Horizontal telescopic disconnecting switch detection method, system and equipment based on semantic segmentation and storage medium

Abstract

The invention discloses a horizontal telescopic disconnecting switch detection method, system and device based on semantic segmentation and a storage medium. The method comprises the steps that image frames are extracted from a video, images are preprocessed, semantic segmentation model training is carried out, and a model with the optimal mIou index is generated; taking the opening and closing state image of the disconnecting switch as a template base map, and marking an opening and closing polygonal region; testing an actual disconnecting switch state image, matching the test image with the opening and closing image, and solving an actual opening and closing area; and loading the trained semantic segmentation model, carrying out tool arm mask acquisition on the test image, calculating the contour of the obtained tool arm mask, and calculating the intersection-union ratio of the contour and the splitting and integrating area to obtain the actual state of the disconnecting switch. According to the method, on the premise that no extra sensor is additionally arranged, the recognition accuracy of the horizontal telescopic disconnecting switch of the transformer substation can be improved, and particularly, the problem of remote recognition of the disconnecting switch is effectively solved under the condition that the preset position of the disconnecting switch deviates.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image semantic segmentation, and in particular to a horizontal telescopic disconnecting switch detection method, system, device and storage medium based on semantic segmentation. BACKGROUND

[0002] The horizontal telescopic disconnecting switch realizes the opening and closing operation through horizontal telescopic movement, is widely used in 220kV and 500kV substations, and is used for isolating power supply, switching operation and small current circuit on-off. The core structure thereof comprises: a conductive arm which is designed in a folding or telescopic manner, forms a horizontal break when being opened, and ensures a visible insulation gap; when being closed, the conductive arm is straightened and contacts a static contact, thereby forming a conductive path; and an insulation support which is usually a double-column structure, one column of which supports the conductive static contact, and the other column of which is provided with a folding movable arm. The insulation support has the functions of supporting and operating, and ensures the insulation performance under high voltage.

[0003] The opening and closing states of the horizontal telescopic disconnecting switch are directly related to the safe operation of the power grid. When the switch is opened with load, if it is mistakenly operated as being in the closed state, arc short circuit may be caused, the equipment may be burned, and even the safety of personnel may be endangered. When the switch is closed with live ground, if the main switch is not confirmed to be opened before the ground switch is closed, the system may be grounded, and large-area power failure may be caused. Before maintenance, it is necessary to confirm that the disconnecting switch is opened and grounded, otherwise the maintenance personnel may contact the live part, and electric shock accident may be caused. SUMMARY

[0004] The present application aims to provide a horizontal telescopic disconnecting switch detection method which does not need to install additional sensors and can effectively realize remote identification of the disconnecting switch under the condition of preset position deviation of the disconnecting switch. Another object of the present application is to provide a horizontal telescopic disconnecting switch detection system, device and storage medium based on semantic segmentation.

[0005] Technical scheme: The horizontal telescopic disconnecting switch detection method based on semantic segmentation comprises the following steps: Step S1: constructing a horizontal telescopic disconnecting switch dataset according to the horizontal telescopic disconnecting switch video in the substation; Step S2: constructing a semantic segmentation model, training the model using the horizontal telescopic disconnecting switch dataset, and selecting the semantic segmentation model with the highest mIou index as an inference model; Step S3: labeling a disconnecting switch opening polygon area and a disconnecting switch closing polygon area for the disconnecting switch opening image and the disconnecting switch closing image, respectively, to obtain disconnecting switch opening / closing area picture templates; Step S4: performing affine transformation on the disconnecting switch opening / closing area picture templates and a to-be-detected picture, respectively, to obtain the opening / closing area of the to-be-detected picture; Step S5: using the inference model to identify the picture to be detected, calculating the knife arm mask; Step S6: generating the contour according to the knife arm mask, calculating the intersection over union of the split / combined area of the picture to be identified, and determining the horizontal telescopic disconnecting switch state through the intersection over union.

[0006] Further, the step S1 includes the following steps: Step S11: collecting the horizontal telescopic disconnecting switch video in the transformer substation, and the video includes the process of the horizontal telescopic disconnecting switch from splitting to combining or from combining to splitting; Step S12: extracting the key frame of the collected horizontal telescopic disconnecting switch video through the opencv function library, and saving the picture; Step S13: performing data preprocessing on the saved picture image data, and the data preprocessing includes correcting the dark or bright image, enhancing the knife arm edge details, scaling, adjusting the contrast, and adjusting the saturation; Step S14: using the open source software labelme to label the disconnecting switch knife arm of the preprocessed horizontal telescopic disconnecting switch picture, and obtaining the horizontal telescopic disconnecting switch dataset.

[0007] Further, the step S2 includes the following steps: Step S21: constructing a semantic segmentation model based on the mmsegmentation framework, and performing semantic segmentation training of the horizontal telescopic disconnecting switch knife; the Swin Transformer is selected as the backbone network in the semantic segmentation model, and the swin_base_patch4_window12_384_22k model trained on the ImageNet-22K is selected as the pre-training weight; Step S22: adopting the Mask2Former algorithm in the DecodeHead stage, setting the window size and embedding dimension of the self-attention calculation, generating the segmentation mask through the Transformer decoder, and selecting the classification loss, dice loss, and mask loss as the loss function; Step S23: in the training stage, the AdamW is selected as the optimizer, the learning rate, weight decay, and momentum are set, the learning rate is adjusted, and the PolyLR strategy is adopted.

[0008] Further, the step S3 is specifically as follows: The open source software labelme is used to label the split / combined area of the disconnecting switch split image and the disconnecting switch combined image, the split area label is named as open_area, the combined area label is named as close_area, the coordinates of the split / combined area are represented by the four vertex coordinates of the clockwise polygon, and the combined area coordinates are , and the split area coordinates are .

[0009] Furthermore, step S4 includes the following steps: Step S41: Set the number of feature points to be detected, the detection threshold, and the radius threshold. Use the Superpoint network to extract key points and descriptors from the images of the opening / closing regions of the disconnector and the images to be detected, respectively. Step S42: Use the Lightglue network to perform local feature matching on the extracted keypoints, and output the index of the matched keypoints, the matching score, and the confidence level. Step S43: Calculate the fundamental matrix based on the key points of the output matching, and filter out mismatched points; wherein, the calculation of the fundamental matrix is ​​to set the maximum distance threshold from the point to the epipolar line, using the cv2.findFundamentalMat function of OpenCV, and the method parameter uses cv2.FM_RANSAC; Step S44: Calculate the homography matrix for the filtered matching points and set the reprojection error threshold; the homography matrix is ​​calculated using the cv2.findHomography function of OpenCV. Step S45: Perform a perspective transformation on the coordinates of the split / combined regions using OpenCV's cv2.perspectiveTransform function to obtain the transformed coordinates of the combined region. Transformed regional coordinates .

[0010] Furthermore, step S6 includes the following steps: Step S61: Preprocess the tool arm mask and extract its contour; Step S62: Use list S to store the extracted contour polygon information. 轮廓 express; Step S63: According to list S 轮廓 The combined coordinates of the transformed region S 合 and the transformed regional coordinates S 分 , please find Iou; Step S64: If Iou(S 轮廓， S 合 If Iou(S) > 0.9, then the task isolation switch is in the "closed" state; if Iou(S) > 0.9, then the task isolation switch is in the "closed" state; 轮廓， S 分 If the value is greater than 0.9, the task disconnect switch is in the "open" state; otherwise, the disconnect switch is in the "intermediate state".

[0011] Furthermore, the mask preprocessing involves using the function cv2.approxPolyDP to approximate the contours into polygons, filtering contours with an area less than the threshold of 30, smoothing the edges, and then extracting the contours.

[0012] The semantic segmentation-based horizontal telescopic disconnector detection system of the present invention includes: The dataset module is used to construct a dataset of horizontal telescopic disconnect switches based on videos of horizontal telescopic disconnect switches in substations. The model building module is used to build a semantic segmentation model. It is trained using a horizontally scalable disconnector dataset and the semantic segmentation model with the highest mIou index is selected as the inference model. The annotation and segmentation module is used to annotate the polygonal regions of the disconnector switch when it is open and when it is closed, respectively, to obtain images of the open / closed regions of the disconnector switch. The acquisition module is used to perform affine transformations on the images of the open / closed regions of the disconnector switch and the images to be detected, respectively, to obtain the open / closed regions of the image to be identified. The mask calculation module is used to identify the image to be detected using an inference model and calculate the knife arm mask. The calculation and judgment module is used to generate a contour based on the knife arm mask, calculate the intersection / union ratio with the split / combined region of the image to be identified, and determine the state of the horizontal telescopic disconnect switch by the intersection / union ratio.

[0013] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0014] The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0015] Beneficial effects: Compared with the prior art, the significant advantages of this invention are: This invention accurately determines the open / closed state of a horizontal telescopic disconnector switch through image semantic segmentation when only a complete open / closed video segment is acquired, even with minor changes in open / closed position; no manual intervention is required to adjust the pan / tilt position, as the algorithm automatically completes key point extraction, offset calculation, and image correction, ensuring stable and accurate determination of the disconnector switch's open / closed state even in complex field environments (such as old substations where pan / tilt offset is prone to occur); simultaneously, since no manual on-site camera adjustment or data re-acquisition is needed, the frequency and workload of manual inspections are significantly reduced, lowering the labor intensity and safety risks for maintenance personnel, and significantly improving the automation level and economic benefits of power equipment operation and maintenance. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0017] The method for detecting horizontally retractable disconnect switches based on semantic segmentation as described in this invention includes the following steps: Step S1: Construction of Horizontal Telescopic Disconnect Switch Dataset: Collect videos of horizontal telescopic disconnect switches in the substation, extract keyframes, and generate JSON files through data preprocessing and data augmentation. The specific steps are as follows: Step S11: Conduct video capture of the operation process of the horizontal retractable disconnector switch within the substation. Specifically, high-definition video data containing the complete operating cycle of the disconnector switch (i.e., from the open state to the closed state, or from the closed state to the open state) needs to be acquired. The captured video files should be stored in the widely compatible MP4 format to ensure the universality and ease of processing of the video data. Simultaneously, to ensure the clarity and detail of the video image, the video resolution is set to 1920 pixels wide × 1280 pixels high. Step S12: Based on the OpenCV open-source computer vision library, keyframe intelligent extraction processing is performed on the substation horizontal telescopic disconnector operation video acquired in Step S11. Specifically, an inter-frame difference algorithm combined with a timed sampling strategy is used, performing frame sequence filtering at 5-frame intervals—that is, after processing 5 consecutive video images, only the 5th frame is retained as the representative frame, and the remaining frames are automatically discarded. This sampling interval has been experimentally verified to reduce the amount of data to be labeled by 80% while ensuring the continuous monitoring of the action process, significantly reducing the workload of subsequent manual labeling.

[0018] Step S13: Perform multi-dimensional preprocessing on the JPEG format image data saved in step S12, specifically including: using an adaptive brightness enhancement algorithm based on Retinex theory to perform nonlinear stretching on overly dark areas and brightness compression to preserve local details in overexposed areas; and using the Canny edge detection operator combined with the Laplacian operator for sharpening to improve the gradient intensity of the blade edge. Step S14: Enhance the image using a combination of random scaling (uniform sampling within the range of 0.8×~1.2×), random rotation (-15°~+15° angle range), brightness adjustment (±30% dynamic range), contrast enhancement (γ value 0.7~1.3), and saturation change (±25%) to simulate the actual state of the disconnector switch under outdoor lighting conditions. Step S15: Use the open-source software labelme to annotate the enhanced disconnector image from Step S14. Annotate the main area of ​​the disconnector arm uniformly, excluding non-critical components such as the operating mechanism box and insulators. The label is “switch” and the annotation type is an irregular quadrilateral annotation.

[0019] Step S2: Model Design and Training: Design a semantic segmentation model, train it using the dataset from Step S1, and generate the inference model with the highest mIou. The specific steps are as follows: Step S21: The semantic segmentation model referred to in this step adopts a dedicated deep learning model built on the mmsegmentation open-source framework. This framework integrates a multi-scale feature fusion mechanism and dynamic computation graph optimization technology, and is specifically optimized for the semantic segmentation task of the horizontally telescopic disconnector switch area (pixel proportion of 8%~15%) in power equipment images. The model input layer is configured as a three-channel RGB image (resolution 384×384), and the output layer generates a single-channel probability map (continuous values ​​in the range of 0~1) with the same size as the input, which is then converted into a binary segmentation mask through thresholding (default threshold 0.7). Step S22: Select the Swin Transformer as the backbone network for feature extraction, with the following specific configuration: Window partitioning strategy: A 12×12 pixel non-overlapping sliding window mechanism is adopted, with the window displacement step size set to 1 / 2 of the window size (i.e., 6 pixels) to achieve cross-window information interaction; Embedding dimension control: After the input image is processed by a 4×4 pixel linear embedding layer, a 192-dimensional feature vector is generated; Pre-training transfer: The weights of the swin_base_patch4_window12_384_22k model trained on the ImageNet-22K dataset (14 million images, 22K categories) are loaded. These weights transfer high-level semantic features to the sub-domain of power equipment through knowledge distillation, which can significantly improve the initial convergence speed of the model; Step S23: In the DecodeHead stage, the Mask2Former (Mask-based Transformer) algorithm is used to achieve end-to-end segmentation. Its design includes: a Transformer decoder configured with 6 layers of cross-attention modules, each containing 12 attention heads (head dimension=64), enhancing spatial awareness through Dynamic Position Encoding (DPE); and the introduction of multi-task loss functions: classification loss (Focal Loss, γ=2.0) to address class imbalance; Dice loss (ε=1e-6) to optimize boundary region overlap; and mask loss (L1 Loss) to ensure segmentation continuity. The three loss functions are weighted and combined with a weight ratio of 0.4:0.3:0.3, resulting in an F1 score of 98.7% for the blade edge (width ≤ 5 pixels). Step S3: Generation of the split and merge regions: Use the open-source software Labelme to label and collect two images of the disconnector switch, one for splitting and one for merging. The split region is labeled "open_area," and the merge region is labeled "close_area." The coordinates of the split and merge regions are represented by the coordinates of the four vertices of a clockwise polygon. The coordinates of the merge region are... Regional coordinates ; Step S4: Perform affine transformations on the images to be detected using the images obtained from step S3, respectively, to determine the regions to be separated or merged in the image to be identified. The specific steps are as follows: Step S41: Using the Superpoint neural network architecture based on deep learning, feature point detection and descriptor extraction operations are performed on the standard image of the disconnector switch in the closed state and the image to be detected, respectively. Specific parameter configurations are as follows: Number of feature points to be detected: set to 4096 per image to ensure that the feature space distribution density meets the subsequent matching requirements; Detection confidence threshold: set to 0.05 to balance feature point detection sensitivity and noise suppression capability; Neighborhood radius threshold: set to 2 pixels for local feature point selection in non-maximum suppression processing. This step outputs a feature point set containing the spatial coordinates of the feature points and a 256-dimensional descriptor vector, providing basic data for subsequent feature matching.

[0020] Step S42: An improved Lightglue feature matching network is used to perform cross-domain matching on the feature point sets of the two images extracted in Step S41. This network enhances the correlation of local features through an attention mechanism, outputting a matching result set containing the following elements: (1) Matching point pair index: Records the index mapping relationship between the corresponding feature points in the standard image and the image to be inspected; (2) Matching similarity score: The quality assessment value of matching in the 0-1 interval calculated using cosine similarity; (3) Match confidence: The probability value of matching reliability in the 0-1 interval predicted by the neural network; (4) The matching results are sorted by confidence level, and the top 80% of high-confidence matching pairs are retained for subsequent processing.

[0021] Step S43: Use the cv2.findFundamentalMat function from the OpenCV library to perform the fundamental matrix calculation. The specific implementation parameters are as follows: Estimation method: The robust RANSAC (Random Sample Consensus) algorithm is selected; Distance threshold: Set the maximum distance for epipolar geometry constraints to 3 pixels; Number of iterations: The optimal number of iterations is automatically calculated based on the number of feature points; This step constructs epipolar geometric constraints using a fundamental matrix, eliminates mismatched point pairs that deviate from the constraints by more than a threshold, and outputs a geometrically validated set of high-quality matching points.

[0022] Step S44: Based on the matched point set filtered in step S43, the image homography transformation matrix is ​​calculated using OpenCV's cv2.findHomography function. Key parameter settings include: Reprojection error threshold: set to 2 pixels to ensure that the transformation accuracy meets the equipment detection requirements; Solution method: The Direct Linear Transform (DLT) algorithm is combined with RANSAC optimization. The calculated 3×3 homography matrix H is used to describe the projection transformation relationship between the standard image and the image to be inspected.

[0023] Step S45: Using the homography matrix H obtained in step S44, perform a perspective transformation on the coordinates of the key equipment area defined in step S32: Input coordinates: includes the set of vertex coordinates of the closing area S_close_in and the set of vertex coordinates of the opening area S_open_in; Transformation function: OpenCV's cv2.perspectiveTransform function is used; Output coordinates: Generates the transformed set of spatial coordinates; Closing area: S_close_out=[[x'1 close,y'1 close],[x'2 close,y'2 close],[x'3 close,y'3 close],[x'4 close,y'4 close]]; Opening area: S points_out=[[x'1 points,y'1 points],[x'2 points,y'2 points],[x'3 points,y'3 points],[x'4 points,y'4 points]]; The transformation results are used for subsequent device status identification and position accuracy verification to ensure that the coordinate mapping error is controlled within ±1.5 pixels.

[0024] Step S5: Use the model trained in step S2 to identify the image to be detected and obtain the knife arm mask; Step S6: Application of the disconnector detection system: Generate a contour based on the mask obtained in step S5, and calculate the intersection-exchange ratio with the separation and combination areas in step S4. Determine the actual state of the disconnector through the intersection-exchange ratio to realize the state detection of the horizontal telescopic disconnector. The specific steps are as follows: Step S61: Use the model trained in step S2 to infer and generate a knife arm mask from the image to be detected, preprocess the mask and extract its contours; Step S62: The mask preprocessing in step S61 mainly includes: using the function cv2.approxPolyDP to approximate the contour with polygons to reduce the number of contour points; filtering contours with an area less than the threshold of 30 to remove noise; and edge smoothing, first applying Gaussian blur to the binary mask and then extracting the contour to reduce jagged edges. Step S63: The contour polygon information generated in step S61 is represented by a list, S contour = [[x1,y1],[x2,y2]......[xn,yn]], where n represents the number of contour points; Step S64: Use the S-contour from step S63 and the S-combination and S-divide from step S45 to calculate Iou, where the formula for calculating Iou is as follows: ; Where A and B represent two regions (which can be bounding boxes, segmentation masks, contours, etc.), A represents the intersection area of ​​A and B (the area of ​​the overlapping part), and B represents the union area of ​​A and B (the total area covered by the two regions). In this example, A represents the S contour, and B represents the S union and S division.

[0025] Step S65: If Iou(Scontour, Sclose) > 0.9, the task isolation switch is in the "closed" state; if Iou(Scontour, Sclose) > 0.9, the task isolation switch is in the "closed" state; otherwise, the isolation switch is determined to be in the "intermediate state" (i.e., neither closed nor closed).

[0026] Step S66: Based on the Flask framework in Python, encapsulate the semantic segmentation model into an HTTP interface and design a horizontally scalable disconnector detection system. Receive JSON messages sent from the webpage via a POST interface. The image data of the disconnector switch to be tested is encoded in Base64; Step S67: The system receives Base64 encoded information, converts it into OpenCV NumPy, and finally outputs the status of the isolation switch and returns it to the front-end interface.

[0027] The semantic segmentation-based horizontal telescopic disconnector detection system of the present invention includes: The dataset module is used to construct a dataset of horizontal telescopic disconnect switches based on videos of horizontal telescopic disconnect switches in substations. The model building module is used to build a semantic segmentation model. It is trained using a horizontally scalable disconnector dataset and the semantic segmentation model with the highest mIou index is selected as the inference model. The annotation and segmentation module is used to annotate the polygonal regions of the disconnector switch when it is open and when it is closed, respectively, to obtain images of the open / closed regions of the disconnector switch. The acquisition module is used to perform affine transformations on the images of the open / closed regions of the disconnector switch and the images to be detected, respectively, to obtain the open / closed regions of the image to be identified. The mask calculation module is used to identify the image to be detected using an inference model and calculate the knife arm mask. The calculation and judgment module is used to generate a contour based on the knife arm mask, calculate the intersection / union ratio with the split / combined region of the image to be identified, and determine the state of the horizontal telescopic disconnect switch by the intersection / union ratio.

[0028] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0029] The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

Claims

1. A method for detecting a horizontally telescopic disconnector based on semantic segmentation, characterized in that, The method comprises the following steps: Step S1: constructing a horizontal telescopic disconnecting switch dataset according to horizontal telescopic disconnecting switch videos in a transformer substation; Step S2: constructing a semantic segmentation model, training the semantic segmentation model using the horizontal telescopic disconnecting switch dataset, and selecting a semantic segmentation model with the highest mIou index as an inference model; Step S3: labeling a disconnecting switch open polygon region and a disconnecting switch closed polygon region respectively for a disconnecting switch open image and a disconnecting switch closed image to obtain a disconnecting switch open / closed region picture template; Step S4: performing affine transformation on the disconnecting switch open / closed region picture template and a to-be-detected picture respectively to obtain a disconnecting switch open / closed region of the to-be-detected picture; Step S5: identifying the to-be-detected picture using the inference model to calculate a blade mask; Step S6: generating a contour according to the blade mask, calculating an intersection-over-union of the contour and the disconnecting switch open / closed region of the to-be-detected picture, and determining a horizontal telescopic disconnecting switch state through the intersection-over-union.

2. The method according to claim 1, wherein, The step S1 comprises the following steps: Step S11: collecting horizontal telescopic disconnecting switch videos in a transformer substation, wherein the videos include processes of the horizontal telescopic disconnecting switch from opening to closing or from closing to opening; Step S12: performing key frame extraction on the collected horizontal telescopic disconnecting switch videos through an opencv function library to save pictures; Step S13: performing data preprocessing on the saved picture image data, wherein the data preprocessing includes correcting dark or bright images, enhancing blade edge details, scaling, adjusting contrast, and adjusting saturation; Step S14: labeling disconnecting switch blades using open source software labelme for the preprocessed horizontal telescopic disconnecting switch pictures to obtain a horizontal telescopic disconnecting switch dataset.

3. The method according to claim 1, wherein the method is characterized by, The step S2 comprises the following steps: Step S21: constructing a semantic segmentation model based on a mmsegmentation framework to perform semantic segmentation training of a horizontal telescopic disconnecting switch; in the semantic segmentation model, a Swin Transformer is selected as a backbone network, and a swin_base_patch4_window12_384_22k model trained on an ImageNet-22K is selected as pre-training weights; Step S22: in a DecodeHead stage, a Mask2Former algorithm is adopted, a window size and an embedding dimension for self-attention calculation are set, a segmentation mask is generated through a Transformer decoder, and a classification loss, a dice loss, and a mask loss are selected as loss functions; Step S23: in a training stage, an AdamW is selected as an optimizer, a learning rate, a weight decay, and a momentum are set, the learning rate is adjusted, and a PolyLR strategy is adopted.

4. The method according to claim 1, wherein, The step S3 is specifically as follows: The open-close switch split image and the open-close switch combined image are annotated using open source software labelme software. The split region label is named open_area, the combined region label is named close_area, the coordinates of the split / combined region are represented by the four vertex coordinates of the clockwise polygon, and the combined region coordinates , the split region coordinates .

5. The method according to claim 1, wherein, The step S4 comprises the following steps: Step S41: setting a feature point detection quantity, a detection threshold, and a radius threshold, and extracting key points and descriptors from a disconnecting switch open / closed region picture and a to-be-detected picture using a Superpoint network; Step S42: local feature matching is performed on the extracted key points using the Lightglue network, and the matched key point index, matching score and confidence are output; Step S43: according to the output matched key points, the fundamental matrix is calculated, and the false matching points are filtered; wherein, the fundamental matrix calculation is set to the maximum distance threshold of the point to the epipolar line, and the cv2.findFundamentalMat function of opencv is used, and the method parameter uses cv2.FM_RANSAC; Step S44: homography matrix calculation is performed on the filtered matching points, and a reprojection error threshold is set; wherein, the homography matrix calculation is calculated using the cv2.findHomography function of opencv; Step S45: Perspective transformation is performed on the split / merge region coordinates, and the cv2.perspectiveTransform function of opencv is used to obtain the transformed joint region coordinates , the transformed split region coordinates .

6. The method according to claim 5, wherein the method further comprises: The step S6 comprises the following steps: Step S61: pretreat the knife arm mask and perform contour extraction; Step S62: The extracted contour polygon information is represented in a list S 轮廓 ; Step S63: According to list S 轮廓 with the transformed combined region coordinates S 合 and the transformed sub-region coordinates S 分 , Iou is calculated; Step S64: If Iou(S 轮廓， S 合 )>0.9, the task isolating switch is in the "closed" state; if Iou(S 轮廓， S 分 )>0.9, the task isolating switch is in the "open" state; otherwise, it is determined that the isolating switch is in the "intermediate state".

7. The method according to claim 6, wherein the method further comprises: The mask preprocessing is to use the function cv2.approxPolyDP to perform polygon approximation on the contour, filter the contour with an area less than a threshold of 30, perform edge smoothing processing, and then extract the contour.

8. A semantic segmentation based horizontal telescopic disconnector detection system, characterized in that, Comprise A data set module configured to construct a horizontal telescopic disconnector data set according to horizontal telescopic disconnector videos in a substation; A model construction module configured to construct a semantic segmentation model, train the semantic segmentation model using the horizontal telescopic disconnector data set, and select a semantic segmentation model with the highest mIou index as an inference model; A label division module configured to label a disconnector-open polygonal region and a disconnector-closed polygonal region in a disconnector-open image and a disconnector-closed image, respectively, to obtain disconnector-open / closed region pictures; A calculation module configured to perform affine transformation on the disconnector-open / closed region pictures and a to-be-detected picture, respectively, to calculate a disconnector-open / closed region of the to-be-detected picture; A mask calculation module configured to identify the to-be-detected picture using the inference model and calculate a knife arm mask; A calculation and judgment module configured to generate a contour according to the knife arm mask, calculate an intersection-over-union of the contour and the disconnector-open / closed region of the to-be-detected picture, and determine a horizontal telescopic disconnector state by the intersection-over-union.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method of 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 method of any one of claims 1 to 7.

Images