A cross-check based detection method for overhead power lines

Abstract

A cross-check based power overhead wire detection method, which inputs a power overhead wire image, uses DeepLabv3+ segmentation network to detect the power overhead wire image, performs binarization operation to obtain a binary image, uses MCMLSD line detection algorithm to detect the power overhead wire, outputs the final detected straight line segment, and obtains the coordinates of the starting point and end point of the straight line segment; discretizes the coordinates of the starting point and end point of the straight line segment, determines the coordinate values ​​of the starting point and end point of the straight line segment, and detects the power overhead wire image in a binarized image. I The corresponding straight line segments are drawn and marked; the marked image is output to obtain the final complete overhead power line detection image. This invention uses a cross-verification method to achieve complete extraction of overhead power lines, avoiding the problems of false detection, missed detection, and breakage of power lines in traditional methods, and provides technical support for the automatic and accurate detection of overhead power lines by drones.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a cross-check-based power overhead wire detection method. Background Art

[0002] Overhead conductors are a key component of power transmission systems. Ensuring their proper operation requires extensive human and material resources for power line inspections. Manual inspections are time-consuming and labor-intensive, and quality assurance is difficult for long-distance, high-voltage transmission lines. The rapid development of drone technology in recent years has enabled drone inspections to gradually replace manual inspections, playing a particularly important role in the inspection of long-distance, high-voltage transmission systems. In this field, accurately extracting overhead conductors from drone-captured images or videos is crucial and a prerequisite for automated, high-precision drone inspections.

[0003] Currently, traditional methods for detecting overhead power lines fall into two main categories: one is extraction methods based on edge detection operators, which primarily use edge detection algorithms such as Sobel, Canny, and Gabor to process the features of overhead power lines and extract them from the background. The other is based on the direction and shape characteristics of object edges in the image. Taking advantage of the fact that power lines have straight or line-like characteristics, line detection algorithms such as LSD, CannyLines, and Hough transform are used to extract power lines or fit power line segments. In recent years, deep learning technology has also been widely used in the power industry. Semantic segmentation networks, such as DeepLabv3+, can achieve pixel-level classification of overhead power lines. Compared to traditional image processing methods, they are less affected by the linear morphological features in the image and have better scene robustness.

[0004] The shortcomings of the existing technical solutions are:

[0005] 1. Traditional power conductor detection methods based on image processing suffer from poor real-time performance, low efficiency, and inability to adapt to a variety of complex inspection scenarios (such as forests, fields, and cities). For example, the principle of edge detection algorithms is to use edge detection operators to extract shallow target features for judgment, thereby achieving conductor detection. When the conductor target features in the inspection image are not significant or the scene is complex, the shallow target features are easily interfered with, resulting in suboptimal detection of overhead power conductors. Similarly, line detection algorithms represented by Hough transforms are easily interfered with by the straight line information of non-overhead conductors in the image, resulting in a large number of false detections and increasing the difficulty of conductor detection.

[0006] 2. Although deep semantic segmentation networks represented by DeepLabv3+ are robust to complex scenes and are not easily affected by factors such as the quality of the acquired image and the scene, their wire segmentation results may show discontinuities and breaks in the segmented areas, resulting in the incompleteness of the detected wires and affecting the subsequent positioning of the overhead wires by drones. Summary of the Invention

[0007] The present invention is applied to the automatic detection of overhead power conductors in complex backgrounds during drone inspections. The cross-verification method is used to achieve complete extraction of overhead power conductors, avoiding problems such as false detection, missed detection, and breakage of power lines in traditional methods, and providing technical support for the automatic and accurate detection of overhead power conductors by drones.

[0008] A cross-check based power overhead wire detection method comprises the following steps:

[0009] Step S1: Input a 512×512 image of overhead power lines, marked as I;

[0010] Step S2: Use the DeepLabv3+ segmentation network to detect the power overhead wire image I to obtain the detected overhead wire image I1;

[0011] Step S3: Binarize the detected overhead wire image I1 to obtain a binary image I2. The set of image pixels with a value of 1 in I2 is marked as P. 0 , this set is the straight line identification area;

[0012] Step S4: Detect the overhead power line using the MCMLSD line detection algorithm on the overhead power line image I, output the final detected line segment, and obtain the coordinates of the starting point and the ending point of the line segment;

[0013] Step S5: Discretize the coordinates of the starting point and the ending point of the straight line segment. Through discretization, the coordinate information of the straight line segment is retained, and the obtained straight line segment set is marked as P i , P i Represents the discretized coordinate information of the i-th line segment, i = 1, 2, 3...n;

[0014] Step S6: Let i = 1, and solve the set of straight line segments P 1 With P 0 The intersection of i=1;

[0015] Step S7: P i Each line segment in the process is operated according to step S6 to obtain the line segment set P i With P 0 The set of intersections of i=1,2,3...n;

[0016] Step S8: Calculate the set In each subset with P 0 The number of intersections, get the set

[0017] i=1,2,3...n;

[0018] Step S9: Removal The subset with zero in is sorted in descending order, and the set is marked as

[0019] Step S10: Select The first 75% of the subset is indexed by the superscript index line segment number of i in the subset, and the P in step S5 i Determine the coordinate values ​​of the starting and ending points of the straight line segment, draw the corresponding straight line segment on the power overhead wire image I, and mark it;

[0020] Step S11: Output the marked image to obtain the final complete power overhead wire detection image I0.

[0021] Preferably, in step S2 of the present invention, the power overhead wire image I is subjected to power overhead wire detection using a DeepLabv3+ segmentation network to obtain a detected overhead wire image I1; the specific process is as follows:

[0022] The DeepLabv3+ segmentation network is an encoder-decoder structure consisting of an encoding region and a decoding region. In the encoding region, a deep convolutional neural network is first used to extract features from the original input image to obtain two different feature information. The spatial pyramid pooling module is then used to perform multi-scale dilated convolution sampling on the higher-level feature information extracted by the deep convolutional neural network. Several feature layers obtained by dilated convolution with different expansion rates are stacked together and combined to perform 1×1 convolution to form a high-level semantic feature layer.

[0023] In the decoding area, the lower-level feature information extracted from the deep convolutional neural network is subjected to a 1×1 convolution to form a low-level semantic feature layer, which is then fused with the high-level semantic feature layer after 4x upsampling to supplement the image position information. The fused complete semantic feature layer undergoes feature extraction through a 3×3 convolution, and then the output image is resized to the same size as the input image using 4x upsampling to obtain the detected overhead wire image I1.

[0024] Preferably, in step S3 of the present invention, a binarization operation is performed on the detected overhead wire image I1 to obtain a binarized image I2; the specific process is:

[0025] Image binarization calculates the global threshold by using the grayscale average value based on the histogram, iterative optimal threshold and OTSU algorithm, and superimposes the global threshold with the background distribution to form a local threshold. Binarization is performed based on the obtained global and local thresholds to determine the pixel value of each pixel in the image. The values ​​1 and 0 correspond to pixels identified by the network as overhead wires and non-overhead wires, respectively, and finally a binary image I2 is obtained.

[0026] Preferably, in step S4 of the present invention, the MCMLSD line detection algorithm is used to detect the power overhead wire image I, output the final detected straight line segment, and obtain the coordinates of the starting point and the ending point of the straight line segment;

[0027] First, the global probabilistic Hough transform is used to detect the set of straight line segments in the power overhead wire image I. All the detected straight line segments in the image are analyzed, and the position of the line segments corresponding to the peak in the Hough domain is determined to obtain a set of candidate line segments. The edge points on each candidate line segment are modeled using a linear Markov chain model, and the corresponding probabilistic optimal labels are obtained. According to the probabilistic optimal labels, all candidate line segments are ranked by significance. Finally, the final detected straight line segments are output according to the sorting results, and the coordinates of the starting and ending points of the straight line segments are obtained.

[0028] Preferably, in step S5 of the present invention, the coordinates of the starting point and the ending point of the straight line segment are discretized; the specific process is:

[0029] The coordinates of the starting point of the marked straight line segment are (x i 1 ,y i 1 ), the coordinates of the end point of the straight line segment are Calculate the number of pixel intervals between two points as M; in the x direction and The row index coordinates with intervals of 1 are labeled x k Similarly, we get the column index coordinate set in the y direction, marked as y k , k=1,2,3...M.

[0030] Overhead conductor detection is a crucial component of power line inspections. Traditional overhead conductor detection methods can effectively identify the morphological features of overhead conductors, but their robustness is limited and they struggle to eliminate interference from the surrounding environment. Overhead conductor detection based on deep semantic segmentation is a recently emerging detection method with good robustness. However, in practice, such methods suffer from discontinuous segmentation in their detection results. To address these issues, this paper combines a traditional line detection algorithm with a deep learning network, employing the MSMLSD algorithm (Markov Chain Marginal Line Segment Detector) and a DeepLabv3+ segmentation network to detect overhead conductors. This paper proposes a cross-validation-based overhead conductor detection method that innovatively combines the overhead conductor coordinate information extracted by the traditional line detection algorithm with the overhead conductor image information obtained by the deep learning network. By unifying and comparing these different data sets, the method successfully extracts the complete overhead conductor and mitigates the influence of the surrounding line segment structure on the detection results. Experiments demonstrate that this method exhibits excellent robustness in scenarios such as sky, land, and forests, providing a new approach for automated overhead conductor inspection using drones and advancing technological development in this field. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the implementation framework of the present invention.

[0032] Figure 2 It is the DeepLabv3+ network structure model.

[0033] Figure 3 It is a flow chart of the detection method of the present invention.

[0034] Figure 4 It is a comparison diagram of the effects of using various existing detection methods and the detection method of the present invention.

[0035] Figure 5 This is a comparison of images of overhead power lines obtained using different methods under complex backgrounds. DETAILED DESCRIPTION

[0036] The implementation framework of the power overhead wire detection of the present invention is as follows Figure 1 As shown in the figure, the original image of overhead power lines captured by a drone is input. The DeepLabv3+ segmentation network is used to perform semantic segmentation on the original image. The resulting semantic segmentation image is binarized to produce a binary power line segmentation map. The MCMLSD line detection algorithm is then used to extract power lines from the original image, producing a power line detection map. Finally, a cross-validation method is used to compare and fuse the binary power line segmentation map with the power line detection map, ultimately detecting the complete overhead power line area.

[0037] The specific implementation process of the present invention is as follows Figure 3 The steps are described as follows:

[0038] Step S1: Input a 512×512 image of overhead power lines, marked as I.

[0039] Step S2: Use the DeepLabv3+ segmentation network to detect the power overhead wire image I to obtain the detected overhead wire image I1.

[0040] The structure of the DeepLabv3+ segmentation network model is as follows Figure 2 As shown in the figure, this network model is an encoder-decoder structure, consisting of two main parts: the encoder and the decoder. In the encoder, a deep convolutional neural network (DCNN) is first used to extract features from the original input image, obtaining two different types of feature information. The Atrous Spatial Pyramid Pooling (ASPP) module then performs multi-scale dilated convolution sampling on the higher-level feature information extracted by the DCNN. Several feature layers obtained through dilated convolutions with different dilation rates are stacked together and combined for a 1×1 convolution to form a high-level semantic feature layer.

[0041] In the decoder, the lower-level features extracted by the DCNN undergo a 1×1 convolution to form a low-level semantic feature layer. This is then fused with the higher-level semantic feature layer obtained through 4x upsampling to supplement the image's position information. The fused semantic feature layer undergoes feature extraction using a 3×3 convolution, and the output image is resized to the same size as the input image using 4x upsampling, resulting in the detected overhead wire image I1.

[0042] Step S3: Binarize the detected overhead wire image I1. Image binarization calculates the global threshold through the traditional histogram-based grayscale average, iterative optimal threshold and OTSU algorithm, and superimposes the global threshold with the background distribution to form a local threshold. Binarization is performed based on the obtained global threshold and local threshold to determine the pixel value of each pixel in the image. The values ​​1 and 0 correspond to the pixels identified as overhead wires and non-overhead wires by the network, respectively. Finally, a binary image I2 is obtained. The set of image pixel values ​​​​in I2 with a value of 1 is marked as P 0 , this set is the straight line recognition area.

[0043] Step S4: Detect overhead power lines using the MCMLSD line detection algorithm on the power line image I. First, a global probabilistic Hough transform is used to detect a set of straight line segments in the image. All detected straight line segments in the image are analyzed, and the line segment positions corresponding to peaks in the Hough domain are determined. This yields a set of candidate line segments. A linear Markov chain model is then used to model the edge points on each candidate line segment, obtaining the corresponding probabilistic optimal labels. Based on these probabilistic optimal labels, all candidate line segments are ranked by significance. Finally, the final detected line segments are output based on the ranking results, along with the coordinates of the line segment's starting and ending points.

[0044] Step S5: Discretize the coordinates of the starting point and the ending point of the straight line segment. Mark the coordinates of the starting point of the straight line segment as (x i 1 ,y i 1 ), the coordinates of the end point of the straight line segment are Calculate the number of pixel intervals between two points as M; in the x direction and The row index coordinates with intervals of 1 are labeled x k Similarly, we get the column index coordinate set in the y direction, marked as y k , k=1,2,3...M. Through discretization, the coordinate information of the straight line segment is retained, and the obtained straight line segment set is marked as P i , P i Represents the discretized coordinate information of the i-th line segment, i = 1, 2, 3...n.

[0045] Step S6: Let i = 1, and solve the set of straight line segments P 1 With P 0 The intersection of i=1;

[0046] Step S7: P i Each line segment in the process is operated according to step S6 to obtain the line segment set P i With P 0 The set of intersections of i=1,2,3...n;

[0047] Step S8: Calculate the set In each subset with P 0 The number of intersections, get the set

[0048] i=1,2,3...n;

[0049] Step S9: Removal The subset with zero in is sorted in descending order, and the set is marked as

[0050] Step S10: Select The first 75% of the subset is indexed by the superscript index line segment number of i in the subset, and the P in step S5 i Determine the coordinate values ​​of the starting and ending points of the straight line segment, draw the corresponding straight line segment on the power overhead wire image I, and mark it;

[0051] Step S11: Output the marked image to obtain the final complete power overhead wire detection image I0.

[0052] 1. Comparison of the effectiveness of the present invention and different power line detection algorithms

[0053] like Figure 4 The detection results of five different line detection algorithms are shown in the figure. The overhead wires are detected using traditional classic line detection algorithms Hough line detection, Cannylines line detection and MCMLSD line detection. The detection results show that the MCMLSD line detection algorithm can directly detect the entire overhead wire target, while the other two line detection algorithms cannot detect the complete target. The MCMLSD algorithm has an advantage in detecting overhead wires and retaining the complete wire target. The results after detection are also convenient for the implementation of subsequent algorithms. Although the classic line detection algorithm can detect overhead wire targets, there is a lot of interference in the detection process, and the wire target cannot be accurately detected, and the detection accuracy is low. Figure 4 As shown in Figure 2, the DeepLabv3+ semantic segmentation method alone can detect overhead wire targets while reducing interference targets, but the detection result of this method is a discontinuous straight line, which cannot fully reflect the detection target. This paper combines the MCMLSD line detection algorithm with the DeepLabv3+ segmentation network and performs cross-validation. The effect is shown in Figure 2. Figure 4 As shown, not only can all the interference in the background be removed, but the power overhead wires can also be completely extracted, which can well complete the detection and extraction of the power overhead wires.

[0054] 2. Application of the present invention in complex inspection scenarios

[0055] To further verify the effectiveness of the present invention, we added multiple complex environments to the single backgrounds of sky, land and forest to verify the detection ability of the method under complex backgrounds. We used the MCMLSD line detection algorithm, DeepLabv3+ semantic segmentation detection algorithm and the present invention for comparison. The experimental results are shown in Figure 2. Figure 5 Experiments have shown that the present invention can eliminate interference and fully detect and extract overhead power lines in complex backgrounds such as urban roads, factories, woods, grasslands, and parks.

Claims

1. A method for detecting overhead power lines based on cross-checking, characterized in that The steps include: Step S1: Input a 512×512 image of overhead power lines, marked as I; Step S2: Use the DeepLabv3+ segmentation network to detect the power overhead wire image I to obtain the detected overhead wire image I1; Step S3: Binarize the detected overhead wire image I1 to obtain a binary image I2. The set of image pixels with a value of 1 in I2 is marked as P. 0 , this set is the straight line identification area; Step S4: Detect the overhead power line using the MCMLSD line detection algorithm on the overhead power line image I, output the final detected line segment, and obtain the coordinates of the starting point and the ending point of the line segment; Step S5: Discretize the coordinates of the starting point and the ending point of the straight line segment. Through discretization, the coordinate information of the straight line segment is retained, and the obtained straight line segment set is marked as P i , P i Represents the discretized coordinate information of the i-th line segment, i = 1, 2, 3...n; Step S6: Let i = 1, and solve the set of straight line segments P 1 With P 0 The intersection of i=1; Step S7: P i Each line segment in the process is operated according to step S6 to obtain the line segment set P i With P 0 The set of intersections of Step S8: Calculate the set In each subset with P 0 The number of intersections, get the set Step S9: Removal The subset with zero in is sorted in descending order, and the set is marked as Step S10: Select The first 75% of the subset is indexed by the superscript index line segment number of i in the subset, and the P in step S5 i Determine the coordinate values ​​of the starting and ending points of the straight line segment, draw the corresponding straight line segment on the power overhead wire image I, and mark it; Step S11: Output the marked image to obtain the final complete power overhead wire detection image I0.

2. The cross-check based power overhead wire detection method according to claim 1, characterized in that In the above step S2, the power overhead wire image I is detected using the DeepLabv3+ segmentation network to obtain the detected overhead wire image I1; the specific process is: The DeepLabv3+ segmentation network is an encoder-decoder structure consisting of an encoding region and a decoding region. In the encoding region, a deep convolutional neural network is first used to extract features from the original input image to obtain two different feature information. The spatial pyramid pooling module is then used to perform multi-scale dilated convolution sampling on the higher-level feature information extracted by the deep convolutional neural network. Several feature layers obtained by dilated convolution with different expansion rates are stacked together and combined to perform 1×1 convolution to form a high-level semantic feature layer. In the decoding area, the lower-level feature information extracted from the deep convolutional neural network is subjected to a 1×1 convolution to form a low-level semantic feature layer, which is then fused with the high-level semantic feature layer after 4x upsampling to supplement the image position information. The fused complete semantic feature layer undergoes feature extraction through a 3×3 convolution, and then the output image is resized to the same size as the input image using 4x upsampling to obtain the detected overhead wire image I1.

3. The cross-check based power overhead wire detection method according to claim 2 is characterized in that In step S3, the detected overhead wire image I1 is binarized to obtain a binarized image I2. The specific process is as follows: Image binarization calculates the global threshold by using the grayscale average value based on the histogram, iterative optimal threshold and OTSU algorithm, and superimposes the global threshold with the background distribution to form a local threshold. Binarization is performed based on the obtained global and local thresholds to determine the pixel value of each pixel in the image. The values ​​1 and 0 correspond to pixels identified by the network as overhead wires and non-overhead wires, respectively, and finally a binary image I2 is obtained.

4. The method for detecting overhead power lines based on cross-checking according to claim 3 is characterized in that In step S4, the MCMLSD line detection algorithm is used to detect the overhead power line image I, output the final detected straight line segment, and obtain the coordinates of the starting point and the ending point of the straight line segment; First, the global probabilistic Hough transform is used to detect the set of straight line segments in the power overhead wire image I. All the detected straight line segments in the image are analyzed, and the position of the line segments corresponding to the peak in the Hough domain is determined to obtain a set of candidate line segments. The edge points on each candidate line segment are modeled using a linear Markov chain model, and the corresponding probabilistic optimal labels are obtained. According to the probabilistic optimal labels, all candidate line segments are ranked by significance. Finally, the final detected straight line segments are output according to the sorting results, and the coordinates of the starting and ending points of the straight line segments are obtained.

5. The method for detecting overhead power lines based on cross-checking according to claim 4, characterized in that In step S5, the coordinates of the starting point and the ending point of the straight line segment are discretized; the specific process is as follows: The coordinates of the starting point of the marked straight line segment are (x i 1 ,y i 1 ), the coordinates of the end point of the straight line segment are (x i 2 ,y i 2 ), calculate the number of pixel intervals between two points as M; in the x direction and The row index coordinates with intervals of 1 are labeled x k Similarly, we get the column index coordinate set in the y direction, marked as y k , k=1,2,3…M.

Images