics include,

[0163] The data source is RGB images of icing on power transmission lines taken by dual-light cameras, and the images are labeled with LabelImg. The dataset label format is PNG images.

[0164] Each PNG image label corresponds to a mask of the target region in the original image. All PNG image labels are binary images, with white parts representing the foreground (region of interest) of the corresponding source image and black parts representing the background.

[0165] A custom program can be used to convert XML files generated by LabelImg annotations into PNG image tags.

[0166] The mAP value is often used to measure the performance of object detection algorithms and instance segmentation algorithms;

[0167] The mAP value is defined as the average of the AP (Average Precision) values ​​for all categories, calculated as follows:

[0168]

[0169] Here, C represents the number of target categories. This paper includes two categories: pig and pig_bad. AP value, also known as average precision, is commonly used in object detection algorithms and instance segmentation algorithms, such as Faster R-CNN, SSD, and Mask R-CNN, to quantitatively evaluate the performance of the algorithm. The AP value is obtained by calculating the recall and average precision between 0 and 1.

[0170] S3: Edge detection is performed by designing a multi-directional Sobel operator template, and binarization is performed using the Otsu method and nonlinear operators are used for filtering and denoising. The boundary chain code array is extracted based on the boundary tracing algorithm of the topology structure to determine the MASK of the original transmission line.

[0171] S3.1: Multi-directional Sobel edge operators include,

[0172] The direction of power transmission lines in infrared images is uncertain, and the classic Sobel operator only uses templates in the horizontal and vertical directions, resulting in low accuracy in extracting the edges of power transmission lines.

[0173] To more accurately extract edge information of transmission lines in infrared images, this project uses a multi-directional Sobel operator. This involves adding two templates with different orientations—0°, 90°, 45°, and 135°—to the traditional Sobel operator to extract edge information. The templates are shown in the following four matrices:

[0174]

[0175] The Sobel templates in the four directions mentioned above are applied to the input infrared image for convolution operations. The closer the value is to the pixel, the greater the weighting.

[0176] Calculate the weighted difference of the neighborhood of each pixel in the image, determine the extreme value reached by the weighted difference, and set the extreme value as the threshold.

[0177] When the weighted difference of pixels is greater than the threshold, the point is retained as an edge point;

[0178] When the weighted difference of pixels is less than the threshold, the pixel is removed, and the target edge can be determined.

[0179] The edge results obtained by processing with the multi-directional Sobel operator are used for subsequent region masking to improve the accuracy of region extraction.

[0180] S3.2: Image binarization processing based on Otsu's method includes...

[0181] In order to solve the boundary of the transmission line region in the infrared image, the infrared image needs to be converted to grayscale and then binarized.

[0182] To obtain the best binarized image, Otsu's method was used for processing;

[0183] The grayscale image is divided into L grayscale levels, and the probability of each grayscale level is defined as follows:

[0184]

[0185] Where, n i p is the number of pixels with gray level i, N is the total number of pixels in the image, and p i Let i be the probability of gray level i.

[0186] Binarization processing includes,

[0187] The grayscale image is divided into two categories based on its grayscale value. When 1 ≤ grayscale value ≤ k, the grayscale image is classified into one category of grayscale image, denoted as C0.

[0188] When k+1≤gray value≤L, the gray image is divided into two classes of gray images, denoted as C1;

[0189] The probability of each class occurring is:

[0190]

[0191] Where ω0 is the probability of a grayscale image C0 of class 1, ω1 is the probability of a grayscale image C1 of class 2, and ω(k) is the cumulative probability of grayscale values ​​from grayscale level 1 to k.

[0192] At this point, the class mean is:

[0193]

[0194] Where μ0 is the C0 mean of the first-class grayscale image, μ1 is the C1 mean of the second-class grayscale image, and μ(k) is the cumulative average value. T The overall mean of the image;

[0195] Class variance is:

[0196]

[0197] in, Let C0 be the variance. Let C1 be the variance;

[0198] Three criteria are defined to evaluate the quality of binarization:

[0199]

[0200]

[0201] in, For within-class variance, For inter-class variance, Let λ be the population variance, λ be the ratio of between-class variance to within-class variance, κ be the ratio of population variance to within-class variance, and η be the ratio of between-class variance to population variance.

[0202] To obtain the maximum values ​​of λ, κ, and η, further discussion of their relationships reveals that...

[0203] κ=1+λ

[0204] η=λ / (1+λ)

[0205] It can be seen that as long as λ is maximized, all three discrimination indicators will be at their maximum values;

[0206] Since the above relationships exist, we can conclude that:

[0207]

[0208] From this we can know that, Regardless of the choice of the binarization threshold k, it is only necessary to make The maximum variance is the key factor, which is why Otsu's method exhibits the characteristic of maximizing inter-class variance.

[0209] Ultimately, the relationship between inter-class variance and binarization threshold can be derived:

[0210]

[0211] The optimal binarization threshold is found by iterating through the k values. The specific calculation formula is as follows:

[0212]

[0213] Therefore, the above method can be used to achieve optimal infrared image binarization processing.

[0214] S3.3: Image filtering based on nonlinear operators includes,

[0215] During the process of infrared sensor capturing images, when the imaging element or transmission system is unstable, the visualized infrared image is prone to generating dark spots and bright spots of varying sizes and locations. These types of spots are generally referred to as noise.

[0216] Noise can interfere with the acquisition of subsequent edge arrays. To reduce the impact of noise, image filtering is needed to denoise the image.

[0217] S3.4: Boundary tracing algorithms for topology structures include,

[0218] The boundary search algorithm consists of three parts: the determination of the starting point, the boundary discrimination criterion, and the search termination criterion.

[0219] Determining the starting point is crucial for boundary search in certain images, as different starting points may lead to tracking different boundaries.

[0220] Boundary discrimination criteria are the core of the boundary search algorithm, guiding the algorithm's next search direction and affecting the accuracy of the boundary search algorithm. They are usually based on changes in gray values ​​or gradient information.

[0221] The search termination criteria include determining when to end boundary tracking, which could be returning to the starting point or reaching a predetermined boundary length, etc.

[0222] This invention uses the Square algorithm for calculation.

[0223] S3.5: Original transmission area mask extraction includes,

[0224] The boundary tracing algorithm can extract the boundary of the power transmission line in the infrared image in the form of a chain code array to obtain the edge data of the power transmission line;

[0225] Guided by the acquired edge data, the correct boundary of the transmission line area, namely the boundary chain code array, was selected using the edge data;

[0226] Find the minimum bounding rectangle of the boundary chain code array to obtain the optimal original transmission line mask.

[0227] S4: Acquire images of the transmission line using a dual-light camera and perform camera calibration. Then, perform image registration and fusion on the calibrated images of the transmission line and output the calculated results of the ice thickness of the transmission line using a trained neural network.

[0228] S4.1: Physical model for calculating the icing thickness of transmission lines. Based on DL / T5158-2012(4.4.1) and national standard GB / T 35235-2017, the equivalent icing thickness of transmission lines is calculated using icing images.

[0229] The research is mainly based on the principle of pixel similarity, that is, the ratio of the number of pixels corresponding to the icing area to the line area is equal to the ratio of the equivalent thickness of the icing area to the diameter of the line.

[0230] S4.2: Camera calibration includes,

[0231] In order to correct and align the optical RGB image to infrared image, this project will conduct research on the camera's intrinsic and extrinsic parameter calibration process. Camera calibration mainly includes two aspects: intrinsic parameters and extrinsic parameters.

[0232] Determining camera intrinsic parameters mainly involves the following steps:

[0233] Prepare a calibration board with a series of points with known coordinates. The coordinates of these points can be obtained by measurement or generated by software.

[0234] Point the camera at the calibration board and take a series of pictures that cover the entire plane of the calibration board. Fix the camera on a tripod and take 30 pictures of the chessboard calibration board from different distances and angles.

[0235] Feature points on the calibration board are extracted from the captured images. These feature points can be obtained through various feature extraction algorithms, such as SIFT and SURF.

[0236] The camera's intrinsic parameters are calculated by minimizing the reprojection error. This process requires the use of optimization algorithms, such as the Levenberg-Marquardt algorithm.

[0237] By following the steps above, you can obtain the camera's intrinsic parameters, including focal length, principal point coordinates, pixel size, skew coefficient, and distortion.

[0238] With these parameters, the camera's imaging process can be accurately described, enabling various complex image processing tasks to be performed.

[0239] Methods for obtaining camera extrinsic parameters based on vanishing points include,

[0240] In computer vision and image processing, camera extrinsic parameters are parameters that describe the position and orientation of the camera in the world coordinate system. These parameters typically include rotation and translation matrices.

[0241] In this example, we only need to consider the camera's rotation, namely the pitch angle and yaw angle. These two angles can completely describe the camera's orientation, and its specific location can be obtained by using a laser rangefinder to determine the distance between the camera and the power line.

[0242] When a three-dimensional image is projected onto a two-dimensional plane, perspective distortion may occur, meaning that two lines that were originally parallel in three-dimensional space intersect at a single point in the two-dimensional plane, which is called the vanishing point.

[0243] When two parallel lines in three-dimensional space are parallel to the image plane, the vanishing point is located at infinity on the image plane; otherwise, the vanishing point is located at a finite distance on the image plane. That is, when there is a yaw angle, the x-coordinate of the vanishing point will be shifted.

[0244] When there is a pitch angle, the ordinate of the vanishing point will be shifted. Assume that m and n are the imaging plane dimensions, f is the camera focal length, 2α and 2β are the camera's vertical and horizontal field of view angles, and θ and γ are the camera's pitch and yaw angles, respectively.

[0245] Assuming the dashed rectangular area is the standard imaging plane, meaning the camera has no pitch or yaw angles, the vanishing point pixel coordinates are assumed to be V(u0,v0);

[0246] The black rectangular area is the imaging plane when there are yaw and pitch angles. In this case, the vanishing point coordinates are V(u1,v1).

[0247] Based on geometric relationships, the formula for calculating the yaw angle is as follows:

[0248]

[0249] Similarly, when the camera has a pitch angle, the calculation formula is as follows:

[0250]

[0251] Therefore, based on the above calculation method and combined with the camera's intrinsic parameters, the yaw angle and pitch angle during camera shooting can be calculated.

[0252] S4.3: Image registration and fusion include,

[0253] Because images of the same object obtained by different sensors and different imaging modes may have relative translation, rotation, different scaling ratios, and distortion;

[0254] Therefore, strict geometric registration must be ensured before using optical RGB images and infrared images to perform mask fusion of the transmission line area;

[0255] The main limitation of image registration in parallel-axis optical RGB-infrared binocular imaging devices is that image registration is only possible at a specific target distance D. target Effective, registration error σ x The mathematical expression (in pixels) is as follows:

[0256]

[0257] Where f is the focal length, I pix d is the pixel size. c D is the baseline length. optimal For the target distance, i.e., the image alignment error is 0, if the optical axes are parallel, i.e., D optimal =∞, only when the target object is very far away from the camera will it be accurately aligned; for objects closer to the camera (D... target Registration errors will occur when the value is <∞.

[0258] The calibration method used is the one described above. It is important to note the image scale factor and the offset between the infrared and visible light images.

[0259] S4.4: Calculation of standard icing thickness for transmission lines includes,

[0260] Due to the diversity of icing types on transmission lines and the differences in shooting angles, this project will combine the identified icing type results with the equivalent thickness of icing on transmission lines to calculate the standard thickness of icing on transmission lines, so as to provide an important reference for subsequent transmission line maintenance decisions.

[0261] According to the "Specification for Icing Survey of Overhead Transmission Lines" (DL / T 5509—2015), the formula for calculating the standard icing thickness B0 of transmission lines is as follows:

[0262]

[0263] Where ρ is the estimated icing density, r is the original transmission line radius in mm, and K s R is the icing shape coefficient, and R is the icing radius in mm.

[0264] In particular, the icing shape coefficient K s The determination should be based on the local measured icing data. For areas without measured data, the selection can be made according to the provisions of the table below, taking into account the results of the identified icing types.

[0265] It should be noted that when small icing occurs, the shape factor K... s A lower value is preferable; when large-scale icing occurs, the shape factor K should be used. s A higher value should be selected.

[0266]

[0267] Example 2

[0268] As an embodiment of the present invention, a deep learning-based transmission line icing thickness detection system is provided, comprising: a data acquisition module, an image processing module, an edge detection module, a camera calibration module, and an image fusion module;

[0269] The data acquisition module is used for image acquisition and image preprocessing;

[0270] The image processing module is used for image segmentation, instance segmentation, and feature extraction;

[0271] The edge detection module is used to perform edge detection, and uses Otsu's method for binarization, nonlinear operators for filtering and noise reduction, and extracts boundaries;

[0272] The camera calibration module is used to obtain and calibrate the camera's intrinsic and extrinsic parameters;

[0273] The image fusion module is used to perform image registration and fusion on the transmission line images after camera calibration, and uses a trained neural network to output the calculation results of the ice thickness.

[0274] Specific limitations regarding the deep learning-based transmission line icing thickness detection system can be found in the limitations of the deep learning-based transmission line icing thickness detection method described above, and will not be repeated here. Each module in the aforementioned deep learning-based transmission line icing thickness detection system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0275] Example 3

[0276] Reference Figure 3 This is the third embodiment of the present invention. Based on the first two embodiments, this embodiment provides a computer device, which can be a server, and its internal structure diagram can be as follows. Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media.

[0277] The computer device's database is used to store motion detection data. The computer device's network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in any of the above embodiments of the sparse tensor computation acceleration method.

[0278] Those skilled in the art will understand that Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0279] In one embodiment, the present invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in any of the above embodiments of the sparse tensor operation acceleration method.

[0280] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0281] Example 4

[0282] As an embodiment of the present invention, a method for detecting the icing thickness of transmission lines based on deep learning is provided. To verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0283] To verify the effectiveness of the deep learning-based method for detecting icing thickness on transmission lines proposed in this invention, researchers, with the cooperation of a provincial power company, selected several typical icing-prone areas of transmission lines within the province for field tests. The testing period spanned from November 2023 to February 2024, covering a complete winter icing cycle.

[0284] First, the research team used a high-precision dual-light optical system (model XL-2000, resolution 0.1mm) to acquire images at 10 selected test points. 100 icing images were acquired at each test point, for a total of 1000 raw images. These images cover the transmission line status under different weather conditions, at different times, and with varying degrees of icing.

[0285] Subsequently, the acquired raw images were preprocessed. The preprocessing steps included: (1) adjusting all images to a uniform size of 1920×1080 pixels; (2) adjusting brightness using histogram equalization; (3) enhancing image color saturation using an adaptive color enhancement algorithm; (4) enhancing image contrast using contrast-limited adaptive histogram equalization (CLAHE) technology; and (5) applying a bilateral filtering algorithm to denoise the images, so as to preserve edge information while reducing noise.

[0286] Next, the researchers used an improved YOLACT++ network architecture for image segmentation. This network, based on the ResNet101 backbone, adds an attention mechanism and depthwise separable convolutions to improve feature extraction capabilities and computational efficiency. The network training employed a transfer learning strategy, first pre-training on the COCO dataset and then fine-tuning using 800 manually annotated images of icing on power transmission lines. During training, the Adam optimizer was used with an initial learning rate of 0.001, decaying by a factor of 10 every 50 epochs, for a total of 200 epochs.

[0287] To improve the accuracy of edge detection, the research team designed a novel multi-directional Sobel operator template. In addition to the traditional 0° and 90° directions, templates for 45° and 135° directions were added to better capture oblique edge information. After edge detection, an improved Otsu method was used for adaptive binarization. This method dynamically adjusts the weights of inter-class and intra-class variances, allowing the binarization threshold to better adapt to images under different lighting conditions.

[0288] In the area of ​​boundary tracing, researchers have developed a topology-based boundary tracing algorithm. This algorithm first determines a starting point, then traces the boundary point by point according to a pre-defined boundary discrimination criterion until the search termination criterion is met. This method can effectively handle the problem of transmission line boundary extraction in complex backgrounds.

[0289] Finally, a meticulously designed dual-light camera calibration process ensured accurate registration of visible light and infrared images. The calibration process used a 15×11 checkerboard calibration board, acquiring 50 sets of calibration images at different angles and distances. Camera intrinsic parameters were calculated using Zhang's method, and extrinsic parameters were estimated using vanishing point detection technology. Image registration employed a feature-point-based method, using the SIFT algorithm to extract feature points, the RANSAC algorithm to remove mismatched points, and finally, affine transformation to achieve image alignment.

[0290] After completing the above preparations, the research team used a trained deep learning model to infer the ice thickness of the transmission lines from the test set images. Meanwhile, to verify the accuracy of the algorithm, the researchers also used traditional manual measurement methods as a control group.

[0291] Table 1 Performance Evaluation Table

[0292]

[0293] A thorough analysis of the data in Table 1 leads to the following key conclusions:

[0294] The deep learning-based method for detecting icing thickness on transmission lines proposed in this invention demonstrates high accuracy at various test points. As shown in the table, the relative error of this method ranges from 0.63% to 1.92%, with an average relative error of 1.19%.

[0295] In comparison, the relative error of traditional manual measurement methods ranges from 2.34% to 5.77%, with an average relative error of 3.99%. This indicates that the accuracy of this method is significantly improved compared to traditional methods, with an average improvement of 2.8 percentage points.

[0296] The test data shows that the method maintains stable high accuracy under different icing thickness conditions. Whether under thinner icing conditions (such as 5.2 mm at test point 1) or thicker icing conditions (such as 25.6 mm at test point 6), the relative error of the method is controlled within 2%, which indicates that the method has good adaptability and can cope with various icing conditions.

[0297] The detection time data in the table shows that the detection speed of this method is very fast, with an average detection time of only 1.12 seconds. Considering that traditional manual measurement methods usually take several minutes or even longer, the efficiency advantage of this method is obvious. This high efficiency makes large-scale, real-time monitoring of icing on transmission lines possible.

[0298] By observing the relative error data of this method, it can be found that its fluctuation range is smaller (1.29 percentage points), while the relative error fluctuation of the traditional method is larger (3.43 percentage points). This indicates that this method has better consistency and stability and is less susceptible to human or environmental factors.

[0299] The data shows that the relative error of this method decreases slightly with increasing ice thickness. This may be because the features in the image are more obvious under thicker ice conditions, which is beneficial for the deep learning model to make more accurate identification and measurement.

[0300] Considering the actual application scenarios of transmission line icing detection, the accuracy of this method fully meets the engineering requirements. According to relevant State Grid standards, the allowable error for measuring the thickness of transmission line icing is usually ±5mm. The maximum absolute error of this method is 0.3mm (test point 3), which is far below this standard, proving its reliability in practical applications.

[0301] This method, by combining deep learning technology, multi-directional Sobel operators, and improved Otsu's method, successfully overcomes the limitations of traditional methods under complex backgrounds and different lighting conditions. In particular, the introduction of multi-directional Sobel operators effectively improves the accuracy of edge detection, laying a good foundation for subsequent ice thickness calculation.

[0302] Considering the serious economic losses that may result from icing on transmission lines, the high precision and efficiency of this method are expected to significantly reduce the risks of power grid operation. If it is promoted and applied nationwide, the annual savings in maintenance costs and the avoidance of accident losses could be a considerable figure.

[0303] In summary, the deep learning-based method for detecting icing thickness on transmission lines proposed in this invention demonstrates significant advantages in terms of accuracy, efficiency, adaptability, and consistency. It not only overcomes the limitations of traditional methods but also provides a reliable technical solution for intelligent monitoring of power systems. The application of this method will help improve the safe operation of the power grid and reduce economic losses caused by icing, which is of great practical significance for the stable operation and sustainable development of my country's power system.

[0304] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A power line icing thickness detection method based on deep learning, characterized in that, The application relates to a method for detecting ice thickness on a power transmission line. The method comprises the following steps: collecting ice image of the power transmission line by using a standard double-light optical device, and performing image preprocessing on the ice image to generate an original data set; The method comprises the following steps: performing image segmentation on the original data set, and training a deep neural network model by using a lightweight YOLACT network, and evaluating the trained deep neural network model; The method comprises the following steps: performing edge detection by designing a multi-directional Sobel operator template, performing binarization processing by using the Otsu method, and performing filter denoising by using a nonlinear operator, extracting a boundary chain code array based on a topological structure boundary tracking algorithm, and determining a mask of the original power transmission line; The method comprises the following steps: collecting the power transmission line image by using a double-light camera and performing camera calibration, performing image registration and fusion on the power transmission line image after the camera calibration, and outputting a calculation result of the ice thickness on the power transmission line by using the trained neural network. 2.The power line icing thickness detection method based on deep learning according to claim 1, wherein: The image preprocessing comprises size transformation, brightness transformation, color enhancement, contrast enhancement and image denoising; The image segmentation comprises feature extraction, mask generation, target detection and instance segmentation; The method realizes high-precision instance segmentation of the ice image on the power transmission line by using a single-stage instance segmentation algorithm and a polar coordinate representation coding mask method. The single-stage instance segmentation algorithm comprises the following steps: Feature extraction is performed by using a backbone network, and the extracted features are divided into two branches, i.e. a classification branch and a regression branch, for processing; The classification branch comprises category classification and polar coordinate center degree prediction, and is used for judging which category a pixel belongs to and the reliability of the distance of the pixel to the center; One point in the feature map is taken as a polar coordinate center point, N rays are generated, and the ray distance of each target edge to the center point is calculated; The ray distance is used for reconstructing the target edge, and the best mask shape is determined; 1000 best samples are reserved in each FPN layer, and sorting and filtering are performed based on the product of the classification confidence score and the polar coordinate center degree score; According to the known ray distance, the coordinates of the edge points are connected, and the outline of the target is formed by sequentially connecting from 0 degrees; After inputting the remote sensing image, image features are extracted in the feature backbone network; The image features are input into two branches, a prototype mask generation network branch generates a group of prototype masks with the same size as the original image through a full convolution network; The mask coefficient branch is used for generating mask coefficients of each instance, and the category and boundary box position of the candidate box are predicted; The prototype mask and the corresponding mask coefficient are linearly combined to generate an instance mask; The instance segmentation result is obtained by cutting the mask according to the predicted boundary box. 3.The power line icing thickness detection method based on deep learning according to claim 2, wherein: The steps of training the deep neural network model comprise the following steps: The input image is uniformly preprocessed and uniformly adjusted in size; Multi-scale features are obtained by passing the data through a ResNet backbone network and an FPN layer, and the multi-scale features are input into parallel detection modules for coefficient combination prediction; The p3 feature map output by the feature pyramid is input into a prototype network to generate a pre-segmentation model, and the prototype mask of the image is predicted; The fast non-maximum suppression method is used to remove interference detection boxes, and the prototype mask is used to generate a prediction box and a mask; The loss functions of the detection box, the classification and the mask are calculated respectively, and the weights are updated in a gradient backpropagation manner. Repeat the above steps until the model meets the stability condition; The evaluation of the trained deep neural network model comprises, The deep neural network model is evaluated by using a training evaluation index; The training evaluation index comprises, Each PNG image label corresponds to a mask of the target region of the original image, and the PNG image label is a binary image; The white part represents the foreground of the corresponding source image, and the black part represents the background; The XML file labeled by LabelImg is converted into a PNG image label by a self-programmed program; The target detection algorithm and the instance segmentation algorithm use the mAP value as the training evaluation index; The calculation of the mAP value is as shown in the following formula: where C is the number of classes of the target, AP i is the average precision of the i-th class. 4.The power line icing thickness detection method based on deep learning according to claim 3, wherein: The design of the multi-directional Sobel operator template comprises, Two templates of different directions, 0°, 90°, 45°, and 135°, are added to the traditional Sobel operator; Edge information is extracted through the four-directional template, and the four-directional template matrix is as shown in the following formula: The four-directional Sobel templates are applied to the input infrared image for convolution operation; The weighted difference of the neighborhood of each pixel point on the image is calculated, the extreme value of the weighted difference is determined, and the extreme value is set as a threshold value; When the weighted difference of the pixel point is greater than the threshold value, the pixel point is retained as an edge point; When the weighted difference of the pixel point is less than the threshold value, the pixel point is removed, and the target edge is determined; The binarization processing by the Otsu method comprises, The infrared image is grayed, and the gray image is divided into L gray levels, and the probability of each gray level is defined, and the calculation formula is as follows: where n i is the number of pixels with gray level i, N is the total number of pixels in the image, p i is the probability of gray level i. The gray image is divided into two types according to the gray value, when 1≤gray value≤k, the gray image is divided into a type of gray image, denoted as C0; When k+1≤gray value≤L, the gray image is divided into two types of gray images, denoted as C1; The calculation formula of the probability of each type is as follows: Wherein, ω0 is the probability of the type of gray image C0, ω1 is the probability of the appearance of the type of gray image C1, and ω(k) is the cumulative probability of the gray value from the gray level 1 to k; The calculation formula of the class mean is as follows: wherein μ0 is the mean value of the first class of gray scale images C0, μ1 is the mean value of the second class of gray scale images C1, μ(k) is the cumulative mean value, μ T is the total mean value of the images; The calculation formula of the class variance is as follows: wherein is the variance of class C0, is the variance of class C1; Three evaluation indexes for judging the quality of binarization are defined, and the calculation formula is as follows: wherein, is the within-class variance, is the between-class variance, is the population variance, λ is the ratio of the between-class variance to the within-class variance, κ is the ratio of the population variance to the within-class variance, and η is the ratio of the between-class variance to the population variance; The best binarization threshold value is found by traversing the k value, and the calculation formula is as follows: wherein is the optimal binarization threshold value, k is the number of iterations. The best binarization processing of the infrared image is realized. 5.The power line icing thickness detection method based on deep learning according to claim 4, wherein: The boundary tracking algorithm based on the topological structure comprises the determination of the starting point, the boundary discrimination criterion, and the search termination criterion; The determination of the MASK of the original power transmission line comprises, The boundary of the power transmission line in the infrared image can be extracted in the form of a chain code array through the boundary tracking algorithm, and the edge data of the power transmission line are obtained; Through the obtained edge data, the correct boundary of the power transmission line region, i.e., the boundary chain code array, is screened out; The minimum circumscribed rectangle of the boundary chain code array is solved, and the best mask of the original power transmission line is obtained. 6.The power line icing thickness detection method based on deep learning according to claim 5, wherein: The camera calibration comprises the camera intrinsic parameters and the camera extrinsic parameters; The camera intrinsic parameters comprise the focal length, the principal point coordinates, the pixel size, the skew coefficient, and the distortion; The steps for solving the camera intrinsic parameters comprise, The coordinates of the coordinate points are obtained by measurement or software generation. A plurality of pictures are taken by aiming the camera at the calibration board, the pictures including the plane of the entire calibration board, the camera is fixed on a tripod, and 30 pictures of the chessboard calibration board are taken at different distances and angles; Feature points on the calibration board are extracted from the taken pictures, and the feature points are obtained through a feature extraction algorithm; The camera internal parameters are calculated by minimizing the re-projection error and using an optimization algorithm; The steps of calculating the camera external parameters based on the vanishing point include, The distance between the camera and the transmission line is obtained by a laser range finder, and the specific positions of the pitch angle and the yaw angle are obtained; When the three-dimensional space image is projected onto a two-dimensional plane, the image is perspective distorted, that is, two originally parallel lines in the three-dimensional space intersect at a point in the two-dimensional plane, which is the vanishing point; When the two parallel lines in the three-dimensional space are parallel to the image plane, the vanishing point is located at the infinite distance of the image plane, otherwise, the vanishing point is located at the finite distance of the image plane, that is, when there is a yaw angle, the horizontal coordinate of the vanishing point will be offset; When there is a pitch angle, the vertical coordinate of the vanishing point will be offset; If the dashed rectangular area is a standard imaging plane, that is, the camera does not have a pitch angle and a yaw angle, the vanishing point pixel coordinate is V(u0, v0); If the black rectangular area is the imaging plane when there is a pitch angle and a yaw angle, the vanishing point coordinate is V(u1, v1); When the camera has a yaw angle, according to the geometric relationship, the calculation formula of the yaw angle is as follows: Wherein, m is the size of the imaging plane, f is the focal length of the camera, γ is the yaw angle of the camera, and l is the distance between the vertical coordinate of the vanishing point and the vertical coordinate of the center point of the image plane. When the camera has a pitch angle, according to the geometric relationship, the calculation formula of the pitch angle is as follows: Wherein, θ is the pitch angle of the camera, and n is the size of the imaging plane. 7.The power line icing thickness detection method based on deep learning according to claim 6, wherein: The image registration and fusion includes, The optical RGB image and the infrared image are used for mask registration and fusion of the transmission line region; Registration error σ x The formula for calculating σ is as follows: where f is the focal length, I pix is the pixel size, d c is the baseline length, D optimal is the target distance, D target is the specific target distance; The transmission line icing standard thickness calculation includes, According to the transmission line icing standard thickness calculation formula, the calculation is carried out; The transmission line icing standard thickness B0 calculation formula is as follows: where p is the estimated density of ice, r is the original transmission line radius, K s is the ice shape coefficient, and R is the ice radius.

8. A system employing the deep learning-based ice thickness detection method for a power transmission line according to any one of claims 1 to 7, characterized by It includes: A data acquisition module, an image processing module, an edge detection module, a camera calibration module, and an image fusion module; The data acquisition module is used for image acquisition and image preprocessing; The image processing module is used for image segmentation, instance segmentation, and feature extraction; The edge detection module is used for edge detection, binarization processing by Otsu method, filtering and denoising by using a nonlinear operator, and boundary extraction; The camera calibration module is used for obtaining camera internal and external parameter calibration; The image fusion module is used for image registration and fusion of the transmission line image after camera calibration, and the calculation result of the icing thickness output by the trained neural network is used. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the deep learning-based transmission line icing thickness detection method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the deep learning-based transmission line icing thickness detection method in any one of claims 1 to 7.

Images