A method for locating drainage clamps for live working in distribution networks based on computer vision

Abstract

The present application relates to a method for locating drain wire clamps for live-line work on a distribution network based on computer vision, comprising the following steps: grouping collected images, constructing a DeepLabv3+ semantic segmentation convolutional neural network model, processing the collected images using image enhancement technology to divide them into a training set, a test set, and a validation set, training the DeepLabv3+ semantic segmentation convolutional neural network model using a five-fold cross-validation method, and using the trained model to locate and identify the center screw of the collected drain wire clamp image. The present invention utilizes a camera capable of remotely transmitting images included in a live-line working system to capture images of the drain wire clamp screw, and automatically segments the center screw of the drain wire clamp using the DeepLabv3+ semantic segmentation convolutional neural network model to achieve positioning of the clamp hole, thereby overcoming the defect of the prior art that it has a high dependence on lens pixels and light sources, and is suitable for outdoor work.

Description

Technical Field

[0001] The present application relates to the field of image recognition of power grid drain wire clamps, and specifically to a method for locating drain wire clamps for live distribution network operations based on computer vision. Background Art

[0002] Grid diversion operations involve connecting branches to existing transmission lines. Due to the high voltage and current of the main lines, a power outage can significantly impact residents' lives and business operations, causing numerous inconveniences. Therefore, most grid diversion operations are performed live. However, these operations, performed near high-voltage lines, pose numerous safety risks, including falls from heights, electric shocks, and other casualties.

[0003] To improve the efficiency and automation of live-line work, ensure the safety of workers, and reduce their workload, the development of a live-line distribution network working robot is crucial. However, in certain environments, live-line distribution network working robots still require human intervention. In particular, guiding the drain wire into the holes of a universal clamp requires manual feeding. Accurately determining the location of the clamp hole is crucial for unmanned insertion of the drain wire into the clamp.

[0004] During operation, the center of the hole where the drain wire and busbar are located is the locking screw, and the position of the hole and the screw are relatively fixed. If the position of the center screw can be determined, the position of the wire clamp hole can be indirectly determined. Therefore, image segmentation of the center screw has become a method for locating the wire clamp hole.

[0005] At present, there are two main methods for image segmentation of screws. The first is the detection of thread parameters. First, the NCC (normalized cross correlation) normalized matching algorithm is used to identify the threads on the screw and find the center position on the screw. Then, through ROI cropping design, a rectangular image is obtained. Finally, the thread features are extracted to obtain the parameters of the major diameter, minor diameter and pitch of the thread. The second is the identification of thread defects. First, the image must be professionally processed, such as pixel grayscale adjustment and denoising, to make the image as clear as possible. Then, the grayscale gradient matrix is ​​used for calculation, and various data of the external thread on the screw surface are collected for comparison with the standard data.

[0006] Both of the above methods collect data on the screw threads and then analyze their specific conditions. While thread detection technology effectively ensures accuracy, it places high demands on the hardware configuration of the inspection system, such as requiring a high-pixel industrial camera and a relatively expensive fixed-focus lens. The measurement environment also significantly impacts accuracy. In a relatively open environment, the light source does not illuminate the part surface perpendicularly, resulting in large shadows or overexposure, significantly reducing image quality. These conditions can lead to suboptimal thread image recognition results, hindering the segmentation of the center screw image and making it impossible to locate the clamp hole. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for locating drainage wire clamps for live distribution network operations based on computer vision, which uses a camera capable of remotely transmitting images contained in a live distribution network operation robot to capture images of the drainage wire clamp screw, and automatically segments the center screw of the drainage wire clamp through the DeepLabv3+ semantic segmentation convolutional neural network model to achieve the positioning of the wire clamp hole, overcoming the defect of the existing technology that it has a high dependence on lens pixels and light sources, and is suitable for outdoor operations.

[0008] The technical solution adopted by the present invention is: a method for locating a drainage clamp for live working on a distribution network based on computer vision, comprising the following steps:

[0009] S100: randomly shuffling the collected drainage clip image data and evenly dividing them into five groups;

[0010] S200: Construct a DeepLabv3+ semantic segmentation convolutional neural network model, which includes an encoding part and a decoding part; the encoding part includes a DCNN network, an ASPP network and a first convolutional layer; the DCNN network uses an improved Xception network as a base network, and the improved Xception network uses a dilated convolution with stride to replace the maximum pooling operation in the original Xception network, and the DCNN network is used to expand the number of feature channels; the ASPP network includes five branches, the first branch is convolution, the second to fourth branches are a combination of depthwise separable convolution and dilated convolution, and the fifth branch is global average pooling. The ASPP network is used to output five feature layers, and the first convolutional layer is used to cascade five feature layers; the decoding part includes a second convolutional layer, a third convolutional layer, a first 4x upsampling module and a second 4x upsampling module, and the feature layer with expanded feature channels and the five feature layers are cascaded after being processed by the second convolutional layer and the first 4x upsampling module, and then processed by the third convolutional layer and the second 4x upsampling module in sequence;

[0011] S300: The DeepLabv3+ semantic segmentation convolutional neural network model is trained using a five-fold cross-validation method, using one set of drainage clip image data as a test set, and expanding the remaining four sets of drainage clip image data using image enhancement technology. 10% of the expanded drainage clip image data is used as a validation set, and the rest is used as a training set; the training set is input into the DeepLabv3+ semantic segmentation convolutional neural network model for training, and the training process includes an encoding process and a decoding process;

[0012] The encoding process is handled by the encoding part of the DeepLabv3+ semantic segmentation convolutional neural network model. The DCNN network performs downsampling and pooling operations to increase the number of feature channels of the network. Then, the ASPP network is processed to obtain five feature layers. The five feature layers are cascaded and convolved through the first convolutional layer to obtain the feature image.

[0013] The decoding process is handled by the decoding part of the DeepLabv3+ semantic segmentation convolutional neural network model. The five feature layers obtained during the encoding process are upsampled 4x through the first 4x upsampling module. The processed feature layers are cascaded with the feature layers amplified by the DCNN network, and the features are refined through the third convolutional layer. Finally, the predicted image of the center screw of the drainage clamp is obtained by processing through the second 4x upsampling module.

[0014] S400: Use the validation set to verify the recognition effect of the trained DeepLabv3+ semantic segmentation convolutional neural network model, completing the training of the DeepLabv3+ semantic segmentation convolutional neural network model;

[0015] S500: Use the trained DeepLabv3+ semantic segmentation convolutional neural network model to locate and identify the central screw of the drainage clamp image.

[0016] Furthermore, the specific method of expanding the drainage clip image data by using the image enhancement technology in step S300 is as follows:

[0017] (1) Randomly rotate the image with a rotation angle of 0 to 40°;

[0018] (2) Randomly shift the image horizontally or vertically, with the shift range being -20% to 20% of the image length or height;

[0019] (3) Perform random shear transformation on the image, and the image tilt angle is [-0.2°, 0.2°];

[0020] (4) Randomly scale the image with a scaling ratio of 0.8 to 1.2 times;

[0021] (5) Randomly flip half of the image horizontally;

[0022] (6) Use the nearest neighbor interpolation algorithm to randomly fill the newly created pixels on the image.

[0023] Furthermore, the dilated convolution increases the size of the convolution kernel by inserting a hole in the convolution kernel. The effective size K' of the convolution kernel after adding the hole is:

[0024] K'=K+(K-1)×(D-1)

[0025] Among them, K is the size of the convolution kernel before adding the hole, and D is the expansion rate;

[0026] The receptive field of the dilated convolution is:

[0027] F i+1 =(2 i+2 -1)*(2 i+2 -1)

[0028] Among them, F i+1 Represents the receptive field of layer i+1, where i is the number of layers of the receptive field.

[0029] The beneficial effects of the present invention are: the drainage wire clamp hole image is recognized by the DeepLabv3+ semantic segmentation convolutional neural network model, and a deep separable convolution is added to the DeepLabv3+ semantic segmentation convolutional neural network model to optimize the speed of the segmentation network and reduce the convolution calculation amount; and the parallel void convolution is adopted to increase the receptive field without losing information, so that the detailed spatial information is more certain, more data details can be obtained, and the recognition accuracy is improved; the entire recognition process includes encoding and decoding processes, which makes it easier to segment the wire clamp from the complex background. Compared with the existing PspNet, U-Net and SegNet networks, the present invention has different degrees of improvement in sensitivity, DICE similarity, IOU and PA indicators, the separation performance is more stable, the accuracy of processing the edge details of the drainage wire clamp is better, and the ability to segment the central screw is significantly enhanced, and the model loss value of the present application is the lowest, with good fitting effect and robustness; the method described in the present invention successfully realizes the segmentation of the drainage wire clamp from the complex background, and can achieve the purpose of automatic positioning without excessive reliance on high-definition cameras or light sources, providing a strong technical guarantee for the intelligentization of power grid drainage operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 This is a schematic diagram of the drainage wire passing through the wire clamp;

[0032] Figure 2 Schematic diagram of the structure of the DeepLabv3+ semantic segmentation convolutional neural network model in an embodiment of the present invention;

[0033] Figure 3 Schematic diagram of the structure of the improved Xception network in an embodiment of the present invention;

[0034] Figure 4 Schematic diagram of the structure of the dilated convolution in an embodiment of the present invention;

[0035] Figure 5 Schematic diagram of the structure of depth-wise separable convolution in an embodiment of the present invention;

[0036] Figure 6 Schematic diagram of an image after data enhancement and its corresponding label image in an embodiment of the present invention;

[0037] Figure 7 The figure shows the distribution of quantitative evaluation results of the embodiment of the present invention and three existing convolutional neural networks in a 5-fold experiment.

[0038] Figure 8 The loss curves for semantic segmentation achieved by the embodiment of the present invention and three existing convolutional neural networks are shown. DETAILED DESCRIPTION

[0039] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0040] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by persons of ordinary skill in the field described in this application. "First", "second" and similar words used in this patent application specification and claims do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, "one" or "an" and similar words do not indicate a quantity limitation, but rather indicate the existence of at least one. "Connected" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0041] like Figure 1 As shown, the drainage clamp consists of a first clamping block 1, a second clamping block 2, a screw support 3, and a clamping drive motor (not shown). When the drainage clamp reaches the connection position, the clamping drive motor 4 starts to operate, driving the screw support 3 to rotate, thereby bringing the first clamping block 1 and the second clamping block 2 closer together until the drainage line is clamped, achieving the purpose of connecting the main branch line.

[0042] To accurately locate the hole of the drain clamp, the embodiment of the present invention adopts a method for locating the drain clamp during live work on a distribution network based on computer vision, which segments the central screw, specifically including the following steps:

[0043] S100: randomly shuffling the collected drainage clip image data and evenly dividing them into five groups;

[0044] S200: Build DeepLabv3+ semantic segmentation convolutional neural network model, the specific structure is as follows Figure 2 As shown in FIG, the DeepLabv3+ semantic segmentation convolutional neural network model includes an encoding part and a decoding part.

[0045] The encoding part includes a DCNN network, an ASPP network, and the first convolutional layer. The DCNN network uses a modified Xception network as its base network. The modified Xception network replaces the maximum pooling operation in the original Xception network with strided dilated convolutions. The DCNN network is used to expand the number of feature channels. The ASPP network has five branches. The first branch is a 1x1 convolution. The second to fourth branches are a combination of depthwise separable convolutions with dilation rates of 18, 12, and 6, respectively, and dilated convolutions with a 3x3 kernel. This reduces the amount of computation while ensuring good segmentation. The fifth branch is global average pooling. The ASPP network outputs five feature layers. The first convolutional layer is a 1x1 convolution, which is used to cascade the five feature layers.

[0046] The decoding part includes the second convolutional layer, the third convolutional layer, the first 4x upsampling module and the second 4x upsampling module. The second convolutional layer and the third convolutional layer are 1*1 convolution and 3*3 convolution respectively. The feature layer with amplified feature channels and the five feature layers are processed by the second convolutional layer and the first 4x upsampling module respectively, then cascaded, and then processed by the third convolutional layer and the second 4x upsampling module in turn.

[0047] S300: The DeepLabv3+ semantic segmentation convolutional neural network model is trained using a five-fold cross-validation method. One set of drainage clip image data is used as a test set. The test set is used to evaluate the performance of the trained DeepLabv3+ semantic segmentation convolutional neural network model. The remaining four sets of drainage clip image data are expanded using image enhancement technology. 10% of the expanded drainage clip image data is used as a validation set, and the rest as a training set. This process is repeated five times until all collected drainage clip image data has been used once as a training set and a validation set. The training set is then input into the DeepLabv3+ semantic segmentation convolutional neural network model for training. The training process includes encoding and decoding.

[0048] The specific method for expanding the drainage clip image data by using image enhancement technology in the embodiment of the present invention is as follows:

[0049] (1) Randomly rotate the image with a rotation angle of 0 to 40°;

[0050] (2) Randomly shift the image horizontally or vertically, with the shift range being -20% to 20% of the image length or height;

[0051] (3) Perform random shear transformation on the image, and the image tilt angle is [-0.2°, 0.2°];

[0052] (4) Randomly scale the image with a scaling ratio of 0.8 to 1.2 times;

[0053] (5) Randomly flip half of the image horizontally;

[0054] (6) Use the nearest neighbor interpolation algorithm to randomly fill the newly created pixels on the image.

[0055] Figure 6These are representative enhanced images randomly selected and their corresponding label images. The image data used in the embodiments of the present invention were all taken on-site using the camera on the robot, totaling 187 images. Each image was taken strictly according to the perspective of the aerial live-working robot. The 187 original images collected were randomly shuffled and evenly divided into five groups, each with about 37 images. The first group of image data is the test set, and the remaining four groups of image data are expanded using image enhancement technology, 10% of which are validation sets and the rest are test sets, and so on, completing a 5-fold cross-validation scheme. The original resolution of each image is 4512*4512. In order to adapt to CPU performance, shorten training time, and improve training efficiency, the image resolution is adjusted to 512*512 during training. Since the camera used to identify the drain clamp is installed on the robotic arm of the aerial live-working robot, its position is fixed, so the positions of the drain clamp screw and busbar in each image are not much different. In order to improve the generalization ability of the model, each image in the training set has adopted data augmentation methods such as center cropping, random cropping, horizontal flipping, vertical flipping, scaling enhancement, random rotation, shearing and brightness change, which has increased the number of images in the training set by 30 times.

[0056] The encoding process is handled by the encoding part of the DeepLabv3+ semantic segmentation convolutional neural network model. The DCNN network performs downsampling and pooling operations to increase the number of feature channels of the network. Then, it is processed by the ASPP network to obtain five feature layers. The five feature layers are cascaded and convolution processing is performed through the first convolutional layer to obtain the feature image.

[0057] In the embodiment of the present invention, the DCNN network gradually increases the number of feature channels from 64, 128, 256, 728, 1024, 1536 to 2048. As the number of feature channels increases, the feature calculation amount of each layer increases sharply. To solve this technical problem, the embodiment of the present invention uses the following in the DCNN network: Figure 5 The improved Xception network shown in the figure serves as the base network and replaces the maximum pooling operation in the Xception network with strided dilated convolutions. This enables strided dilated convolutions to extract features from images of any resolution. Furthermore, after each 3x3 depth-dilated convolution, additional data normalization (Batchnormalization) and ReLU activation operations are added to improve the overall performance of the network.

[0058] During the ASPP network processing, the first branch's 1x1 convolution fuses information at different scales, while the second through fourth branches are used to capture features at different scales. This process doesn't capture the entire image, necessitating the addition of the first and fifth branches to overcome the gradual loss of detail and reduced image resolution caused by continuous downsampling. At the same time, each output convolution encompasses a wider range of information without increasing computational complexity, thus increasing the receptive field, speeding up semantic segmentation, and further improving recognition accuracy.

[0059] like Figure 4 As shown in the figure, the dilated convolution increases the size of the convolution kernel by inserting holes in it. Without losing information, it increases the receptive field and allows each convolution output to contain a wider range of information. After inserting D-1 holes between every two elements of the convolution kernel, the effective size K' of the convolution kernel is:

[0060] K'=K+(K-1)×(D-1) (1)

[0061] Among them, K is the size of the convolution kernel before adding the hole, and D is the number of holes;

[0062] The receptive field of the dilated convolution is:

[0063] F i+1 =(2 i+2 -1)*(2 i+2 -1) (2)

[0064] Among them, F i+1 represents the receptive field of layer i+1, where i is the number of layers in the receptive field. When the dilation ratio i+1 = 2, the receptive field increases from 3*3 to 7*7 while keeping the convolution kernel size unchanged. As can be seen from the above formula, dilated convolution is used when a network layer requires a larger receptive field but cannot increase the number of convolution kernels.

[0065] like Figure 5 As shown in the figure, depthwise separable convolution decomposes the standard convolution into depthwise convolution and pointwise convolution, where depthwise convolution performs spatial convolution on each input channel independently, and pointwise convolution is used to combine the output of depthwise convolution. The ratio of the computational effort of depthwise separable convolution to that of standard convolution can be calculated by the following formula:

[0066]

[0067] Among them, C in and C out are the number of channels of the input image and the output image, H out and W out are the height and width of the output image, D F is the convolution kernel size of depth-wise separable convolution. The above formula can be simplified to:

[0068]

[0069] From the above formula, we can see that depth-wise separable convolution is highly applicable in solving the problem of trainable parameters or model overfitting.

[0070] The decoding process is handled by the decoding part of the DeepLabv3+ semantic segmentation convolutional neural network model. The five feature layers obtained in the encoding process are upsampled 4 times through the first 4x upsampling module. The feature layer with 256 output feature channels in the DCNN network is processed by 1*1 convolution of the second convolutional layer, and then cascaded with the feature layer after 4x upsampling. The features are refined by 3*3 convolution of the third convolutional layer. Finally, the image with extracted features is restored to the original input image size through the second 4x upsampling module to obtain the predicted image of the center screw of the drainage clamp.

[0071] S400: Use the validation set to verify the recognition effect of the trained DeepLabv3+ semantic segmentation convolutional neural network model, and complete the training of the DeepLabv3+ semantic segmentation convolutional neural network model.

[0072] S500: Use the trained DeepLabv3+ semantic segmentation convolutional neural network model to locate and identify the central screw of the drainage clamp image.

[0073] The following experiments illustrate the recognition effect of the embodiment of the present invention. This experiment selected three mainstream networks, U-Net, SegNet, and PspNet, for comparative testing with the embodiment of the present invention. Four indicators, sensitivity, intersection-over-union (IOU), similarity coefficient (DICE), and pixel accuracy (PA), were used for evaluation. The specific calculation formula is as follows:

[0074]

[0075]

[0076]

[0077]

[0078] Where TP represents the number of line-clip images in the test sample that can be accurately predicted as line-clip images, TN represents the number of non-line-clip images in the test sample that can be accurately predicted as non-line-clip images, FP represents the number of non-line-clip images in the test sample that can be accurately predicted as line-clip images, and FN represents the number of line-clip images in the test sample that can be accurately predicted as non-line-clip images. Sensitivity represents the proportion of images in the test sample that can be accurately identified. The intersection-over-union (IOU) metric represents a standard metric for semantic segmentation. It represents the ratio of the correctly predicted pixel area to the calibrated pixel area and the predicted pixel area, that is, the ratio of the intersection to the union of two sets. The similarity coefficient (DICE) is a set similarity metric that calculates the ratio of twice the area of ​​the overlapping area between the true and predicted values ​​to the sum of their respective areas. As the DICE value approaches 0, the two contour areas are less similar; as the DICE value approaches 1, the two contour areas are more similar. Pixel accuracy (PA) represents the ratio of pixels with correctly predicted categories to the total number of pixels.

[0079] The average results of the four indicators of the embodiment of the present invention and the U-Net, SegNet and PspNet networks in the experiment are shown in Table 1:

[0080] Table 1 Average results of PspNet / U-Net / SegNet / DeepLabv3+ in four quantitative evaluations

[0081]

[0082] As shown in Table 1, based on the pixel accuracy (PA) metric, all networks successfully segmented the center screw of the power grid drain clamp. Evaluations of the DICE (Dissimilarity Index) and the IOU (Intersection Over Union) metric show that the 5-fold average segmentation performance of the embodiment of the present invention is 93.88% and 96.84%, respectively. These improvements are 11.37% and 5.90% higher for the DICE and 3.31% and 1.54% higher for the PspNet, U-Net, and SegNet networks, respectively, as well as 3.74% and 1.94% higher for the IOU.

[0083] In order to quantitatively evaluate and compare the segmentation performance of the four convolutional neural networks, Figure 7 The distribution of four convolutional neural networks (CNNs) in a five-fold cross-validation experiment shows the sensitivity, intersection-over-union (IOU), similarity coefficient (DICE), and pixel accuracy (PA). These results demonstrate that the segmentation performance of the embodiment of the present invention is more stable. This demonstrates that the multi-scale, wide-field-of-view DeepLabv3+ network has stronger semantic segmentation performance and is more suitable for segmenting and locating power grid drain clamp screws.

[0084] like Figure 8 As shown in Figure 2, after 100 rounds of training, the loss curves of the embodiment of the present invention are compared with those of the U-Net, SegNet, and PspNet networks. Figure 8 As can be seen, the loss values ​​of the embodiment of the present invention and the U-Net, SegNet, and PspNet networks all dropped to a near-convergence level after around 20 rounds. After another 80 rounds of gradient descent, the losses of all four network models converged. A comparison of the loss functions shows that the embodiment of the present invention has the lowest loss value, approximately 0.02; the U-Net and SegNet networks have intermediate loss values, while the PspNet network has the highest loss value. This demonstrates that the embodiment of the present invention achieves better fitting results and exhibits superior robustness.

[0085] Compared with the existing PspNet, U-Net and SegNet networks, the embodiments of the present invention, after fully utilizing the channel information of the feature dimension, have better accuracy in processing the edge details of the drainage wire clamp, which significantly enhances the semantic segmentation capability of the drainage wire clamp, and is conducive to the subsequent automatic installation of the power grid drainage line, solving the problems of high labor intensity and low operation efficiency of manual operation.

[0086] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for locating drainage clamps for live-line work in distribution networks based on computer vision, characterized in that: The steps include: S100: randomly shuffling the collected drainage clip image data and evenly dividing them into five groups; S200: Construct a DeepLabv3+ semantic segmentation convolutional neural network model, which includes an encoding part and a decoding part; the encoding part includes a DCNN network, an ASPP network and a first convolutional layer; the DCNN network uses an improved Xception network as a base network, and the improved Xception network uses a dilated convolution with stride to replace the maximum pooling operation in the original Xception network, and the DCNN network is used to expand the number of feature channels; the ASPP network includes five branches, the first branch is convolution, the second to fourth branches are a combination of depthwise separable convolution and dilated convolution, and the fifth branch is global average pooling. The ASPP network is used to output five feature layers, and the first convolutional layer is used to cascade five feature layers; the decoding part includes a second convolutional layer, a third convolutional layer, a first 4x upsampling module and a second 4x upsampling module, and the feature layer with expanded feature channels and the five feature layers are cascaded after being processed by the second convolutional layer and the first 4x upsampling module, and then processed by the third convolutional layer and the second 4x upsampling module in sequence; S300: The DeepLabv3+ semantic segmentation convolutional neural network model is trained using a five-fold cross-validation method, using one set of drainage clip image data as a test set, and expanding the remaining four sets of drainage clip image data using image enhancement technology. 10% of the expanded drainage clip image data is used as a validation set, and the rest is used as a training set; the training set is input into the DeepLabv3+ semantic segmentation convolutional neural network model for training, and the training process includes an encoding process and a decoding process; The encoding process is handled by the encoding part of the DeepLabv3+ semantic segmentation convolutional neural network model. The DCNN network performs downsampling and pooling operations to increase the number of feature channels of the network. Then, the ASPP network is processed to obtain five feature layers. The five feature layers are cascaded and convolved through the first convolutional layer to obtain the feature image. The decoding process is handled by the decoding part of the DeepLabv3+ semantic segmentation convolutional neural network model. The five feature layers obtained during the encoding process are upsampled 4x through the first 4x upsampling module. The processed feature layers are cascaded with the feature layers amplified by the DCNN network, and the features are refined through the third convolutional layer. Finally, the predicted image of the center screw of the drainage clamp is obtained by processing through the second 4x upsampling module. S400: Use the validation set to verify the recognition effect of the trained DeepLabv3+ semantic segmentation convolutional neural network model, completing the training of the DeepLabv3+ semantic segmentation convolutional neural network model; S500: Use the trained DeepLabv3+ semantic segmentation convolutional neural network model to locate and identify the central screw of the drainage clamp image.

2. A method for locating a drainage clamp for live-line operation in a distribution network based on computer vision according to claim 1, characterized in that: The specific method of expanding the drainage clip image data by using the image enhancement technology in step S300 is as follows: (1) Randomly rotate the image with a rotation angle of 0 to 40°; (2) Randomly shift the image horizontally or vertically, with the shift range being -20% to 20% of the image length or height; (3) Perform random shear transformation on the image, and the image tilt angle is [-0.2°, 0.2°]; (4) Randomly scale the image with a scaling ratio of 0.8 to 1.2 times; (5) Randomly flip half of the image horizontally; (6) Use the nearest neighbor interpolation algorithm to randomly fill the newly created pixels on the image.

3. The method for locating a drainage clamp for live-line operation in a distribution network based on computer vision according to claim 1, characterized in that: The dilated convolution increases the size of the convolution kernel by inserting a hole in the convolution kernel. The effective size K' of the convolution kernel after adding the hole is: K'=K+(K-1)×(D-1) Among them, K is the size of the convolution kernel before adding the hole, and D is the expansion rate; The receptive field of the dilated convolution is: F i+1 ＝(2 i+2 -1)*(2 i+2 -1) Among them, F i+1 Represents the receptive field of layer i+1, where i is the number of layers of the receptive field.

Images