A transmission line channel hidden danger identification post-processing method

Abstract

The present application relates to a kind of transmission line passage hidden danger identification post-processing method, comprising: obtaining transmission line passage picture;According to the tile height and width of pre-set, from top to bottom cutting transmission line passage picture, filter out invalid tile;Tile is input semantic segmentation model, obtain mask chart;Eliminate background information in mask chart, respectively for each class of hidden danger in mask chart create a binary graph, and extract the contour of object in binary graph;The contour of surface object is profiled, and the local coordinates of contour are converted into global pixel coordinates;Merge the same surface object;Calculate the barycentric coordinates of point object, convert the barycentric coordinates into global pixel coordinates;Density clustering is carried out to point object;Surface object generates polygon object, point object generates point object, respectively polygon object and point object generate shp file.The present application directly gives semantic segmentation model identification to the tile cut with array format, effectively improves processing speed, reduces the transmission loss of computer on IO.

Description

TECHNICAL FIELD

[0001] The application relates to a transmission line channel hidden danger identification post-processing method and belongs to the technical field of transmission line channel hidden danger processing. BACKGROUND

[0002] As an important means of overhead transmission line channel inspection, a UAV has been widely applied in the power industry. The data generated by the UAV overhead transmission line channel inspection is very large, and a single line often needs to take thousands of super high-definition images. After years of vigorous promotion and application of the UAV inspection, a large amount of data has been accumulated. At present, the UAV overhead transmission line channel inspection image analysis still relies on manual work, which is large in workload, low in efficiency, and difficult to associate with specific geographic locations, and therefore an intelligent processing method needs to be adopted.

[0003] Hidden danger identification is of great significance to the operation of overhead transmission lines. In tree barrier analysis, the situation of the fast-growing forest hidden danger intensive section can guide the reasonable arrangement of tree barrier distance measurement operation planning. In typhoon-prone areas, manual inspection is often needed before the typhoon season to find floating object hidden dangers, and a large-scale treatment of floating objects is needed before the disaster.

[0004] Chinese patent CN112380945B discloses a transmission channel environment hidden danger identification method and system based on satellite remote sensing images. The method uses a particle swarm algorithm and a genetic algorithm in a cooperative manner to train the parameters of a support vector machine classifier, improves the accuracy and efficiency of the support vector machine classifier in identifying environmental hidden dangers, and promotes the construction of a system for using high-resolution satellite remote sensing images to conduct three-dimensional sky-ground inspection of transmission lines. The method only displays the identification results of a single tile and cannot batch manage hidden danger information, track historical data, or automatically evaluate hidden danger results.

[0005] Chinese patent CN111339882A discloses a transmission line hidden danger detection method based on instance segmentation. The method uses visual feature information of hidden danger target faults, enhances fault area features based on image processing means through prior knowledge, fully combines deep learning and traditional image processing technology, and makes the model have strong robustness for hidden danger target fault detection. The method performs instance segmentation on a single image and only gives a pixel coordinate box of the detection result.

[0006] In the prior art, only image classification or target detection is used, and only one image is processed, which cannot completely extract hidden danger surface objects that run through multiple images, but only gives the classification result or the boundary box of the target of each tile. SUMMARY

[0007] In order to overcome the above problems, the application provides a transmission line channel hidden danger identification post-processing method, which is aimed at the transmission channel image taken by the unmanned aerial vehicle, and the cut tiles are directly given to the semantic segmentation model for identification in the array format, so that the processing speed is effectively improved, and the transmission loss of the computer on the IO is reduced. The semantic segmentation model can classify each pixel, and in the case that the hidden danger object is usually intercepted by multiple tiles, the latitude and longitude profile of the entire surface object can be completely extracted.

[0008] The technical scheme of the application is as follows:

[0009] A transmission line channel hidden danger identification post-processing method, comprising the following steps:

[0010] Obtaining a transmission line channel picture, wherein the transmission line channel picture is a tif picture;

[0011] Cutting the transmission line channel picture from top to bottom according to the preset tile height and width, inputting the entire row of grids into a grid filter, and filtering out invalid tiles;

[0012] Inputting the tiles into a semantic segmentation model, and obtaining a mask picture through the semantic segmentation model;

[0013] Removing the background information in the mask picture, creating a binary picture for each identified object in the mask picture, and extracting the profile of the object in the binary picture; the value of each pixel in the mask picture represents a hidden danger type; the object includes a surface object and a point object;

[0014] Removing the surface object with an area less than a threshold T;

[0015] Performing profile fitting on the profile of the surface object, leaving an edge profile, taking the edge profile as the profile of the surface object, and converting the local coordinates of the profile into global pixel coordinates;

[0016] Merging the same surface objects, traversing the surface objects according to the categories, and merging the surface object profiles with overlapping global pixel coordinates through a profile merger;

[0017] Calculating the center of gravity coordinates of the point object, and converting the center of gravity coordinates into global pixel coordinates;

[0018] Performing density clustering on the point object;

[0019] Generating a polygon object from the surface object and a point object from the point object, and generating a shp file from the polygon object and the point object respectively.

[0020] Further, it further comprises:

[0021] Obtaining the width and height of the transmission line channel picture;

[0022] If the width and height of the transmission line corridor image are not integer multiples of the tile width and height, then the left and bottom sides of the transmission line corridor image will be padded with black images to make them integer multiples of the tile width and height.

[0023] Furthermore, the step of inputting the entire row of grid cells into the grid filter to filter out invalid tiles specifically involves:

[0024] Input the entire row of raster cells into the raster filter, and the raster filter will then... The height is read as a pixel image with a height of 1 and a width of w; tilehight is the height of the tile, and w is the width of the transmission line channel image;

[0025] The pixel image is cut according to the width of the tile; tilewidth is the width of the tile.

[0026] Determine whether the pixel image contains pixels; if so, output the tile corresponding to the pixel image.

[0027] Furthermore, a binary image is created for each type of identified object in the mask image, specifically as follows:

[0028] Iterate through the values ​​in the mask image; each value except 255 represents a different object to be identified.

[0029] The values ​​other than 255 are extracted and assigned the value 0, and the remaining part is assigned the value 255, resulting in a binary image of 255 identified objects.

[0030] Furthermore, the outline of the surface object is fitted, leaving the edge outline. The fitting parameters adaptively change with the area of ​​the surface object. The formula for the adaptive fitting parameters is as follows:

[0031] epsilong = e * ln(area);

[0032] Where area is the pixel area and epsilong is the fitting parameter.

[0033] The formula for converting the local coordinates of a contour to global pixel coordinates is as follows:

[0034]

[0035] Among them, overlap rate The overlap ratio of the tiles is given by , where img_column and img_row are the row and column numbers of the tile, respectively, x_local and y_local are the local coordinates, and x_global and y_global are the global pixel coordinates.

[0036] Further, before merging the same kind of face objects, each face object needs to be traversed to ensure that there is no intersection between two sides in each face object, specifically:

[0037] Converting the face object into a geometry object;

[0038] Merging the geometry object and the starting coordinates, and eliminating geometry objects that cannot be merged.

[0039] Further, the point objects are density clustered, specifically:

[0040] The cluster center of each type of point object is extracted using the DBScan clustering algorithm, and the risk level coefficient is calculated by weighting the number of hidden danger points and the distance from the cluster center. The top one-third is high risk, the middle one-third is medium risk, and the last one-third is low risk.

[0041] The risk level coefficient Λ weighting algorithm is as follows:

[0042]

[0043] Where n is the number of hidden danger points, and distance is the distance between the hidden danger point and the cluster center.

[0044] Further, the cluster center of each type of point object is extracted using the DBScan clustering algorithm, specifically:

[0045] DBScan searches for clusters by checking the Eps neighborhood of each hidden danger point in the data set. If the Eps neighborhood of point p contains more than MinPts hidden danger points, a cluster is created with hidden danger point p as the core object.

[0046] DBScan iteratively aggregates objects that are directly density reachable from the core object, and merges density reachable clusters.

[0047] When no new hidden danger point is added to any cluster, the extraction of the cluster center of each type of object is ended.

[0048] Further, the face object is generated as a polygon object, and the point object is generated as a point object. The polygon object and the point object are respectively generated in the shp file. The formula for converting pixel coordinates to projection coordinates is as follows:

[0049]

[0050] Where X0, Y0 are, X1, Y1 are pixel coordinates, X2, Y2 are projection coordinates, S is the pixel height of the non-overlapping part of the cutout, crop_h is the height of the cutout, and overlap_rate is the overlap rate.

[0051] The present application has the following beneficial effects:

[0052] 1. The method is aimed at power transmission channel images taken by unmanned aerial vehicles. The cut tiles are directly given to the semantic segmentation model in array format for identification, effectively improving the processing speed and reducing the transmission loss of the computer on the IO.

[0053] 2. The semantic segmentation model can classify each pixel. In the case where the hidden danger object is usually intercepted by multiple tiles, the latitude and longitude profile of the entire surface object can be completely extracted.

[0054] 3. The method uses ROI density clustering. If there are multiple grave areas or construction sites in the identified area, the density clustering will calculate the clustering center of gravity of each type of point object, and evaluate the high, medium and low risk levels of the hidden danger points according to the number and distance. The inspection personnel prefer to handle areas with high risk levels. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The present application is a method flowchart. DETAILED DESCRIPTION

[0056] The present application will be described in detail below in combination with the drawings and specific embodiments.

[0057] Embodiment one

[0058] A power transmission line channel hidden danger identification post-processing method, comprising the following steps:

[0059] Obtaining a power transmission line channel picture, wherein the power transmission line channel picture is a tif picture;

[0060] According to the preset tile height and width, the power transmission line channel picture is cut from top to bottom, and the entire row of grids is input into a grid filter to filter out invalid tiles;

[0061] The tiles are input into a semantic segmentation model, and a mask picture is obtained through the semantic segmentation model;

[0062] The background information in the mask picture is removed, a binary picture is created for each type of identified object in the mask picture, and the contour of the object in the binary picture is extracted; the value of each pixel in the mask picture represents a hidden danger type; the object includes a surface object and a point object; each pixel value is a type of identified object, and one hidden danger type can correspond to multiple pixel values, i.e., multiple identified objects;

[0063] The surface object with an area smaller than a threshold T is removed;

[0064] In an embodiment of the present application, the contour area function is calculated using the Green formula to calculate the area inside the contour, and the formula is as follows:

[0065]

[0066] The contour of the surface object is subjected to contour fitting, leaving an edge contour, taking the edge contour as the contour of the surface object, and converting the local coordinates of the contour into global pixel coordinates;

[0067] The same surface objects are merged, the surface objects are traversed according to categories, and the surface object contours with coinciding global pixel coordinates are coincided through a contour merger;

[0068] The barycentric coordinates of the point object are calculated, and the barycentric coordinates are converted into global pixel coordinates;

[0069] In an embodiment of the present application, the barycentric coordinates of the point object are calculated according to the following formula:

[0070]

[0071] The moments of an image mainly represent the geometric features of an image region, and are also called geometric moments. Since the moments have the invariance of rotation, translation and scale, they are also called invariant moments. The first-order moments m01 and m10 are the moments of an image about the x-axis and the y-axis respectively, and can be used to determine the centroid of a target region.

[0072] The point objects are subjected to density clustering;

[0073] The surface objects are generated into polygon objects, and the point objects are generated into point objects, and the polygon objects and the point objects are respectively generated into shp files.

[0074] The semantic segmentation model used in the power line hidden danger identification post-processing is prior art, and is not first proposed in the present application.

[0075] In an embodiment of the present application, it further comprises:

[0076] The width and height of the power line channel picture are acquired;

[0077] If the width and height of the power line channel picture are not integer multiples of the tile width and height, the left side and the bottom of the power line channel picture are completed with black pictures to integer multiples of the tile width and height.

[0078] The power line channel picture is usually not an integer multiple of the tile height and width, and cannot be completely cut, so it needs to be completed.

[0079] Embodiment two

[0080] The power transmission line channel hidden danger identification post-processing method, on the basis of embodiment one, the whole row grid is input into the grid filter, and invalid tiles are filtered out, specifically:

[0081] The whole row grid is input into the grid filter, and the grid filter reads a pixel image with a height of 1 and a width of w from the height of tilehight is the height of the tile, and w is the width of the power transmission line channel image;

[0082] The pixel image is cut according to the width tilewidth; tilewidth is the width of the tile;

[0083] It is judged whether the pixel image contains pixels, and if so, the tile corresponding to the pixel image is output.

[0084] Since some hidden dangers may penetrate through multiple tiles, the display effect of identification is related to the edge of the tile, and the overlapping rate cutting is used to facilitate the outline merging of the same identification object.

[0085] Embodiment three

[0086] A power transmission line channel hidden danger identification post-processing method, on the basis of embodiment one, a binary image is created for each identification object in the mask image, specifically:

[0087] The values of the mask image are traversed, and each value other than 255 represents an identification object;

[0088] The values other than 255 are extracted and assigned to 0, and the remaining part is assigned to 255, to obtain 255 identification object binary images.

[0089] Embodiment four

[0090] A power transmission line channel hidden danger identification post-processing method, on the basis of embodiment three, the outline of the surface object is fitted, and the edge outline is left, and the fitting parameter changes adaptively with the area of the surface object, and the adaptive fitting parameter formula is as follows:

[0091] epsilong=e*ln(area);

[0092] Wherein, area is the pixel area, and epsilong is the fitting threshold.

[0093] Fitting the outline of the surface object helps to reduce the amount of storage information and coordinate conversion operation. In one specific embodiment of the present application, the coordinate points of the outline of the fitted surface object are reduced by more than seven-eighths.

[0094] The local coordinates are the coordinates of the surface object on the corresponding mask image.

[0095] Global pixel coordinate is the coordinate on the whole aerial photograph.

[0096] The greater the epsilong is, the more simplified the you and effect of the contour is, resulting in more loss of coordinates, and the smaller the epsilong is, the closer the you and effect of the contour is to the original contour, resulting in more retention of coordinates.

[0097] The face objects of each hazard type have different threshold values, and the following threshold values with the best screening effect are given by the present application after a large number of tests:

[0098] Greenhouse and mulch (500), color steel tile (300), garbage pile (100) and bamboo (600).

[0099] The formula for converting the local coordinates of the contour to the global pixel coordinates is as follows:

[0100]

[0101] Wherein, overlap rate is the overlap rate of the tile, img_column and img_row are the row and column numbers of the tile, x_local and y_local are the local coordinates, and x_global and y_global are the global pixel coordinates.

[0102] Example five

[0103] A transmission line channel hazard identification post-processing method, on the basis of example one, before merging the same face objects, each face object needs to be traversed to ensure that there is no intersection between two sides in each face object, specifically:

[0104] Convert the face object to a geometry object;

[0105] Merge the geometry object and the starting coordinates, and remove the geometry objects that cannot be merged.

[0106] When merging, there may be a point conflict caused by the intersection of two sides of a polygon. In order to prevent the occurrence of point conflict, it is necessary to ensure that there is no intersection between two sides in each face object.

[0107] Example six

[0108] A transmission line channel hazard identification post-processing method, on the basis of example three, the density clustering of the point object is specifically:

[0109] The DBScan clustering algorithm is used to extract the clustering center of each type of point object, and a risk level coefficient is obtained by weighting the number of hidden danger points and the distance from the clustering center, and the first third is high risk, the middle third is medium risk, and the last third is low risk.

[0110] The risk level coefficient Lambda weighting algorithm is as follows:

[0111]

[0112] Wherein, n is the number of hidden danger points, and distance is the distance between the hidden danger point and the clustering center.

[0113] In an embodiment of the present application, the DBScan clustering algorithm is used to extract the clustering center of each type of point object, specifically:

[0114] DBScan searches for clusters by checking the Eps neighborhood of each hidden danger point in the data set, and if the Eps neighborhood of point p contains more than MinPts hidden danger points, a cluster with hidden danger point p as the core object is created.

[0115] DBScan iteratively aggregates objects directly density reachable from the core object, and merges density reachable clusters.

[0116] When there is no new hidden danger point added to any cluster, the extraction of the clustering center of each type of object is ended.

[0117] Example Seven

[0118] A transmission line passage hidden danger recognition post-processing method, on the basis of example six, the polygon object is generated from the surface object, the point object is generated from the point object, and the polygon object and the point object are respectively generated in the shp file, and the formula for converting pixel coordinates to projection coordinates is as follows:

[0119]

[0120] Wherein, X0, Y0 is, X1, Y1 is pixel coordinate, X2, Y2 is projection coordinate, S is the pixel height of the non-overlapping part of the cutting map, crop_h is the height of the cutting map, and overlap_rate is the overlap rate.

[0121] The above is only an embodiment of the present application, and does not limit the patent range of the present application, and any equivalent structure made by using the content of the present application specification and drawings, or directly or indirectly used in other related technical fields, is also included in the patent protection range of the present application.

Claims

1. A method for post-identification and processing of hidden dangers in power transmission line corridors, characterized in that, Includes the following steps: Obtain images of power transmission line corridors, wherein the images of power transmission line corridors are TIFF images; Cut the image of the power transmission line channel from top to bottom according to the preset tile height and width, input the entire row of grid into the grid filter, and filter out invalid tiles; The tiles are input into a semantic segmentation model, and a mask image is obtained through the semantic segmentation model. Background information is removed from the mask image, and a binary image is created for each type of object in the mask image. The outline of the object in the binary image is then extracted. The value of each pixel in the mask image represents the type of hazard. The objects include surface objects and point objects. Remove face objects whose area is less than the threshold T; The contour of the face object is fitted, leaving the edge contour. The edge contour is used as the contour of the face object, and the local coordinates of the contour are converted into global pixel coordinates. Merge similar face objects, traverse face objects by category, and use the outline merger to make the outlines of face objects with overlapping global pixel coordinates overlap. Calculate the centroid coordinates of the point object and convert them into global pixel coordinates; Perform density clustering on point objects; Generate polygon objects from the face objects and point objects from the point objects. Generate shapefiles from the polygon objects and point objects respectively. The process of inputting the entire row of grid cells into the grid filter to filter out invalid tiles specifically involves: Input the entire row of raster cells into the raster filter, and the raster filter will then... The height is read as a pixel image with a height of 1 and a width of w; tilehight is the height of the tile, and w is the width of the transmission line channel image; The pixel image is cut according to the width of the tile; tilewidth is the width of the tile. Determine whether the pixel image contains pixels; if so, output the tile corresponding to the pixel image. The outline of the surface object is fitted, leaving the edge outline. The fitting parameters adaptively change with the area of ​​the surface object. The formula for the adaptive fitting parameters is as follows: epsilong = e * ln(area); Where area is the pixel area and epsilong is the fitting parameter; The formula for converting the local coordinates of a contour to global pixel coordinates is as follows: Among them, overlap rate The overlap ratio of the tiles is given by , where img_column and img_row are the row and column numbers of the tile, respectively, x_local and y_local are the local coordinates, and x_global and y_global are the global pixel coordinates. The face objects are generated into polygon objects, and the point objects are generated into point objects. The polygon and point objects are then generated into shapefiles. The formula for converting pixel coordinates to projected coordinates is as follows: Where X0 and Y0 are the coordinates of the upper left corner of the remote sensing image, X1 and Y1 are the pixel coordinates, X2 and Y2 are the projection coordinates, S is the pixel height of the non-overlapping part of the crop, and crop_h is the height of the crop.

2. The method for post-identification and processing of hidden dangers in transmission line corridors according to claim 1, characterized in that, Also includes: Obtain the width and height of the transmission line corridor image; If the width and height of the transmission line corridor image are not integer multiples of the tile width and height, then the left and bottom sides of the transmission line corridor image will be padded with black images to make them integer multiples of the tile width and height.

3. The method for post-identification and processing of hidden dangers in transmission line corridors according to claim 1, characterized in that, Create a binary image for each type of identified object in the mask image, specifically as follows: Iterate through the values ​​in the mask image; each value except 255 represents a different object to be identified. The values ​​other than 255 are extracted and assigned the value 0, and the remaining part is assigned the value 255, resulting in a binary image of 255 identified objects.

4. The method for post-identification and processing of hidden dangers in transmission line corridors according to claim 1, characterized in that, Before merging similar face objects, it is necessary to traverse each face object to ensure that there are no cases where two sides of each face object intersect. Specifically: Convert a face object to a geometry object; Merge the geometry objects and starting coordinates, and discard geometry objects that cannot be merged.

5. The method for post-identification and processing of hidden dangers in transmission line corridors according to claim 3, characterized in that, The density clustering of the point objects is specifically as follows: The DBScan clustering algorithm is used to extract the cluster centers of each type of point object, and the risk level coefficient is calculated by weighting the number of potential hazards and their distance from the cluster centers. The first third is considered high risk, the middle third is considered medium risk, and the last third is considered low risk. The weighted algorithm for the risk level coefficient Λ is as follows: Where n is the number of potential hazards, and distance is the distance between a potential hazard and the cluster center.

6. The method for post-identification and processing of hidden dangers in transmission line corridors according to claim 5, characterized in that, The DBScan clustering algorithm is used to extract the cluster centers of each class of point objects. Specifically: DBScan searches for clusters by examining the Eps neighborhood of each potential hazard point in the dataset. If the Eps neighborhood of point p contains more than MinPts potential hazard points, a cluster is created with potential hazard point p as the core object. DBScan iteratively aggregates objects that are directly density-reachable from the core object and merges density-reachable clusters; The extraction of cluster centers for each type of object ends when no new potential vulnerabilities are added to any cluster.

Images