A power transmission line defect identification method and system based on multi-frame checking

Abstract

The application discloses a kind of power transmission line defect identification method and system based on multi-frame check, and the application relates to defect identification technical field, the method obtains continuous multiple frames of power transmission line images, carries out wire segment locking processing to continuous multiple frames of images, obtains the same visible wire segment and its interframe corresponding relationship between adjacent frames;Determine the wire center line in each frame according to the interframe corresponding relationship, generate wire arc length normalization coordinates, establish image coordinate mapping relationship and generate wire local motion quantity, form wire with body check reference;Defect preliminary screening processing is carried out to continuous multiple frames of images, obtain suspected defect area and generate suspected defect with body position sequence;Then based on the continuity of cross-frame position and motion homology is carried out joint gate confirmation, and outputs defect identification result;Can improve the stability and accuracy of defect identification under the condition of wire swing, view angle change and complex background, reduce false abnormality false detection rate, improve defect confirmation credibility.

Description

Technical Field

[0001] This invention relates to the field of defect identification technology, specifically to a method and system for identifying transmission line defects based on multi-frame verification. Background Technology

[0002] As transmission line inspections evolve from manual inspections to drone inspections, online video inspections, and intelligent image analysis, automatic identification technologies for attachment-type defects such as conductor strand breakage, wear and stripping, and foreign matter attachment have gradually evolved from single-frame threshold detection, edge detection, and texture analysis to image recognition methods that combine target detection, semantic segmentation, temporal tracking, and multi-frame verification.

[0003] In the existing identification process, abnormal areas are usually detected first in the image coordinate system, and then it is determined whether they are defects by simple cross-frame overlay or target tracking. This type of process does not establish a body reference relationship around the same physical conductor. When the conductor swings, bends, the drone's perspective changes, or the background structure interferes, reflections, shadows, bird shadows, background cables and tower material edges are easily confused with real defects. Furthermore, it is difficult to continuously determine whether the same abnormality is always attached to the same conductor position and whether its movement is consistent with the local movement of the conductor. Therefore, there are problems such as unstable cross-frame correspondence, a large number of false detections, and insufficient confirmation reliability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for identifying transmission line defects based on multi-frame verification, which solves the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution, comprising the following steps: S1. Acquire multiple consecutive frames of images of the transmission line, perform conductor segment locking processing on the multiple consecutive frames of images of the transmission line, and obtain the same visible conductor segment and the inter-frame correspondence between the same visible conductor segment in adjacent frames. S2. Determine the center line of the same visible conductor segment in each frame image according to the inter-frame correspondence, generate the conductor arc length normalized coordinates along the conductor center line, establish the image coordinate mapping relationship of the conductor arc length normalized coordinates in consecutive multi-frame images, generate the local motion of the conductor according to the image coordinate mapping relationship, and associate the conductor arc length normalized coordinates, the image coordinate mapping relationship and the local motion of the conductor as the conductor body verification benchmark. S3. Perform initial defect screening on multiple consecutive frames of images of the transmission line to obtain suspected defect areas, and map the suspected defect areas to the conductor body verification benchmark to generate a suspected defect body position sequence; the suspected defect body position sequence includes the arc length normalized position, normal offset and attachment range of the suspected defect area in multiple consecutive frames. S4. Based on the cross-frame position continuity in the suspected defect satellite position sequence, generate position stability gating results, and generate suspected defect motion based on the position change of the suspected defect region in consecutive multi-frame images. Generate motion homology gating results based on the suspected defect motion and the local motion of the conductor. S5. Perform dual-gating fusion processing based on the position stability gating result and the motion homology gating result, and output the defect identification result.

[0006] Preferably, in S1, the inspection video stream of the same inspection section of the transmission line is acquired by the acquisition terminal, and N consecutive frames of images are extracted from the inspection video stream according to the preset time window and preset frame interval, and a frame sequence identifier is configured for each frame of image to obtain multiple consecutive frames of images of the transmission line, where N is an integer greater than or equal to 3. For each frame of a series of images of a transmission line, candidate regions for conductors are extracted to obtain candidate regions for conductors in each frame. Centerline fitting is performed on the candidate regions of conductors in each frame of the image to obtain the candidate centerlines of the conductors. Match the candidate centerlines of the conductors in adjacent frames, and determine the candidate conductor regions that have the same direction, corresponding end positions, consistent relative order along the line, and continuous coverage area as continuously matched candidate conductor regions. Merge consecutively matched conductor candidate regions to obtain the same visible conductor segment, and establish inter-frame correspondence based on the start and end points and relative positions along the conductor candidate centerline in adjacent frames for the same visible conductor segment.

[0007] Preferably, in S2, the start and end points of the same visible conductor segment are determined in each frame image according to the inter-frame correspondence. In each frame of the image, the center line of the conductor is obtained by fitting the conductor region between the start and end ends. The direction of the conductor's centerline from the starting end to the ending end is determined as the arc length calculation direction; Determine multiple centerline positions along the arc length calculation direction, calculate the distance along the line from each centerline position to the starting end, and calculate the ratio of the distance along the line to the total length of the conductor centerline to obtain the normalized coordinates of the conductor arc length corresponding to each centerline position; Generate a sequence of arc length normalized coordinate points on the conductor arc length normalized coordinate according to a preset coordinate interval; Extract the image coordinates corresponding to each arc length normalized coordinate point in the arc length normalized coordinate point sequence in each frame image, and associate the image coordinates corresponding to the same arc length normalized coordinate points according to the frame order to obtain the image coordinate mapping relationship.

[0008] Preferably, in S2, the image coordinates of the same arc length normalized coordinate point in adjacent frame images are obtained based on the image coordinate mapping relationship; The coordinate displacement vector is obtained by performing differential processing on the image coordinates of the same arc length normalized coordinate point in adjacent frame images; Motion decomposition of the coordinate displacement vector yields the local motion direction, local motion amplitude, and local motion velocity of the conductor. Among them, the direction of local motion of the conductor corresponds to the direction of the coordinate displacement vector, the amplitude of local motion of the conductor corresponds to the length of the coordinate displacement vector, and the speed of local motion of the conductor corresponds to the ratio of the amplitude of local motion of the conductor to the time interval between adjacent frames. The local motion direction, local motion amplitude, and local motion velocity of the conductor are combined to form the local motion quantity of the conductor. Using the arc length normalized coordinate point as an index, the arc length normalized coordinate point, the corresponding image coordinate mapping relationship, and the local motion of the conductor corresponding to the arc length normalized coordinate point are associated to obtain the conductor body verification benchmark.

[0009] Preferably, in S3, a preliminary defect screening process is performed on multiple consecutive frames of images of the transmission line to obtain suspected defect areas. The preliminary defect screening process divides the same visible conductor segment into multiple arc length normalization detection windows based on the conductor arc length normalization coordinates in the conductor body verification benchmark. Based on the image coordinate mapping relationship, the conductor neighborhood image block corresponding to each arc length normalization detection window is extracted from multiple consecutive frames of images of the transmission line; Based on the local motion of the conductor, the neighboring image blocks of the conductor corresponding to the same arc length normalized detection window are aligned with the body to obtain a body-aligned image block group. Normal contour response of conductor is generated in the body-aligned image block group, and edge break response, local protrusion response, texture abrupt response and grayscale abrupt response are calculated based on the normal contour response of conductor to obtain defect candidate response map. The pixel regions whose response values ​​in the defect candidate response map reach the initial screening threshold are connected to obtain the initial screening connected regions. The initially screened connected regions that simultaneously satisfy the conductor adjacency condition, area condition, length along the conductor condition, and response intensity condition are identified as suspected defect regions. Among them, the conductor adjacency condition is that the shortest distance from the initial screening connected domain to the conductor centerline is not greater than the adjacency distance threshold; the area condition is that the area of ​​the initial screening connected domain is within the preset area range; the line length condition is that the length of the initial screening connected domain along the conductor centerline reaches the line length threshold; and the response intensity condition is that the average response value in the initial screening connected domain reaches the response intensity threshold.

[0010] Preferably, in S3, after obtaining the suspected defect area, the region center point, the region boundary point set, and the arc length normalized detection window to which the suspected defect area belongs are determined. Project the center point of the region along the normal direction of the conductor centerline to the conductor centerline to obtain the center projection point, and read the arc length normalized coordinates of the conductor corresponding to the center projection point from the conductor body calibration datum to obtain the arc length normalized position. Calculate the directed normal distance between the center point of the region and the center projection point to obtain the normal offset; Project the boundary point set of the region onto the center line of the traverse to obtain the boundary projection point set, and read the normalized coordinates of the traverse arc length corresponding to the boundary projection point set from the traverse body verification datum. The minimum arc length normalized coordinate of the conductor in the boundary projection point set is taken as the starting arc length normalized position, the maximum arc length normalized coordinate of the conductor in the boundary projection point set is taken as the ending arc length normalized position, and the interval between the starting arc length normalized position and the ending arc length normalized position is taken as the attachment range. According to the frame sequence identifier, suspected defect areas that meet the following conditions are associated and determined as the same suspected defect target: the difference of arc length normalization position is not greater than the preset arc length association threshold, the difference of normal offset is not greater than the preset normal association threshold, and the overlap ratio of attachment range is not less than the preset range association threshold. The arc length normalization position, normal offset and attachment range corresponding to the same suspected defect target are then arranged in time sequence to generate a suspected defect body position sequence.

[0011] Preferably, in S4, the arc length normalized position, normal offset and attachment range corresponding to M consecutive frames are read from the suspected defect satellite position sequence, where M is an integer greater than or equal to 3; Calculate the arc length normalized position difference between adjacent frames, and take the maximum value among multiple arc length normalized position differences as the arc length position change; Calculate the difference in normal offset between adjacent frames, and take the maximum value among multiple normal offset differences as the change in normal offset; Calculate the attachment range overlap ratio between adjacent frames, and take the minimum value among multiple attachment range overlap ratios as the attachment range retention amount; The changes in arc length position, normal offset, and attachment range retention are normalized to obtain stable components of arc length, normal offset, and range. The position stability gating value is obtained by weighted fusion of the arc length stable component, the normal stable component, and the range stable component. When the position stability gating value reaches the position stability threshold, a position stability gating result that satisfies the position stability condition is generated; When the position stability gating value does not reach the position stability threshold, a position stability gating result that does not meet the position stability condition is generated.

[0012] Preferably, in S4, the suspected defect regions corresponding to consecutive M frames are read from the suspected defect satellite position sequence, and the region center point of the suspected defect region in each frame is obtained. The displacement of the center point between the center points of the regions in adjacent frames is calculated according to the frame order to obtain the suspected defect displacement sequence. The suspected defect motion quantity is generated based on the suspected defect displacement sequence. The suspected defect motion quantity includes the suspected defect motion direction, suspected defect motion amplitude, and suspected defect motion velocity. Read the arc length normalized position corresponding to each frame from the suspected defect satellite position sequence, and interpolate the local motion direction, local motion amplitude and local motion velocity of the conductor based on the adjacent arc length normalized coordinate points corresponding to the arc length normalized position from the conductor satellite verification benchmark to obtain the local motion quantity of the conductor corresponding to the arc length normalized position. The local motion quantity of the conductor includes the local motion direction, local motion amplitude and local motion velocity of the conductor. Calculate the angle between the direction of movement of the suspected defect and the direction of local movement of the conductor, and obtain the directional homogeneous components based on the angle. Calculate the relative difference in amplitude between the suspected defect motion amplitude and the local motion amplitude of the conductor, and obtain the common amplitude components based on the relative difference in amplitude; Calculate the relative velocity difference between the suspected defect's movement velocity and the local movement velocity of the conductor, and obtain the velocity homogeneous components based on the relative velocity difference; The relative motion residual is calculated based on the direction homogeneous component, the amplitude homogeneous component, and the velocity homogeneous component, and the motion homogeneous gating value is generated based on the relative motion residual; When the motion homology gate value reaches the motion homology threshold, a motion homology gate result that satisfies the motion homology condition is generated. When the motion homology gating value does not reach the motion homology threshold, a motion homology gating result that does not meet the motion homology condition is generated.

[0013] Preferably, in S5, the position stability gating value in the position stability gating result is read, and the motion homogeneous gating value in the motion homogeneous gating result is read; Dual-gated fusion processing obtains a dual-gated weighted value by weighting and fusing position-stable gating values ​​and motion-homogeneous gating values. The minimum value between the position stability gate value and the motion homogeneity gate value is read to obtain the dual-gated constraint value; The body homology confirmation value is obtained by weighting the dual-gated weighted value, dual-gated constraint value, confirmation fusion coefficient and constraint fusion coefficient. The sum of the confirmation fusion coefficient and the constraint fusion coefficient is 1. The dual-gated constraint value is the minimum value between the position stability gating value and the motion homology gating value. The homogeneity confirmation value of the satellite is compared with the defect confirmation threshold and the defect exclusion threshold, and dual-gated fusion processing is performed, wherein the defect confirmation threshold is greater than the defect exclusion threshold. When the homology confirmation value of the satellite reaches the defect confirmation threshold, the defect confirmation result is output. When the homology confirmation value is lower than the defect elimination threshold, the defect elimination result is output. When the satellite homology confirmation value reaches the defect elimination threshold but is lower than the defect confirmation threshold, the result to be reviewed is output. The defect identification results include the same visible conductor segment corresponding to the suspected defect area, arc length normalization position, normal offset, attachment range, position stability gate value, motion homology gate value, and body homology confirmation value.

[0014] A transmission line defect identification system based on multi-frame verification includes an image locking module, a body reference module, a preliminary screening and positioning module, a dual gating module, and a confirmation output module.

[0015] This invention provides a method and system for identifying transmission line defects based on multi-frame verification. It has the following advantages: (1) This method continuously locks the same visible conductor segment and establishes a conductor body verification benchmark formed by the conductor arc length normalized coordinate, image coordinate mapping relationship and conductor local motion quantity association, so that the conductor position, conductor shape and conductor local motion in multiple consecutive frames can be uniformly expressed under the same body reference, thereby solving the problem of easy drift of defect position and unstable cross-frame correspondence under the conditions of conductor swing, viewpoint change and background interference in the prior art.

[0016] (2) This method converts the suspected defect area into a sequence of suspected defect body positions containing arc length normalization position, normal offset and attachment range, and performs joint judgment on the suspected defect based on position stability gating result and motion homology gating result. This enables the system to not only determine whether the abnormal area is continuously attached to the conductor, but also to determine whether the abnormal area is consistent with the local movement of the conductor, thereby solving the problem that false anomalies such as reflection, shadow, bird shadow and background cable are difficult to effectively eliminate in the prior art.

[0017] (3) This method combines the conductor body verification benchmark with position stability gating and motion homogeneity gating, so that the defect identification process has both conductor body positioning capability and multi-frame homogeneity confirmation capability, thereby solving the problem of many false detections and low confirmation reliability caused by simply relying on single-frame anomaly detection or simple multi-frame superposition in the prior art. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2This is a schematic diagram of the module of the present invention; Figure 3 This is a flowchart of the multi-frame wire detection and matching process of the present invention; Figure 4 This is a schematic diagram of the dual-gated fusion confirmation and result output of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1

[0021] Please see Figure 1 , Figure 3 and Figure 4 This invention provides a method for identifying transmission line defects based on multi-frame verification, comprising the following steps: S1. Acquire multiple consecutive frames of images of the transmission line, perform conductor segment locking processing on the multiple consecutive frames of images of the transmission line, and obtain the same visible conductor segment and the inter-frame correspondence between the same visible conductor segment in adjacent frames. S2. Determine the center line of the same visible conductor segment in each frame image according to the inter-frame correspondence, generate the conductor arc length normalized coordinates along the conductor center line, establish the image coordinate mapping relationship of the conductor arc length normalized coordinates in consecutive multi-frame images, generate the local motion of the conductor according to the image coordinate mapping relationship, and associate the conductor arc length normalized coordinates, the image coordinate mapping relationship and the local motion of the conductor as the conductor body verification benchmark. S3. Perform initial defect screening on multiple consecutive frames of images of the transmission line to obtain suspected defect areas, and map the suspected defect areas to the conductor body verification benchmark to generate a suspected defect body position sequence; the suspected defect body position sequence includes the arc length normalized position, normal offset and attachment range of the suspected defect area in multiple consecutive frames. S4. Based on the cross-frame position continuity in the suspected defect satellite position sequence, generate position stability gating results, and generate suspected defect motion based on the position change of the suspected defect region in consecutive multi-frame images. Generate motion homology gating results based on the suspected defect motion and the local motion of the conductor. S5. Perform dual-gating fusion processing based on the position stability gating result and the motion homology gating result, and output the defect identification result.

[0022] In this embodiment, the video defect recognition scenario of UAV inspection of power transmission lines is applied. Specifically, it is applicable to the identification of attachment-type defects of conductors under conditions of wind deflection, slight galloping, local reflection, background structure interference, and slight changes in viewing angle. Attachment-type defects include strand breakage, wear and peeling, and attached foreign objects.

[0023] In this embodiment, the inspection data comes from the continuous video of the inspection section of the transmission line by the airborne visible light inspection device. The video resolution is 3840×2160 and the sampling frame rate is 25 frames / second. In each defect identification process, 7 consecutive frames of images are extracted from the inspection video stream as a group of continuous multi-frame images of the transmission line.

[0024] It should be noted that the video resolution, sampling frame rate, number of frames, threshold, and statistical quantity involved in this embodiment are only used to illustrate the implementation process of the present invention and do not constitute a limitation on the scope of protection of the present invention. When the inspection equipment, shooting distance, image resolution, or wire background type changes, the defect confirmation threshold, defect elimination threshold, position stability threshold, and motion homology threshold can be recalibrated by manually annotating samples.

[0025] In S1, conductor segment locking processing is performed on multiple consecutive frames of images of the transmission line. By extracting the candidate regions of conductors in each frame, fitting the center line, and matching adjacent frames, the same visible conductor segment of the same physical conductor in multiple consecutive frames is determined, and the inter-frame correspondence of the same visible conductor segment between adjacent frames is established.

[0026] In S2, the center line of the same visible conductor segment in each frame image is determined according to the inter-frame correspondence, and the arc length normalized coordinate of the conductor is generated along the center line. Then, the image coordinates of the same arc length normalized coordinate in different frames are associated to form an image coordinate mapping relationship. Based on the changes in image coordinates between adjacent frames, the local motion of the conductor is generated, thereby obtaining the conductor body verification reference.

[0027] In S3, a preliminary defect screening process is performed on multiple consecutive frames of images of the transmission line to obtain suspected defect areas. The suspected defect areas are then mapped to the conductor body verification reference to obtain the arc length normalized position, normal offset, and attachment range of the suspected defect areas in multiple consecutive frames, thereby generating a sequence of suspected defect body positions.

[0028] In S4, based on the suspected defect satellite position sequence, it is determined whether the suspected defect region maintains a stable attachment position in multiple consecutive frames, and a position stability gating result is generated. At the same time, the displacement vector of the center point of the suspected defect is generated using adjacent frame pairs as the basic comparison unit. The displacement vector of the center point of the suspected defect in each adjacent frame pair is compared with the local motion of the conductor at the corresponding position to generate a motion homogeneous gating result.

[0029] In S5, dual-gated fusion processing is performed based on position stability gating results and motion homology gating results. When the position of the suspected defect area relative to the conductor remains stable and its motion is consistent with the local motion of the conductor, the defect confirmation result is output. When the position stability or motion homology of the suspected defect area does not meet the requirements, the defect exclusion result or the result to be reviewed is output.

[0030] Through the above steps, the present invention first locks the same visible conductor segment, then establishes a conductor body verification benchmark, and then converts the suspected defect area into the conductor body coordinate system to judge the position stability and motion homology. This can reduce false detections caused by false anomalies such as reflections, shadows, bird shadows, background cables and tower material edges, and improve the stability and reliability of the identification of attachment-type defects in transmission lines.

[0031] Example 2

[0032] Please refer to Figure 1 , Figure 3 and Figure 4 Specifically: In S1, the inspection video stream of the same inspection section of the transmission line is acquired through the acquisition terminal. N consecutive frames of images are extracted from the inspection video stream according to the preset time window and preset frame interval, and a frame sequence identifier is configured for each frame image to obtain multiple consecutive frames of images of the transmission line, where N is an integer greater than or equal to 3. For each frame of a series of images of a transmission line, candidate regions for conductors are extracted to obtain candidate regions for conductors in each frame. Centerline fitting is performed on the candidate regions of conductors in each frame of the image to obtain the candidate centerlines of the conductors. Match the candidate centerlines of the conductors in adjacent frames, and determine the candidate conductor regions that have the same direction, corresponding end positions, consistent relative order along the line, and continuous coverage area as continuously matched candidate conductor regions. Merge consecutively matched conductor candidate regions to obtain the same visible conductor segment, and establish inter-frame correspondence based on the start and end points and relative positions along the conductor candidate centerline in adjacent frames for the same visible conductor segment.

[0033] In this embodiment, the application scenario is consistent with that of Embodiment 1, both being applied to a video defect identification scenario where a drone performs close-range visible light inspection of the same inspection section of a power transmission line. Specifically, the continuous image processing process is described under conditions of wind deflection, slight galloping, local reflection, background structure interference, and slight changes in viewing angle. This embodiment is used to illustrate the specific implementation of S1. The basic conditions of the inspection video are consistent with those of Embodiment 1, namely, the inspection video resolution is 3840×2160, the sampling frame rate is 25 frames / second, the inspection height is 18 to 26 meters, the shooting distance is 12 to 21 meters, and in each processing task, 7 consecutive frames are extracted from the inspection video stream according to adjacent frames as a group of continuous multi-frame images of the power transmission line.

[0034] It should be noted that the preset time window, preset frame interval, filtering threshold, matching threshold, and structural parameters involved in this embodiment are only used to more clearly illustrate the implementation process and implementation logic of the corresponding technical solution in this embodiment, and should not be construed as limiting the scope of protection of this invention.

[0035] In this embodiment, after acquiring the inspection video stream of the same inspection section of the transmission line through the acquisition terminal, a preset time window of 0.24 seconds is first used, and a preset frame interval of 1 adjacent frame is used to extract 7 consecutive frames of images from the inspection video stream. The frame sequence identifiers F1, F2, F3, F4, F5, F6 and F7 are configured in sequence to obtain a continuous multi-frame image of the transmission line. Since the conductor usually appears as a thin, narrow, and linear target that is sensitive to changes in local brightness in the UAV inspection image, before extracting the candidate region of the conductor, brightness normalization processing and bilateral filtering processing are performed on each frame of the image to suppress compression noise and local bright reflection. Then, the image is converted into a grayscale image and local contrast enhancement is performed to make the grayscale difference between the conductor and the sky background, mountain background and tower background more obvious, thereby providing a stable input for the subsequent extraction of the candidate region of the conductor.

[0036] When extracting candidate regions for conductors in each frame of a series of images of a transmission line, linear enhancement processing is first performed on the image in multiple directions. Specifically, the fine line response of pixels is calculated in the directions of 0°, 15°, 30°, 45°, 60°, 75°, 90°, 105°, 120°, 135°, 150°, and 165°. Linear features of conductors are extracted at three width scales of 3 pixels, 5 pixels, and 7 pixels. The maximum response value at each direction and scale is then used as the comprehensive linear response value of the current pixel. Next, using a 31×31 pixel local neighborhood as a statistical window, the average response intensity within the neighborhood of each pixel is calculated. Pixels whose comprehensive linear response value is higher than the local average response value and higher than 1.35 times the median response value of the entire image are selected. The pixels are identified as foreground pixels. Then, a morphological closing operation of 9 pixels in length is performed on the binary region formed by the foreground pixels along the local principal direction to connect the small breaks caused by reflection, compression damage or local occlusion. Subsequently, the foreground region after the closing operation is labeled with connected components. For each connected component, the region area, minimum bounding rectangle aspect ratio, principal axis continuous length, principal direction variation standard deviation and the overlap ratio with the tower material edge template are calculated in sequence. Only connected components with an area between 80 pixels and 4200 pixels, a minimum bounding rectangle aspect ratio of not less than 8, a principal axis continuous length of not less than 4% of the short side of the image, a principal direction variation standard deviation of not more than 12 degrees and an overlap ratio with the tower material edge template of not more than 0.18 are retained as candidate conductor regions. Thus, the candidate conductor regions in each frame image are obtained.

[0037] After obtaining the candidate wire regions in each frame of the image, centerline fitting is performed on each candidate wire region. Specifically, each candidate wire region is first thinned to a skeleton path of single-pixel width. Then, spurious branches in the skeleton path are removed. Specifically, short branches with a length less than 12 pixels and an angle greater than 25 degrees with the main path are identified as noise branches and deleted. After obtaining the cleaned skeleton path, pixels with a connectivity of 1 in the skeleton path are counted as endpoint candidates. When there are two endpoint candidates, these two endpoint candidates are directly determined as the start and end points of the candidate wire centerline. When there are more than two endpoint candidates, the centerline is selected along the skeleton path. The pair of candidate endpoints with the longest path length measured by the path are used as the start and end points. When there are fewer than two candidate endpoints, the endpoints are determined according to the extreme values ​​of the projection along the main direction of the skeleton path. After the endpoints are determined, an ordered skeleton point sequence is extracted along the skeleton path with one endpoint as the start point and the other endpoint as the end point. 21 sampling points are extracted from the ordered skeleton point sequence according to the principle of equal arc length. The 21 sampling points are then fitted with a cubic B-spline at least squares to obtain a smooth and continuous candidate centerline for the conductor. To facilitate subsequent cross-frame matching, the overall direction angle, total curve length, start point position, end point position, and the positions of 9 equal arc length sampling points for each candidate centerline for the conductor are also recorded.

[0038] It should be noted that when performing centerline fitting on the candidate regions of conductors in each frame of the image, the input to the least squares fitting is the sequence of center-fitting sampling point coordinates extracted from a single candidate region and arranged in path order, as well as the sequence of sampling parameters corresponding one-to-one with the center-fitting sampling point coordinate sequence. The center-fitting sampling point coordinate sequence is obtained as follows: first, the candidate region of the conductor is represented as a binary region mask; then, the binary region mask is thinned to obtain a main skeleton path with a single pixel width; subsequently, short branches with lengths less than a preset length threshold and angles greater than a preset angle threshold with the main skeleton path are deleted to obtain a purified main skeleton path; then, taking one endpoint of the main skeleton path as the starting point and the other endpoint as the ending point, the arc length is accumulated along the path direction of the main skeleton path, and multiple center-fitting sampling points are selected sequentially from the main skeleton path at equal arc length intervals to form a sequence of center-fitting sampling point coordinates arranged in path order; based on this, a cubic B-spline curve is used as the fitting model. The least-squares fitting method aims to minimize the sum of squared Euclidean distances between the curve points corresponding to each sampling parameter position of the cubic B-spline curve and the coordinates of the center fitting sampling point. The optimal fitting result is obtained by adjusting the coordinates of the control points of the cubic B-spline curve. The output of the least-squares fitting includes the coordinates of the control points of the cubic B-spline curve, the continuous spline curve determined by the control point coordinates, and the coordinates of the starting end, ending end, total curve length, overall direction angle, and multiple equal-arc length sampling points calculated based on the continuous spline curve. The continuous spline curve is used as the candidate centerline of the conductor. The coordinates of the starting and ending ends are used to determine the correspondence of the end positions in adjacent frames, the overall direction angle is used to determine the consistency of the direction, and the coordinates of multiple equal-arc length sampling points are used to determine the consistency of the relative order along the line. This allows the least-squares fitting to not only complete the geometric transformation from the candidate conductor region to the candidate centerline of the conductor, but also provide unified, continuous, and comparable input data for matching the candidate centerlines of the conductor in subsequent adjacent frames.

[0039] When matching candidate centerlines of conductors in adjacent frames, each candidate centerline of conductor in frame Fi is used as a reference to construct candidate matching pairs with all candidate centerlines of conductors in frame F(i+1). Four conditions are then sequentially determined: consistent direction, corresponding end positions, consistent relative order along the line, and continuous coverage area. Consistent direction is determined by comparing the angle difference between the overall direction angles of two candidate centerlines; a difference of no more than 5 degrees satisfies the consistency condition. Corresponding end positions are determined by calculating the sum of endpoint distances for "start to start and end to end" and "start to end and end to start," respectively, and selecting the pair with the smaller sum as the end-to-end correspondence method. If the average distance between the two pairs of corresponding endpoints under this correspondence method is no more than 16 pixels and the distance between any two corresponding endpoints is no more than 22 pixels, the end-to-end correspondence condition is satisfied. Consistent relative order along the line is determined by sampling the nine equal arc lengths on the first candidate centerline. The sample points are projected onto the center line of the second candidate guide line one by one, and the normalized position along the line of the projection point is recorded. When the nine normalized positions along the line remain monotonically increasing, and the average difference between the normalized positions along the line of the corresponding sample points is not greater than 0.08, the condition of consistent relative order along the line is considered to be met. The continuous coverage area is calculated by expanding the two candidate guide areas outward by 4 pixels and calculating their mask overlap ratio. When the overlap ratio is not less than 0.70 and the length ratio of the two is between 0.80 and 1.25, the condition of continuous coverage area is considered to be met. For candidate matching pairs that meet the above four conditions, scores are assigned according to the four indicators of consistent direction, corresponding end positions, consistent relative order along the line, and continuous coverage area. The scores are then comprehensively evaluated with weights of 0.25, 0.30, 0.25, and 0.20. Only the candidate guide areas with a comprehensive matching score of not less than 0.78 and which are both bidirectional best matches are determined as continuously matched candidate guide areas.

[0040] After obtaining consecutively matched candidate conductor regions, these regions are merged to obtain the same visible conductor segment. Specifically, the bidirectional best matching relationship determined in frames F1 to F7 is used as connecting edges to connect the candidate conductor regions in each frame into a matching chain arranged in chronological order. When a matching chain continuously covers 3 or more frames, and any two adjacent frames within the chain satisfy the aforementioned four matching conditions, all candidate conductor regions in the matching chain are considered as the imaging result of the same physical conductor in consecutive frames, and the matching chain is merged and defined as the same visible conductor. To ensure the consistency of the direction of the correspondence between subsequent frames, a unified centerline direction is first determined for the same visible conductor segment in the first frame. When the overall candidate centerline of the conductor is closer to the horizontal direction, the end with the smaller horizontal coordinate is defined as the start end and the other end is defined as the end end. When the overall candidate centerline of the conductor is closer to the vertical direction, the end with the smaller vertical coordinate is defined as the start end and the other end is defined as the end end. In subsequent frames, the direction is kept consistent according to the proximity relationship with the start and end ends of the previous frame, thereby avoiding the reversal of the fitting direction of the candidate centerline of the conductor.

[0041] When establishing inter-frame correspondences based on the start, end, and relative positions along the candidate centerlines of the same visible conductor segment in adjacent frames, the already established start and end correspondences between adjacent frames are first retained. Then, using the total length of the candidate centerlines in each frame as a reference, the corresponding centerline points are read at 11 normalized along-line positions: 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0. The centerline points at the same normalized along-line positions in adjacent frames are then... The centerline points are one-to-one correspondences, thus forming an inter-frame correspondence between adjacent frames consisting of the start end, the end end, and multiple relative position points along the line. The inter-frame correspondence established in the above manner can not only describe the correspondence between the ends of the candidate centerline of the conductor in adjacent frames, but also describe the continuous mapping relationship between multiple positions along the middle of the conductor in adjacent frames. Therefore, it can provide a stable, continuous, and directly callable pre-input for the unified modeling of the conductor centerline, the generation of the conductor arc length normalized coordinates, and the construction of the conductor body verification benchmark in the subsequent embodiment 1.

[0042] When the start or end points of adjacent frames cannot be stably matched due to partial occlusion, the entry or exit of the conductor from the screen, or the obstruction of the end by tower materials, the same visible conductor segment is not directly discarded. Instead, multiple equally visible sampling points of equal arc length are used as matching inputs to calculate the average distance of the sampling points along the line, the consistency of the sampling point order, and the consistency of the local tangent direction between adjacent frames. When the number of visible sampling points is not less than 5 and the comprehensive matching score is not less than 0.78, the candidate matching pair is determined as a locally visible matching pair, and the inter-frame correspondence is supplemented based on the locally visible matching pair.

[0043] Example 3

[0044] Please refer to Figure 1 , Figure 3 and Figure 4 Specifically: In S2, based on the inter-frame correspondence, the start and end points of the same visible conductor segment are determined in each frame image; In each frame of the image, the center line of the conductor is obtained by fitting the conductor region between the start and end ends. The direction of the conductor's centerline from the starting end to the ending end is determined as the arc length calculation direction; Determine multiple centerline positions along the arc length calculation direction, calculate the distance along the line from each centerline position to the starting end, and calculate the ratio of the distance along the line to the total length of the conductor centerline to obtain the normalized coordinates of the conductor arc length corresponding to each centerline position; Generate a sequence of arc length normalized coordinate points on the conductor arc length normalized coordinate according to a preset coordinate interval; Extract the image coordinates corresponding to each arc length normalized coordinate point in the arc length normalized coordinate point sequence in each frame image, and associate the image coordinates corresponding to the same arc length normalized coordinate points according to the frame order to obtain the image coordinate mapping relationship.

[0045] In S2, based on the image coordinate mapping relationship, the image coordinates of the same arc length normalized coordinate point in adjacent frame images are obtained; The coordinate displacement vector is obtained by performing differential processing on the image coordinates of the same arc length normalized coordinate point in adjacent frame images; Motion decomposition of the coordinate displacement vector yields the local motion direction, local motion amplitude, and local motion velocity of the conductor. Among them, the direction of local motion of the conductor corresponds to the direction of the coordinate displacement vector, the amplitude of local motion of the conductor corresponds to the length of the coordinate displacement vector, and the speed of local motion of the conductor corresponds to the ratio of the amplitude of local motion of the conductor to the time interval between adjacent frames. The local motion direction, local motion amplitude, and local motion velocity of the conductor are combined to form the local motion quantity of the conductor. Using the arc length normalized coordinate point as an index, the arc length normalized coordinate point, the corresponding image coordinate mapping relationship, and the local motion of the conductor corresponding to the arc length normalized coordinate point are associated to obtain the conductor body verification benchmark.

[0046] In this embodiment, the application scenario is consistent with that of Embodiment 1. Both are applied to a video defect identification scenario where a drone conducts close-range visible light inspection of the same inspection section of a power transmission line. The basic conditions of the inspection video are consistent with those of Embodiments 1 and 2, namely, the inspection video resolution is 3840×2160, the sampling frame rate is 25 frames / second, the inspection height is 18 to 26 meters, the shooting distance is 12 to 21 meters, and in each processing task, 7 consecutive frames are extracted from the inspection video stream according to adjacent frames as a group of continuous multi-frame images of the power transmission line.

[0047] When determining the start and end points of the same visible conductor segment in each frame image based on the inter-frame correspondence, the inter-frame correspondence established in Example 2 is read first. The inter-frame correspondence includes the start and end point correspondences of the same visible conductor segment in adjacent frames, as well as the correspondences of multiple relative position points along the line. Specifically, in the first frame F1, the front end point of the candidate center line of the conductor, which has been unified in direction in Example 2, is directly determined as the start end, and the rear end point is directly determined as the end end. Then, in the second frame F2, the endpoints corresponding to the start end and the endpoints corresponding to the end end of the first frame F1 are read respectively, and the former is determined as the start end of the second frame F2, and the latter is determined as the end end of the second frame F2. In the same way, in the third frame F3 to the fourth frame F2... In the seven frames F7, the start and end points of the previous frame are used as tracking references in turn. Using the endpoint correspondence established in Example 2, the endpoint corresponding to the start point of the previous frame is determined as the start point of the current frame, and the endpoint corresponding to the end point of the previous frame is determined as the end point of the current frame. When two endpoints in a frame are close to the endpoints of the previous frame, the corresponding order of three relative positions along the line near the two endpoints is further compared. Only the endpoints that are close in distance and have a consistent relative order along the line are determined as the start or end point of the current frame. This ensures that the direction of the same visible conductor segment remains consistent in all seven frames of images, and the start and end points are not interchanged due to local reflection, slight occlusion, or reversal of the centerline fitting direction.

[0048] In each frame of the image, when fitting the conductor centerline based on the conductor region between the start and end points, the specific processing procedure is as follows: When extracting the local conductor region defined by the start point, end point, and the boundary of the same visible conductor segment in the current frame, first read the current frame conductor candidate region mask that has been determined as the same visible conductor segment in Example 2, and determine the line connecting the start and end points as the line reference axis; then determine the normal reference axis with the perpendicular direction of the line reference axis, thereby establishing the line-normal local coordinate system of the local conductor region in the current frame; subsequently, for each foreground pixel in the conductor candidate region mask, calculate the projection position of the foreground pixel on the line reference axis and the projection distance on the normal reference axis, wherein the line projection position is used to characterize the foreground image. The location of the pixel between the start and end points is determined by the normal projection distance, which characterizes the lateral deviation of the foreground pixel relative to the line connecting the start and end points. Then, in the conductor candidate region mask, only foreground pixels whose projection position is between the start and end points and whose normal projection distance is between the minimum and maximum normal projection distances corresponding to the boundary of the current frame conductor candidate region are retained. A morphological closing operation with a radius of 2 pixels is performed on the retained foreground pixels to fill the local boundary gap, thereby obtaining a local conductor region jointly defined by the start point, the end point, and the boundary of the same visible conductor segment. The minimum and maximum normal projection distances are determined by the extreme values ​​of the projection of the boundary point set of the current frame conductor candidate region on the normal reference axis.

[0049] The local conductor region is represented as a binary region mask. Then, the binary region mask is thinned to shrink the conductor region into a skeleton path of single-pixel width. Next, the skeleton path is filtered, removing short branches with a length less than 12 pixels and an angle greater than 25 degrees with the main skeleton direction, retaining only the main skeleton path connecting the start and end points. Then, using the start point as the first path point and the end point as the last path point, all path pixels are extracted along the main skeleton path in the order of path connectivity, resulting in an ordered skeleton point sequence arranged sequentially by path. This ordered skeleton point sequence is then resampled using equal arc lengths. Specifically, the Euclidean distance between adjacent skeleton points is accumulated to obtain the cumulative distance along the line starting from the start point. The total length is then divided by 100 as the sampling step size. At the cumulative distance corresponding to sampling steps 0, 1, 2, up to 100, the sampling steps are further refined. The process involves finding the two closest skeleton points in the ordered skeleton point sequence and interpolating their x and y coordinates according to the distance ratio to obtain 101 centerline sampling points arranged in path order. Finally, using the coordinate sequence of these 101 centerline sampling points as input, a cubic B-spline curve is used as the fitting model. The objective is to minimize the sum of the squared Euclidean distances between the curve points at the corresponding sampling parameter positions and the 101 centerline sampling points. This process yields the control point coordinates of the cubic B-spline curve, which are then used to generate a continuous spline curve, which is then used as the centerline of the traverse. The input to the fitting is the coordinates of the 101 centerline sampling points and their corresponding equally spaced sampling parameter sequence. The output is the control point coordinates, the continuous spline curve, and the starting and ending positions and the continuous geometric trajectory of the entire curve, all read from the continuous spline curve.

[0050] After determining the direction of the conductor centerline from the starting end to the ending end as the arc length calculation direction, multiple centerline positions on the conductor centerline are determined along the arc length calculation direction. The distance along the line from each centerline position to the starting end is calculated, and the ratio of the distance along the line to the total length of the conductor centerline is calculated to obtain the normalized coordinates of the conductor arc length corresponding to each centerline position. The specific processing process is as follows: First, the fitted conductor centerline is discretized into sufficiently dense curve points along the direction from the starting end to the ending end. Sufficiently dense means that the interval between adjacent curve points in the image is no more than 1 pixel. Then, starting from the starting end, the Euclidean distance between each pair of adjacent curve points is accumulated sequentially to obtain the distance along the line from each curve point to the starting end. At the same time, the accumulated distance corresponding to the ending end is determined as the total length of the conductor centerline of that frame. Afterwards, the distance along the line from each curve point to the starting end is divided by the total length of the conductor centerline of that frame to obtain the arc length normalized coordinates of that frame. The arc length normalized coordinates of the conductor corresponding to the curve points are such that the arc length normalized coordinate of the starting end is 0, the arc length normalized coordinate of the ending end is 1, and the arc length normalized coordinate of any intermediate curve point is the proportion of its position along the line to the total length. In this embodiment, in order to ensure the uniformity of subsequent cross-frame mapping, not all discrete curve points are directly used as indices, but an arc length normalized coordinate point sequence is generated on the conductor arc length normalized coordinates according to a preset coordinate interval. Specifically, with 0.01 as the preset coordinate interval, 0.00, 0.01, 0.02, 0.03 up to 1.00 are generated sequentially from 0, for a total of 101 arc length normalized coordinate points. The function of this arc length normalized coordinate point sequence is to uniformly map the center lines of conductors with slightly different lengths and slightly swaying shapes in different frames to the same set of standard along-line positions, so that subsequent image coordinate extraction and local motion calculation can be performed around the same index.

[0051] The image coordinates corresponding to each arc length normalized coordinate point in the arc length normalized coordinate point sequence are extracted from each frame image. The image coordinates corresponding to the same arc length normalized coordinate points are associated according to the frame order to obtain the image coordinate mapping relationship. The specific processing procedure is as follows: For any arc length normalized coordinate point in the current frame, first multiply the arc length normalized coordinate point by the total length of the current frame's conductor centerline to obtain the target distance along the line corresponding to the arc length normalized coordinate point in the current frame; then, among the discrete curve points of the conductor centerline in the current frame, find two curve points whose cumulative distance along the line is located on both sides of the target distance and is closest to the target distance along the line; if the target distance along the line is exactly the same as the cumulative distance along the line of one of the curve points, then directly read the x and y coordinates of that curve point as the image coordinates of the arc length normalized coordinate point in the current frame. If the target's distance along the line lies between two consecutive curve points, then linear interpolation is performed on the x and y coordinates of the two consecutive curve points according to the ratio of the target's distance along the line to the cumulative distance between the two consecutive curve points, to obtain the image coordinates of the arc-length normalized coordinate point in the current frame. Using the above method, the corresponding image coordinates of the 101 arc-length normalized coordinate points in the current frame are extracted sequentially, and this process is repeated in all frames F1 to F7 to obtain the image coordinate sequence of each arc-length normalized coordinate point in F1 to F7. Subsequently, the image coordinates corresponding to the same arc-length normalized coordinate points are associated according to the frame order. Specifically, the arc-length normalized coordinate point is used as the first-level index and the frame order is used as the second-level index to establish a correspondence table of "arc-length normalized coordinate point, frame order and image coordinates", thereby forming an image coordinate mapping relationship.

[0052] First, using any arc-length normalized coordinate point as an index, read the image coordinates of that arc-length normalized coordinate point in the previous and next frames from the image coordinate mapping relationship; then, subtract the abscissa of the previous frame image coordinate from the abscissa of the next frame image coordinate to obtain the lateral displacement, and subtract the ordinate of the previous frame image coordinate from the ordinate of the next frame image coordinate to obtain the longitudinal displacement, and use the lateral displacement and longitudinal displacement together to form the coordinate displacement vector.

[0053] Next, when performing motion decomposition on the coordinate displacement vector, the length of the coordinate displacement vector is first calculated and compared with the motion micro threshold. In this embodiment, the motion micro threshold is 0.50 pixels. When the length of the coordinate displacement vector is less than the motion micro threshold, the local motion amplitude of the conductor is recorded as 0, the local motion velocity of the conductor is recorded as 0, and the local motion direction of the conductor is marked as an invalid direction. When the length of the coordinate displacement vector is greater than or equal to the motion micro threshold, the direction of the coordinate displacement vector is determined as the local motion direction of the conductor, the length of the coordinate displacement vector is determined as the local motion amplitude of the conductor, and the local motion amplitude of the conductor is divided by the time interval between adjacent frames of 0.04 seconds to obtain the local motion velocity of the conductor.

[0054] Using the above method, the local motion direction, local motion amplitude, local motion velocity, and invalid direction indicator of the conductor can be obtained for each arc length normalized coordinate point between F1 and F2, F2 and F3, F3 and F4, F4 and F5, F5 and F6, and F6 and F7, respectively. Finally, the local motion direction, local motion amplitude, local motion velocity, and invalid direction indicator of the conductor corresponding to the same arc length normalized coordinate point between each adjacent frame are combined to form the local motion quantity of the conductor corresponding to that arc length normalized coordinate point.

[0055] When using arc-length normalized coordinate points as indexes to correlate the arc-length normalized coordinate points, the corresponding image coordinate mapping relationships, and the local motion of the conductor corresponding to the arc-length normalized coordinate points to obtain the conductor body calibration benchmark, the specific processing procedure is as follows: Each arc-length normalized coordinate point is used as an independent main index entry. Under this main index entry, the image coordinates of the arc-length normalized coordinate point in F1 to F7 are stored sequentially, along with the conductor's local motion direction, local motion amplitude, local motion velocity, and invalid direction identifier between adjacent frames. When subsequent... When a step requires accessing image coordinates or local motion information at a specific location along the line, simply input the target arc length normalized position. The nearest arc length normalized coordinate point record can be directly read from the traverse body calibration benchmark, or interpolation can be performed between two adjacent arc length normalized coordinate point records to obtain the image coordinate mapping relationship and traverse local motion amount corresponding to that location along the line. Thus, the traverse body calibration benchmark not only completes the standardization of the location along the line of the same visible traverse segment in multiple consecutive frames, but also completes the unified association of image observation and local motion at the same location along the line in multiple consecutive frames.

[0056] Compared with the prior art, this embodiment does not only perform static fitting on the conductor profile, but after determining the start and end points, it further obtains the conductor centerline, unifies the arc length calculation direction, calculates the conductor arc length normalized coordinates, generates an arc length normalized coordinate point sequence, extracts the corresponding image coordinates and establishes an image coordinate mapping relationship, and then calculates the local motion of the conductor and generates a conductor body verification benchmark. In this way, the geometric position changes and local dynamic changes of the same visible conductor segment in multiple consecutive frames are unified into the same body reference system.

[0057] Example 4

[0058] Please refer to Figure 1 , Figure 3 and Figure 4 Specifically: In S3, a preliminary defect screening process is performed on multiple consecutive frames of images of the transmission line to obtain suspected defect areas. The preliminary defect screening process divides the same visible conductor segment into multiple arc length normalization detection windows based on the conductor arc length normalization coordinates in the conductor body verification benchmark. Based on the image coordinate mapping relationship, the conductor neighborhood image block corresponding to each arc length normalization detection window is extracted from multiple consecutive frames of images of the transmission line; Based on the local motion of the conductor, the neighboring image blocks of the conductor corresponding to the same arc length normalized detection window are aligned with the body to obtain a body-aligned image block group. In the body-aligned image block group, an anomaly candidate exclusion mask is established based on the basic contour band of the conductor, and a normal contour response of the conductor is generated at the pixel position that does not fall into the anomaly candidate exclusion mask. Based on the normal contour response of the conductor, the edge breakage response, local protrusion response, texture abrupt response and grayscale abrupt response are calculated to obtain the defect candidate response map; The pixel regions whose response values ​​in the defect candidate response map reach the initial screening threshold are connected to obtain the initial screening connected regions. Connectivity regions that simultaneously satisfy the conditions of wire adjacency, area, length along the line, and response intensity are identified as suspected defect regions.

[0059] In S3, after obtaining the suspected defect area, the region center point, the region boundary point set, and the arc length normalized detection window to which the suspected defect area belongs are determined. Project the center point of the region along the normal direction of the conductor centerline to the conductor centerline to obtain the center projection point, and read the arc length normalized coordinates of the conductor corresponding to the center projection point from the conductor body calibration datum to obtain the arc length normalized position. Calculate the directed normal distance between the center point of the region and the center projection point to obtain the normal offset; Project the boundary point set of the region onto the center line of the traverse to obtain the boundary projection point set, and read the normalized coordinates of the traverse arc length corresponding to the boundary projection point set from the traverse body verification datum. The minimum arc length normalized coordinate of the conductor in the boundary projection point set is taken as the starting arc length normalized position, the maximum arc length normalized coordinate of the conductor in the boundary projection point set is taken as the ending arc length normalized position, and the interval between the starting arc length normalized position and the ending arc length normalized position is taken as the attachment range. According to the frame sequence identifier, suspected defect areas that meet the following conditions are associated and determined as the same suspected defect target: the difference of arc length normalization position is not greater than the preset arc length association threshold, the difference of normal offset is not greater than the preset normal association threshold, and the overlap ratio of attachment range is not less than the preset range association threshold. The arc length normalization position, normal offset and attachment range corresponding to the same suspected defect target are then arranged in time sequence to generate a suspected defect body position sequence.

[0060] In this embodiment, the application scenario is consistent with that of Embodiments 1 to 3. All are applied to the video defect identification scenario of UAV conducting close-range visible light inspection of the same inspection section of a power transmission line. The inspection video resolution is 3840×2160, the sampling frame rate is 25 frames / second, the inspection height is 18 meters to 26 meters, and the shooting distance is 12 meters to 21 meters. In each processing task, 7 consecutive frames are extracted from the inspection video stream according to adjacent frames as a group of consecutive multi-frame images of the power transmission line. Based on the same visible conductor segment, image coordinate mapping relationship, conductor local motion amount and conductor body verification benchmark obtained in Embodiments 1 to 3, this embodiment further explains how to perform defect initial screening and generate a sequence of suspected defect body positions.

[0061] It should be noted that the window length, window sliding step size, normal expansion width, connected component filtering threshold, and association threshold involved in this embodiment are only used to more clearly illustrate the implementation process of this embodiment, and should not be construed as limiting the scope of protection of this invention.

[0062] When dividing the same visible conductor segment into multiple arc length normalization detection windows based on the conductor arc length normalization coordinates in the conductor body verification benchmark, the range of the current visible conductor segment's arc length normalization coordinates from 0 to 1 is first read. A window length of 0.06 is used as the window length of a single arc length normalization detection window, and a window sliding step size of 0.03 is used between adjacent windows. Starting from 0.00, multiple windows are generated sequentially according to the step size. Each window is defined by a starting arc length normalization coordinate and an ending arc length normalization coordinate. Specifically, the first window covers 0.00 to 0.06, the second window covers 0.03 to 0.09, the third window covers 0.06 to 0.12, and so on until the last window is still within the range of 0 to 1. This decomposes the entire visible conductor segment into multiple local and overlapping detection units along the line direction, preventing truncation when defects are located precisely at the window edge.

[0063] When extracting the conductor neighborhood image block corresponding to each arc length normalization detection window from multiple consecutive frames of transmission line images based on image coordinate mapping relationships, for any arc length normalization detection window, first read the starting arc length normalization coordinates, ending arc length normalization coordinates, and the image coordinates of the arc length normalization coordinates of the points in the window in the current frame. Then, determine the direction of the line connecting the image coordinates corresponding to the starting arc length normalization coordinates and the image coordinates corresponding to the ending arc length normalization coordinates as the line direction of the current window, and determine the perpendicular direction of this line direction as the normal direction. Then, use the image corresponding to the starting arc length normalization coordinates as the normal direction. The pixel distance between the image coordinates corresponding to the image coordinates of the image coordinates and the coordinates of the arc length normalization is taken as the line length. Based on this line length, an additional 4 pixels are added to both ends to form a complete line clipping range. At the same time, an additional 18 pixels are added to both sides of the normal direction to form a normal clipping range. This results in a local rectangular clipping region centered on the point in the window, with the long side arranged along the direction of the conductor and the short side arranged along the normal direction. Then, all pixels in the current frame image that fall within this local rectangular clipping region are clipped out to obtain the conductor neighborhood image block corresponding to the arc length normalization detection window in the current frame.

[0064] When performing body alignment on the neighboring image blocks of the conductor corresponding to the same arc length normalization detection window, the neighboring image block of the conductor in the first frame is first determined as the reference image block. Then, for the conductor neighboring image blocks in subsequent frames, the local motion direction and local motion amplitude of the conductor corresponding to the arc length normalization coordinates of the points in the window are read respectively. The local motion direction and local motion amplitude of the conductor are converted into the translation compensation amount of the points in the window in the image. Then, the image blocks of subsequent frames are translated according to the translation compensation amount to the position that coincides with the point in the window of the reference image block. After the translation is completed, the tangent direction of the conductor centerline near the point in the window in the current frame is read, and the angle between it and the corresponding tangent direction in the reference image block is calculated. This angle is used as the rotation compensation angle to perform rotation compensation on the current frame image block, so that the direction of the conductor body is consistent in multiple frame image blocks. In this way, the body alignment of the same arc length normalization detection window in multiple consecutive frames is completed, and a body-aligned image block group is obtained.

[0065] When generating a normal contour response for a conductor in a body-aligned image block group, the method no longer uses the number of edge occurrences as the sole criterion for a normal contour. Instead, it first generates a basic contour band for the conductor based on the conductor centerline, the arc length normalization detection window, and the width of the conductor candidate region. Then, it establishes an anomaly candidate exclusion mask within the basic contour band. The anomaly candidate exclusion mask includes outward expansion positions with an edge extension distance exceeding 2 pixels and a continuous length exceeding 3 pixels, positions with a continuous edge loss length exceeding 3 pixels, positions with a texture normalization difference exceeding 0.55, and positions with a grayscale normalization difference exceeding 0.55.

[0066] After establishing the anomaly candidate exclusion mask, stable edges are counted only at pixel positions in the basic contour band of the conductor that do not fall into the anomaly candidate exclusion mask. The median normal position and median local width of the unexcluded edge points are calculated along each sampling column. The median normal position and median local width of the conductor are then smoothly connected between adjacent sampling columns to form the normal contour response of the conductor. When all sampling columns along a certain line fall into the anomaly candidate exclusion mask, the smooth interpolation results of adjacent sampling columns are used first to form the normal contour of that sampling column, instead of directly including the continuously occurring abnormal edges in the normal contour response of the conductor.

[0067] When the proportion of unexcluded edge points that can be used to form a normal conductor contour response within the same arc length normalization detection window is less than 0.45, the normal conductor contour corresponding to the arc length normalization detection window in the previous arc length normalization detection window, the next arc length normalization detection window, or historical health inspection images is read as a supplementary reference. When a supplementary reference cannot be obtained, the arc length normalization detection window is marked as a low-confidence window for the normal contour, and the suspected defect area generated by the window is directly marked as a candidate area to be reviewed, so as to avoid the continuous attachment of real defects being mistakenly absorbed as normal contours.

[0068] When calculating the edge break response based on the normal contour response of the conductor, the actual conductor edge in the current frame image block is first used as the detection object. Then, the current frame image block is scanned column by column or row by row along the line direction. The actual edge of the current frame is compared with the stable edge in the normal contour response of the conductor. If a continuous gap appears at the position where the stable edge should exist continuously, and the width of the gap along the normal direction does not exceed the local width of the conductor but the length along the line reaches more than 3 pixels, then the position corresponding to the gap is marked as the edge break region. The continuous length of the gap and the gradient attenuation at the gap position are jointly mapped to the edge break response value. The longer the gap and the more obvious the gradient attenuation, the higher the edge break response value.

[0069] When calculating the local bulge response based on the normal contour response of the conductor, the outer boundary of the actual conductor region in the current frame image block is first read. Then, the stable boundary formed by the normal contour response of the conductor is used as the reference boundary. The outward expansion distance of the actual boundary of the current frame relative to the reference boundary is calculated along the normal direction. When the outward expansion distance continuously exceeds 2 pixels and is continuously distributed along the line direction, the outward expansion area is identified as the local bulge area. The magnitude of the outward expansion distance and the continuous length along the line are jointly mapped to the local bulge response value. The larger the outward expansion distance and the longer the continuous length, the higher the local bulge response value.

[0070] When calculating the texture mutation response, a 5×5 pixel local statistical unit is constructed with each candidate pixel in the current frame image block as the center. The gray-level co-occurrence matrix texture features and gradient direction histogram features are calculated in this local statistical unit. Then, a reference region is selected in the normal conductor region corresponding to the normal conductor contour response. The reference region is adjacent to the candidate pixel along the line position and does not fall into the abnormal candidate exclusion mask. The gray-level co-occurrence matrix texture features and gradient direction histogram features of the reference region are calculated in the same way. Then, the differences in contrast, energy, entropy and gradient direction distribution between the candidate region and the reference region are compared. The above differences are normalized and weighted to form the texture mutation response value. The greater the texture difference, the higher the texture mutation response value.

[0071] When a normal reference region that is adjacent to the candidate pixel along the line and does not fall into the abnormal candidate exclusion mask cannot be selected in the normal conductor region corresponding to the normal conductor contour response, the body-aligned normal region in the adjacent frame of the same arc length normalization detection window is selected as the normal reference region. When the body-aligned normal region in the adjacent frame still cannot meet the reference requirements, the normal conductor region in the previous arc length normalization detection window or the next arc length normalization detection window adjacent to the current arc length normalization detection window is selected as the normal reference region. When a normal reference region still cannot be obtained, the texture change response value and grayscale change response value corresponding to the candidate pixel are marked as invalid, and the defect candidate response value of the candidate pixel is generated only based on the edge break response and local protrusion response.

[0072] When the texture abrupt response value and the grayscale abrupt response value are marked as invalid, the defect candidate response value is calculated as 0.55 multiplied by the edge break response and 0.45 multiplied by the local bulge response. When the texture abrupt response value and the grayscale abrupt response value are not marked as invalid, the defect candidate response value is calculated as 0.30 multiplied by the edge break response, 0.25 multiplied by the local bulge response, 0.25 multiplied by the texture abrupt response value, and 0.20 multiplied by the grayscale abrupt response value.

[0073] When calculating the gray-level abrupt response, the candidate pixel neighborhood in the current frame image block is used as the detection area. In the normal guide area corresponding to the normal guide contour response, a normal reference area is selected at the same position along the line and does not fall into the abnormal candidate exclusion mask. The average gray level, gray level variance, and local contrast of the detection area and the normal reference area are calculated respectively. Then, the average gray level difference, gray level variance difference, and local contrast difference are normalized and weighted according to a preset ratio to obtain the gray-level abrupt response value. The more obvious the difference in brightness distribution between the detection area and the normal reference area, the higher the gray-level abrupt response value.

[0074] In this embodiment, the normalization process adopts a truncated linear normalization method. Specifically, when the difference value to be normalized is less than or equal to the lower limit difference value, the normalized value is determined to be 0; when the difference value to be normalized is greater than or equal to the upper limit difference value, the normalized value is determined to be 1; when the difference value to be normalized is greater than the lower limit difference value and less than the upper limit difference value, the difference between the difference value to be normalized and the lower limit difference value is divided by the difference between the upper limit difference value and the lower limit difference value to obtain the corresponding normalized value. The lower limit difference value is determined based on the median of the corresponding feature difference in the normal conductor region, and the upper limit difference value is determined based on the 90th percentile of the corresponding feature difference in the normal conductor region. When the upper limit difference value is equal to the lower limit difference value, the upper limit difference value is adjusted to the lower limit difference value plus a preset small amount.

[0075] When calculating the texture mutation response value, the differences in gray-level co-occurrence matrix contrast, gray-level co-occurrence matrix energy, gray-level co-occurrence matrix entropy, and gradient direction histogram between the candidate region and the normal reference region are calculated separately. The contrast-normalized difference, energy-normalized difference, entropy-normalized difference, and gradient direction-normalized difference are obtained by truncating linear normalization. The texture mutation response value is then calculated as the sum of 0.30 multiplied by the contrast-normalized difference, 0.20 multiplied by the energy-normalized difference, 0.25 multiplied by the entropy-normalized difference, and 0.25 multiplied by the gradient direction-normalized difference.

[0076] When calculating the gray-level abrupt change response value, the average gray-level difference, gray-level variance difference, and local contrast difference between the area to be detected and the normal reference area are calculated respectively. The average gray-level normalized difference, gray-level variance normalized difference, and local contrast normalized difference are obtained by truncating linear normalization. Then, the gray-level abrupt change response value is calculated as 0.40 multiplied by the average gray-level normalized difference, 0.30 multiplied by the gray-level variance normalized difference, and 0.30 multiplied by the local contrast normalized difference.

[0077] After obtaining the edge break response, local bulge response, texture mutation response, and grayscale mutation response, the edge break response is weighted by 0.30, the local bulge response by 0.25, the texture mutation response by 0.25, and the grayscale mutation response by 0.20, and then weighted pixel by pixel to obtain a defect candidate response map. The response value of each pixel in the defect candidate response map represents the degree to which the pixel has the combined characteristics of contour abnormality, shape abnormality, texture abnormality, and grayscale abnormality.

[0078] When performing connectivity processing on pixel regions in the defect candidate response map whose response values ​​reach the initial screening threshold, pixels with response values ​​not less than 0.62 in the defect candidate response map are first marked as foreground pixels. Then, the interconnected foreground pixels are aggregated into multiple initial screening connected regions according to the 8-neighborhood connectivity rule. Each initial screening connected region consists of a group of high-response pixels that are spatially interconnected.

[0079] When initially screening connected components as suspected defect areas, each component is sequentially evaluated for wire adjacency, area, length, and response intensity conditions. The wire adjacency condition is determined by calculating the shortest Euclidean distance from each pixel in the initially screened connected component to the center line of the current frame's wire, and taking the minimum value. If this minimum value is no greater than 6 pixels, the wire adjacency condition is satisfied. The area condition is determined by counting the total number of pixels in the initially screened connected component; if the total number of pixels is between 18 and 210 pixels, the area condition is satisfied. The length condition is determined by projecting all pixels in the initially screened connected component onto the direction of the wire's center line and calculating the distance between the farthest and nearest points after projection; if this distance is no less than 5 pixels, the length condition is satisfied. The response intensity condition is determined by calculating the average response value of all pixels in the initially screened connected component in the defect candidate response map; if this average response value is no less than 0.68, the response intensity condition is satisfied. Only initially screened connected components that simultaneously satisfy all four conditions are identified as suspected defect areas.

[0080] After obtaining the suspected defect area, first determine the region center point, the region boundary point set, and the arc length normalization detection window to which the suspected defect area belongs. The region center point is obtained by averaging the horizontal and vertical coordinates of all pixels in the suspected defect area. The region boundary point set is obtained by extracting all boundary pixels on the outer contour of the suspected defect area. The arc length normalization detection window to which the suspected defect area belongs is directly taken from the current arc length normalization detection window that generated the suspected defect area.

[0081] When projecting the center point of the region onto the center line of the conductor along the normal direction of the conductor center line to obtain the center projection point, first read the tangent direction of the conductor center line at the center position of the arc length normalization detection window in the current frame, and determine the perpendicular direction of the tangent direction as the normal direction. Then, draw a straight line from the center point of the region along the normal direction, and find the intersection point between the normal line and the conductor center line in the current frame. If there is a unique intersection point, the intersection point is determined as the center projection point. If there is no analytical intersection point due to the discrete curve expression method, find the curve point with the smallest distance from the normal line among the discrete curve points of the conductor center line, and determine the curve point as the center projection point.

[0082] When reading the arc length normalized coordinates of the conductor corresponding to the center projection point from the conductor body verification reference to obtain the arc length normalized position, first calculate the cumulative distance along the conductor center line of the current frame of the center projection point, then divide the cumulative distance along the line by the total length of the conductor center line of the current frame to obtain the arc length normalized coordinates of the conductor corresponding to the center projection point, and determine the arc length normalized coordinates of the conductor as the arc length normalized position of the suspected defect area.

[0083] When calculating the normal offset, a preset direction of the normal direction of the conductor centerline is taken as the positive direction and the opposite direction as the negative direction. The displacement vector between the center point of the region and the center projection point is projected onto the normal direction, and the resulting signed projection distance is determined as the normal offset. The normal offset is positive when the center point of the region is on the positive side and negative when the center point of the region is on the opposite side.

[0084] When projecting the set of boundary points of a region onto the center line of a conductor to obtain the set of boundary projection points, the same normal projection method as the center point of the region is used for each boundary point in the set of boundary points. The corresponding boundary projection point is obtained on the center line of the conductor. Then, the cumulative distance along the line of each boundary projection point on the center line of the conductor in the current frame is read, and the cumulative distance along the line is divided by the total length of the center line of the conductor in the current frame to obtain the arc length normalized coordinate of the conductor corresponding to the boundary projection point. This forms the set of arc length normalized coordinates corresponding to the set of boundary projection points.

[0085] When determining the attachment range, the minimum value in the arc length normalized coordinate set corresponding to the boundary projection point set is determined as the starting arc length normalized position, the maximum value in the arc length normalized coordinate set is determined as the ending arc length normalized position, and the interval between the starting arc length normalized position and the ending arc length normalized position is determined as the attachment range. In this way, the coverage range of the suspected defect area along the conductor is transformed into a standardized interval expression.

[0086] When generating the suspected defect body location sequence, candidate association pairs are constructed for suspected defect regions located within adjacent or same arc length normalized detection windows in adjacent frames according to the frame order identifier. The difference in arc length normalized position, the difference in normal offset, and the overlap ratio of the attachment range are calculated for each pair. The overlap ratio of the attachment range is obtained by calculating the ratio of the intersection length to the union length of the two attachment range intervals. When the difference in arc length normalized position is not greater than 0.018, the difference in normal offset is not greater than 4 pixels, and the overlap ratio of the attachment range is not less than 0.55, the candidate association pair is determined as the corresponding result of the same suspected defect target in adjacent frames. Then, the suspected defect regions of multiple frames that continuously meet the above association conditions are concatenated in frame order, and the arc length normalized position, normal offset, and attachment range corresponding to the same suspected defect target are arranged in time sequence to generate the suspected defect body location sequence.

[0087] When multiple suspected defect regions in the same frame simultaneously meet the association conditions with the same suspected defect target in the previous frame, an association cost value is constructed based on the arc length normalized position difference, normal offset difference, and attachment range overlap ratio. The candidate association pair with the smallest association cost value and the best bidirectional relationship is selected as the effective association pair to avoid multiple suspected defect regions being incorrectly linked as the same suspected defect target.

[0088] The above processing method serves to first complete the initial screening of defects based on local windows under the condition of body alignment, and then further convert the suspected defect region in a single frame into a structured body representation that includes arc length normalized position, normal offset and attachment range. Cross-frame association is completed by using the difference in arc length normalized position, the difference in normal offset and the overlap ratio of attachment range. Thus, the suspected defect region is no longer just an abnormal connected region in a single frame image, but is transformed into a sequence of suspected defect body positions that are continuously attached, continuously located and continuously comparable relative to the same visible conductor segment.

[0089] Example 5

[0090] Please refer to Figure 1 , Figure 3 and Figure 4 Specifically: In S4, the arc length normalized position, normal offset and attachment range corresponding to M consecutive frames are read from the suspected defect satellite position sequence, where M is an integer greater than or equal to 3; Calculate the arc length normalized position difference between adjacent frames, and take the maximum value among multiple arc length normalized position differences as the arc length position change; Calculate the difference in normal offset between adjacent frames, and take the maximum value among multiple normal offset differences as the change in normal offset; Calculate the attachment range overlap ratio between adjacent frames, and take the minimum value among multiple attachment range overlap ratios as the attachment range retention amount; The changes in arc length position, normal offset, and attachment range retention are normalized to obtain stable components of arc length, normal offset, and range. The position stability gating value is obtained by weighted fusion of the arc length stable component, the normal stable component, and the range stable component. When the position stability gating value reaches the position stability threshold, a position stability gating result that satisfies the position stability condition is generated; When the position stability gating value does not reach the position stability threshold, a position stability gating result that does not meet the position stability condition is generated; Read the suspected defect regions corresponding to M consecutive frames from the suspected defect satellite location sequence, and obtain the region center point of the suspected defect region in each frame; Using adjacent frame pairs as the basic comparison unit, the displacement vector of the suspected defect center point between the center points of the regions in adjacent frames is calculated, and the movement direction, amplitude, and speed of the suspected defect in the corresponding adjacent frame pairs are obtained based on the displacement vector of the suspected defect center point. Read the local motion direction, local motion amplitude, and local motion velocity of the conductor at the same adjacent frame pairs and the same arc length normalization position from the conductor body calibration reference. The direction homogeneous component is obtained based on the angle between the direction of movement of the suspected defect and the direction of movement of the local conductor; the amplitude homogeneous component is obtained based on the relative difference between the amplitude of movement of the suspected defect and the amplitude of movement of the local conductor; and the velocity homogeneous component is obtained based on the relative difference between the velocity of movement of the suspected defect and the velocity of movement of the local conductor. The motion homology value of adjacent frame pairs is calculated based on the direction homology component, amplitude homology component and velocity homology component. The minimum and average values ​​of the motion homology values ​​of all adjacent frame pairs in a continuous M-frame window are then fused to obtain the motion homology gating value. When the motion homology gate value reaches the motion homology threshold, a motion homology gate result that satisfies the motion homology condition is generated. When the motion homology gating value does not reach the motion homology threshold, a motion homology gating result that does not meet the motion homology condition is generated.

[0091] In this embodiment, the application scenario is consistent with that of Embodiments 1 to 4, and is applied to the video defect identification scenario of a drone conducting close-range visible light inspection of the same inspection section of a power transmission line.

[0092] In this embodiment, the position stability gating result is used to characterize the cross-frame attachment stability of the same suspected defect target in the conductor's body coordinates. The motion homology gating result is used to characterize whether the motion of the same suspected defect target in the image is consistent with the local motion of the conductor corresponding to its attachment position. The arc length position change is the maximum value of the arc length normalized position difference between adjacent frames in M ​​consecutive frames. The normal offset change is the maximum value of the normal offset difference between adjacent frames in M ​​consecutive frames. The attachment range retention is the minimum value of the overlap ratio of the attachment range between adjacent frames in M ​​consecutive frames. The arc length stable component, normal stable component, and range stable component are used to unify the stability indicators of different dimensions to a comparable scale of 0 to 1. The direction homology component, amplitude homology component, and velocity homology component are used to characterize the consistency between the suspected defect motion direction, suspected defect motion amplitude, and suspected defect motion velocity and the conductor's local motion direction, conductor's local motion amplitude, and conductor's local motion velocity, respectively. The adjacent frame motion homology value is used to comprehensively characterize the homology between the suspected defect motion amount and the conductor's local motion amount.

[0093] When reading the arc length normalized position, normal offset, and attachment range corresponding to M consecutive frames from the suspected defect satellite position sequence, all records corresponding to the same suspected defect target are first sorted according to frame order. Then, a continuous subsequence of length M is read from front to back in a sliding window manner. In this embodiment, M is taken as 5. When the number of record frames for the same suspected defect target is greater than 5, multiple consecutive 5-frame windows of frames 1 to 5, 2 to 6, and 3 to 7 are generated in sequence. When the number of record frames for the same suspected defect target is less than 5, the suspected defect target does not participate in the current position stability gating and motion homology gating calculations and is directly marked as a candidate target to be supplemented for subsequent observations, thereby ensuring that the input data participating in the gating calculations have sufficient temporal continuity.

[0094] When the same suspected defect target is marked as a candidate target to be observed in subsequent supplementary observations, the candidate target is synchronously written into the results to be reviewed, and the same visible conductor segment, the obtained arc length normalization position, the obtained normal offset, the obtained attachment range, the obtained frame sequence range, and the insufficient frame count mark are recorded; when subsequent inspection video frames continue to be input and the candidate target is supplemented and associated with no less than M frames, the sequence of suspected defect carrier positions of consecutive M frames is reread and the calculation process of position stability gating results and motion homology gating results is entered.

[0095] When calculating the arc length normalization position difference, normal offset difference, and attachment range overlap ratio between adjacent frames, for two adjacent frames within a consecutive 5-frame window, first read the arc length normalization position of the previous frame and the arc length normalization position of the next frame. Then, subtract the arc length normalization position of the previous frame from the arc length normalization position of the next frame and take the absolute value to obtain the arc length normalization position difference of the adjacent frame pair. Next, read the normal offset of the previous frame and the normal offset of the next frame. Then, subtract the normal offset of the previous frame from the normal offset of the next frame and take the absolute value to obtain the normal offset difference of the adjacent frame pair. After that, read the attachment range interval of the previous frame and the attachment range interval of the next frame respectively. First, calculate the intersection length of the two attachment range intervals, then calculate the union length of the two attachment range intervals, and divide the intersection length by the union length to obtain the attachment range overlap ratio of the adjacent frame pair. When the two attachment range intervals have no intersection, the attachment range overlap ratio is recorded as 0.

[0096] When obtaining the arc length position change, normal offset change, and attachment range retention from multiple adjacent frame pairs, the maximum value among the arc length normalized position differences of all adjacent frame pairs within a consecutive 5-frame window is determined as the arc length position change, the maximum value among the normal offset differences of all adjacent frame pairs is determined as the normal offset change, and the minimum value among the attachment range overlap ratios of all adjacent frame pairs is determined as the attachment range retention. The maximum value is used to characterize the arc length position change and normal offset change in order to record the most unstable line drift and lateral drift of the suspected defective target within a consecutive 5-frame window, and the minimum value is used to characterize the attachment range retention in order to record the most unfavorable attachment coverage retention of the suspected defective target within a consecutive 5-frame window.

[0097] When normalizing the changes in arc length position, normal offset, and adhesion range retention, the changes in arc length position are mapped to a stable arc length component according to the rule that "the smaller the change, the higher the stability." Specifically, when the changes in arc length position are no greater than 0.005, they are directly recorded as 1; when the changes in arc length position are greater than 0.030, they are directly recorded as 0; and when the changes in arc length position are between 0.005 and 0.030, they are mapped to a range between 0 and 1 in a linear decreasing manner. Similarly, the changes in normal offset are mapped to a stable normal component according to the rule that "the smaller the change, the higher the stability." When the displacement change is no more than 1.0 pixel, it is directly recorded as 1; when the normal offset change is greater than 6.0 pixels, it is directly recorded as 0; when the normal offset change is between 1.0 and 6.0 pixels, it is mapped to a range between 0 and 1 in a linearly decreasing manner. The attachment range retention amount is mapped to a range stability component according to the rule that "the higher the overlap, the higher the stability". Specifically, when the attachment range retention amount is not less than 0.90, it is directly recorded as 1; when the attachment range retention amount is less than 0.40, it is directly recorded as 0; when the attachment range retention amount is between 0.40 and 0.90, it is mapped to a range between 0 and 1 in a linearly increasing manner.

[0098] When weighted and fused to generate a position stability gating result from the arc length stable component, normal stable component, and range stable component, the arc length stable component is first multiplied by 0.35, the normal stable component by 0.30, and the range stable component by 0.35. Then, the three are added together to obtain the position stability gating value. When the position stability gating value is not less than 0.72, a position stability gating result that meets the position stability condition is generated, and this position stability gating value and the "meets the position stability condition" judgment flag are written together into the gating record of the current suspected defect target. When the position stability gating value is less than 0.72, a position stability gating result that does not meet the position stability condition is generated, and this position stability gating value and the "does not meet the position stability condition" judgment flag are written together into the gating record of the current suspected defect target.

[0099] When reading the suspected defect regions corresponding to M consecutive frames from the suspected defect location sequence and generating the suspected defect motion, firstly, the suspected defect regions corresponding to each frame within the current 5 consecutive frame window are read, and the center point of the suspected defect region in each frame is obtained. The center point is obtained by averaging the x-coordinates and y-coordinates of all pixels within the suspected defect region of the current frame. Then, multiple adjacent frame pairs within the 5 consecutive frame window are determined according to the frame order. For each adjacent frame pair, the horizontal displacement of the suspected defect is obtained by subtracting the x-coordinate of the previous frame's center point from the x-coordinate of the next frame's center point, and the vertical displacement of the suspected defect is obtained by subtracting the y-coordinate of the previous frame's center point from the x-coordinate of the next frame's center point. The horizontal and vertical displacements of the suspected defect together constitute the displacement vector of the suspected defect center point of that adjacent frame pair. When the suspected defect center point of that adjacent frame pair... When the length of the displacement vector of the defect center point is less than the motion micro threshold, the suspected defect movement direction of the adjacent frame pair is marked as an invalid direction, the suspected defect movement amplitude of the adjacent frame pair is marked as 0, and the suspected defect movement speed of the adjacent frame pair is marked as 0. When the length of the displacement vector of the suspected defect center point of the adjacent frame pair is greater than or equal to the motion micro threshold, the suspected defect movement direction of the adjacent frame pair is determined as the direction of the displacement vector of the suspected defect center point, the suspected defect movement amplitude of the adjacent frame pair is determined as the length of the displacement vector of the suspected defect center point, and the ratio of the suspected defect movement amplitude of the adjacent frame pair to the time interval between adjacent frames is determined as the suspected defect movement speed of the adjacent frame pair. Thus, the suspected defect movement amount corresponding to each adjacent frame pair is obtained, and the suspected defect lateral displacement amount and the suspected defect longitudinal displacement amount together constitute the suspected defect center point displacement vector. When the displacement vector length of the suspected defect center point is less than the motion micro threshold of 0.50 pixels, the motion direction of the suspected defect in the adjacent frame pair is marked as an invalid direction, the motion amplitude of the suspected defect is marked as 0, and the motion speed of the suspected defect is marked as 0.

[0100] When interpolating the local motion direction, amplitude, and velocity of the conductor based on the arc-length normalized coordinate points corresponding to the arc-length normalized position in the conductor body verification reference to obtain the local motion quantity of the conductor corresponding to the arc-length normalized position, the arc-length normalized position of the suspected defect target in the current adjacent frame pair is read first. Then, the previous arc-length normalized coordinate point that is not greater than the arc-length normalized position and closest to the arc-length normalized position, and the next arc-length normalized coordinate point that is not less than the arc-length normalized position and closest to the arc-length normalized position are searched in the conductor body verification reference. If the arc-length normalized position completely coincides with one of the arc-length normalized coordinate points, the arc-length normalized coordinate point in the same phase is directly read. The local motion direction, amplitude, velocity, and invalid direction marker of the conductor in the adjacent frame pair are used as the local motion quantity of the conductor corresponding to the arc length normalization position. If the arc length normalization position is located between the previous arc length normalization coordinate point and the next arc length normalization coordinate point, the position ratio of the arc length normalization position relative to the previous and next arc length normalization coordinate points is first calculated. Then, the local motion amplitude and velocity of the conductor are linearly interpolated according to the position ratio. For the local motion direction of the conductor, unit direction vector interpolation is performed when both directions are valid. When at least one of the directions is an invalid direction marker, the interpolated local motion direction of the conductor is recorded as an invalid direction marker.

[0101] When calculating the directional homogeneous components, if both the suspected defect movement direction and the local conductor movement direction are valid directions, the angle between them is calculated. When the angle is no greater than 10 degrees, the directional homogeneous component is directly recorded as 1; when the angle is greater than 60 degrees, the directional homogeneous component is directly recorded as 0; when the angle is between 10 and 60 degrees, it is mapped to 0 to 1 in a linear decreasing manner. If both the suspected defect movement direction and the local conductor movement direction are invalid direction indicators, and both movement amplitudes are less than the small movement threshold, the directional homogeneous component is recorded as 1. If only one of the movement directions is an invalid direction indicator, the directional homogeneous component is recorded as 0.

[0102] When calculating the amplitude homogeneous components, the larger of the suspected defect motion amplitude and the local motion amplitude of the conductor is first used as the normalized denominator. The normalized denominator is then compared with a small threshold value, which in this embodiment is 0.50 pixels. When the normalized denominator is less than the small threshold value and the absolute difference between the two is less than the small threshold value, the amplitude homogeneous component is recorded as 1. When the normalized denominator is less than the small threshold value and the absolute difference between the two is not less than the small threshold value, the amplitude homogeneous component is recorded as 0. When the normalized denominator is greater than or equal to the small threshold value, the absolute difference between the two is divided by the normalized denominator to obtain the amplitude relative difference. The amplitude homogeneous component is obtained according to the rule that the amplitude relative difference is not greater than 0.10 and is recorded as 1, the amplitude relative difference is greater than 0.60 and is recorded as 0, and the amplitude relative difference decreases linearly between 0.10 and 0.60.

[0103] When calculating the velocity homogeneous components, the larger of the suspected defect movement velocity and the local movement velocity of the conductor is first used as the normalized denominator. The normalized denominator is then compared with a small threshold for the velocity denominator. In this embodiment, the small threshold for the velocity denominator is 12.5 pixels / second. When the normalized denominator is less than the small threshold for the velocity denominator and the absolute difference between the two is less than the small threshold for the velocity denominator, the velocity homogeneous component is recorded as 1. When the normalized denominator is less than the small threshold for the velocity denominator and the absolute difference between the two is not less than the small threshold for the velocity denominator, the velocity homogeneous component is recorded as 0. When the normalized denominator is greater than or equal to the small threshold for the velocity denominator, the absolute difference between the two is divided by the normalized denominator to obtain the velocity relative difference. The velocity homogeneous component is obtained according to the rule that a velocity relative difference not greater than 0.12 is recorded as 1, a velocity relative difference greater than 0.65 is recorded as 0, and a velocity relative difference between 0.12 and 0.65 decreases linearly.

[0104] When calculating the relative motion residual based on the direction homogeneous component, amplitude homogeneous component, and velocity homogeneous component, and generating the motion homogeneous gating value based on the relative motion residual, the direction residual, amplitude residual, and velocity residual are calculated separately for each adjacent frame pair, and fused according to weights of 0.40, 0.30, and 0.30 to obtain the relative motion residual of the adjacent frame pair. Then, the relative motion residual of the adjacent frame pair is subtracted by 1 to obtain the motion homogeneous value of the adjacent frame pair. Within a continuous M-frame window, the minimum and average values ​​of the motion homogeneous values ​​of all adjacent frame pairs are read, and fused according to weights of 0.40 and 0.60 to obtain the motion homogeneous gating value.

[0105] When the motion homology gating value is not less than 0.70, a motion homology gating result that meets the motion homology condition is generated, and the motion homology gating value and the judgment flag of "meeting the motion homology condition" are written together into the gating record of the current suspected defective target. When the motion homology gating value is less than 0.70, a motion homology gating result that does not meet the motion homology condition is generated, and the motion homology gating value and the judgment flag of "not meeting the motion homology condition" are written together into the gating record of the current suspected defective target.

[0106] By using the above processing method, position stability gating results are first constructed using three types of body position parameters: arc length normalized position, normal offset, and attachment range. Then, motion homology gating results are constructed using the direction, amplitude, and velocity relationship between the motion of the suspected defect and the local motion of the conductor. This makes the gating analysis no longer dependent on a single position threshold or a single velocity threshold, but instead verifies the real attachment relationship between the suspected defect target and the conductor through two technical chains: spatial attachment stability and local motion homology. This provides a stable, continuous, and physically meaningful preliminary judgment result for subsequent dual-gating fusion processing.

[0107] Example 6

[0108] Please refer to Figure 1 , Figure 3 and Figure 4 Specifically: In S5, the position stability gating value is read from the position stability gating result, and the motion homogeneous gating value is read from the motion homogeneous gating result; Dual-gated fusion processing obtains a dual-gated weighted value by weighting and fusing position-stable gating values ​​and motion-homogeneous gating values. The minimum value between the position stability gate value and the motion homogeneity gate value is read to obtain the dual-gated constraint value; The body homology confirmation value is obtained by weighting the dual-gated weighted value, dual-gated constraint value, confirmation fusion coefficient and constraint fusion coefficient. The sum of the confirmation fusion coefficient and the constraint fusion coefficient is 1. The dual-gated constraint value is the minimum value between the position stability gating value and the motion homology gating value. Both the position stability gate value and the motion homology gate value are restricted to between 0 and 1. When the position stability gate value or the motion homology gate value is less than 0, it is set to 0; when the position stability gate value or the motion homology gate value is greater than 1, it is set to 1. The dual-gating weighted value is calculated by multiplying the first fusion weight by the position stability gate value and the second fusion weight by the motion homology gate value. The sum of the first fusion weight and the second fusion weight is 1. The value range of the first fusion weight is 0.35 to 0.65, and the value range of the second fusion weight is 0.35 to 0.65. In this embodiment, the first fusion weight is 0.50, and the second fusion weight is 0.50.

[0109] In this embodiment, the dual-gated constraint value is the smaller of the position stability gated value and the motion homology gated value. The body homology confirmation value is calculated by multiplying the confirmation fusion coefficient by the dual-gated weighted value and the constraint fusion coefficient by the dual-gated constraint value. The sum of the confirmation fusion coefficient and the constraint fusion coefficient is 1. The value range of the confirmation fusion coefficient is 0.40 to 0.80, and the value range of the constraint fusion coefficient is 0.20 to 0.60. In this embodiment, the confirmation fusion coefficient is 0.60 and the constraint fusion coefficient is 0.40.

[0110] By incorporating dual-gating constraint values ​​into the calculation process of the body-based homogeneous confirmation value, the body-based homogeneous confirmation value can not only reflect the overall fusion level of the position stability gating value and the motion homogeneous gating value, but also be constrained by the lower of the position stability gating value and the motion homogeneous gating value. This avoids directly outputting the defect confirmation result when one gating value is high and the other gating value is low.

[0111] The homogeneity confirmation value of the satellite is compared with the defect confirmation threshold and the defect exclusion threshold, and dual-gated fusion processing is performed, wherein the defect confirmation threshold is greater than the defect exclusion threshold. The defect confirmation threshold is set to 0.78, and the defect exclusion threshold is set to 0.50. When the satellite homology confirmation value is greater than or equal to 0.78, the defect confirmation result is output. When the satellite homology confirmation value is less than 0.50, the defect exclusion result is output. When the satellite homology confirmation value is greater than or equal to 0.50 and less than 0.78, the result to be reviewed is output.

[0112] In other implementations, the defect confirmation threshold can be adjusted between 0.75 and 0.85 based on the clarity of the inspection image, the amplitude of the wire swing, and the tolerance for missed defects, while the defect exclusion threshold can be adjusted between 0.45 and 0.60 based on the number of false anomalies and the workload of review, and the defect confirmation threshold is always kept greater than the defect exclusion threshold.

[0113] When determining the defect confirmation threshold and defect exclusion threshold, manually labeled inspection video samples are selected as threshold calibration samples. The threshold calibration samples include real defect samples and pseudo-anomaly samples. Real defect samples include at least stray strand samples, wear and peeling samples, and attached foreign matter samples. Pseudo-anomaly samples include at least reflective samples, shadow samples, bird shadow samples, background cable samples, and tower edge samples. First, the position stability gate value, motion homology gate value, and body homology confirmation value corresponding to each sample are calculated respectively. Then, with the constraint that the recall rate of real defect samples is not lower than the preset recall rate requirement and the false confirmation rate of pseudo-anomaly samples is not higher than the preset false confirmation rate requirement, the threshold combination that meets the constraint is selected from multiple candidate defect confirmation thresholds and candidate defect exclusion thresholds.

[0114] After completing the threshold calibration, record the number of threshold calibration samples, the number of real defect samples, the number of false anomaly samples, the defect confirmation threshold, the defect exclusion threshold, the recall rate of real defect samples, the false confirmation rate of false anomaly samples, and the proportion of samples to be reviewed, and store the above statistics as threshold calibration records; when the inspection equipment, shooting distance, image resolution, or wire background type changes, reread the new threshold calibration samples and update the threshold calibration records.

[0115] When the homology confirmation value of the satellite reaches the defect confirmation threshold, the defect confirmation result is output. When the homology confirmation value is lower than the defect elimination threshold, the defect elimination result is output. When the satellite homology confirmation value reaches the defect elimination threshold but is lower than the defect confirmation threshold, the result to be reviewed is output. The defect identification results include the same visible conductor segment corresponding to the suspected defect area, arc length normalization position, normal offset, attachment range, position stability gate value, motion homogeneity gate value, body homogeneity confirmation value, corresponding frame sequence range, and result type. The result types include defect confirmation result, defect exclusion result, and result to be reviewed.

[0116] When the result type is a result to be reviewed, the defect identification result also includes a reason identifier to be reviewed. The reason identifier to be reviewed includes at least one of the following: insufficient frame count identifier, low confidence of normal contour identifier, critical position stability gate value identifier, critical motion homogeneous gate value identifier, and dual gate divergence identifier.

[0117] Example 7

[0118] Please see Figure 2 A transmission line defect identification system based on multi-frame verification includes an image locking module, a body reference module, a preliminary screening and positioning module, a dual gating module, and a confirmation output module. The image locking module acquires multiple consecutive frames of images of the transmission line, performs conductor segment locking processing on the multiple consecutive frames of images of the transmission line, and obtains the same visible conductor segment and the inter-frame correspondence between the same visible conductor segment in adjacent frames; The body reference module determines the center line of the same visible conductor segment in each frame image according to the inter-frame correspondence, generates the conductor arc length normalized coordinates along the conductor center line, establishes the image coordinate mapping relationship of the conductor arc length normalized coordinates in consecutive multi-frame images, generates the local motion of the conductor according to the image coordinate mapping relationship, and associates the conductor arc length normalized coordinates, the image coordinate mapping relationship and the local motion of the conductor as the conductor body verification reference. The initial screening and positioning module performs defect screening on multiple consecutive frames of images of the transmission line to obtain suspected defect areas, and maps the suspected defect areas to the conductor body verification benchmark to generate a suspected defect body position sequence. The suspected defect body position sequence includes the arc length normalized position, normal offset and attachment range of the suspected defect area in multiple consecutive frames. The dual-gating module generates position-stable gating results based on the cross-frame position continuity in the suspected defect satellite position sequence, and generates suspected defect motion based on the position change of the suspected defect region in consecutive multi-frame images. Based on the suspected defect motion and the local motion of the conductor, it generates motion homogeneous gating results. The output module confirms that it performs dual-gating fusion processing based on position stability gating results and motion homogeneity gating results to output defect identification results.

[0119] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A method for identifying transmission line defects based on multi-frame verification, characterized in that: Includes the following steps: S1. Acquire multiple consecutive frames of images of the transmission line, perform conductor segment locking processing on the multiple consecutive frames of images of the transmission line, and obtain the same visible conductor segment and the inter-frame correspondence between the same visible conductor segment in adjacent frames. S2. Determine the center line of the same visible conductor segment in each frame image according to the inter-frame correspondence, generate the conductor arc length normalized coordinates along the conductor center line, establish the image coordinate mapping relationship of the conductor arc length normalized coordinates in consecutive multi-frame images, generate the conductor local motion amount according to the image coordinate mapping relationship, and associate the conductor arc length normalized coordinates, the image coordinate mapping relationship and the conductor local motion amount as the conductor body verification reference; S3. Perform initial defect screening on multiple consecutive frames of images of the transmission line to obtain suspected defect areas, and map the suspected defect areas to the conductor body verification reference to generate a suspected defect body position sequence; the suspected defect body position sequence includes the arc length normalized position, normal offset and attachment range of the suspected defect area in multiple consecutive frames. S4. Generate a position stability gating result based on the cross-frame position continuity in the suspected defect satellite position sequence, and generate a suspected defect motion amount based on the position change of the suspected defect region in consecutive multi-frame images. Generate a motion homology gating result based on the suspected defect motion amount and the local motion amount of the conductor. S5. Perform dual-gating fusion processing based on the position stability gating result and the motion homogeneity gating result, and output the defect identification result.

2. The method for identifying transmission line defects based on multi-frame verification according to claim 1, characterized in that: In S1, the inspection video stream of the same inspection section of the transmission line is acquired by the acquisition terminal. N consecutive frames of images are extracted from the inspection video stream according to the preset time window and preset frame interval, and a frame sequence identifier is configured for each frame image to obtain multiple consecutive frames of images of the transmission line, where N is an integer greater than or equal to 3. For each frame of a series of images of the transmission line, the candidate region of the conductor is extracted to obtain the candidate region of the conductor in each frame of the image. Centerline fitting is performed on the candidate conductor regions in each frame of the image to obtain candidate conductor centerlines; The candidate centerlines of the conductors in adjacent frames are matched, and the candidate conductor regions that have the same direction, corresponding end positions, consistent relative order along the line, and continuous coverage area are determined as continuously matched candidate conductor regions. The continuously matched candidate conductor regions are merged to obtain the same visible conductor segment, and the inter-frame correspondence is established based on the start end, end end and relative position along the conductor candidate center line in adjacent frames.

3. The method for identifying transmission line defects based on multi-frame verification according to claim 1, characterized in that: In S2, based on the inter-frame correspondence, the start and end points of the same visible conductor segment are determined in each frame image; In each frame of the image, the center line of the conductor is obtained by fitting the conductor region between the start end and the end end; The direction of the center line of the conductor from the starting end to the ending end is determined as the arc length calculation direction; Multiple centerline positions on the conductor centerline are determined along the arc length calculation direction. The distance along the line from each centerline position to the starting end is calculated, and the ratio of the distance along the line to the total length of the conductor centerline is calculated to obtain the conductor arc length normalized coordinates corresponding to each centerline position. Generate a sequence of arc length normalized coordinate points on the arc length normalized coordinate of the conductor according to a preset coordinate interval; Extract the image coordinates corresponding to each arc length normalized coordinate point in the arc length normalized coordinate point sequence in each frame image, and associate the image coordinates corresponding to the same arc length normalized coordinate points according to the frame order to obtain the image coordinate mapping relationship.

4. The method for identifying transmission line defects based on multi-frame verification according to claim 3, characterized in that: In S2, based on the image coordinate mapping relationship, the image coordinates of the same arc length normalized coordinate point in adjacent frame images are obtained; The image coordinates of the same arc-length normalized coordinate point in adjacent frame images are differentially processed to obtain the coordinate displacement vector; Motion decomposition of the coordinate displacement vector yields the local motion direction, local motion amplitude, and local motion velocity of the conductor. Wherein, the local motion direction of the conductor corresponds to the direction of the coordinate displacement vector, the local motion amplitude of the conductor corresponds to the length of the coordinate displacement vector, and the local motion speed of the conductor corresponds to the ratio of the local motion amplitude of the conductor to the time interval between adjacent frames; The local motion direction, local motion amplitude, and local motion velocity of the conductor are combined to form the local motion quantity of the conductor; Using the arc length normalized coordinate point as an index, the arc length normalized coordinate point, the image coordinate mapping relationship corresponding to the arc length normalized coordinate point, and the local motion of the conductor corresponding to the arc length normalized coordinate point are associated to obtain the conductor body verification benchmark.

5. The method for identifying transmission line defects based on multi-frame verification according to claim 1, characterized in that: In S3, a preliminary defect screening process is performed on multiple consecutive frames of images of the transmission line to obtain suspected defect areas. The preliminary defect screening process divides the same visible conductor segment into multiple arc length normalization detection windows based on the conductor arc length normalization coordinates in the conductor body verification benchmark. Based on the image coordinate mapping relationship, the conductor neighborhood image block corresponding to each arc length normalization detection window is extracted from multiple consecutive frames of images of the transmission line; Based on the local motion of the conductor, the neighboring image blocks of the conductor corresponding to the same arc length normalization detection window are aligned with the body to obtain a body-aligned image block group; In the body-aligned image block group, a normal contour response of the conductor is generated, and the edge break response, local protrusion response, texture abrupt response and grayscale abrupt response are calculated based on the normal contour response of the conductor to obtain a defect candidate response map. The pixel regions whose response values ​​in the defect candidate response map reach the initial screening threshold are connected to obtain the initial screening connected regions. The initially screened connected regions that simultaneously satisfy the conductor adjacency condition, area condition, line length condition, and response intensity condition are identified as the suspected defect regions. The wire adjacency condition is that the shortest distance from the initial screening connected region to the center line of the wire is not greater than the adjacency distance threshold; the area condition is that the area of ​​the initial screening connected region is within a preset area range; the length along the line condition is that the length of the initial screening connected region along the center line of the wire reaches the length along the line threshold; and the response intensity condition is that the average response value within the initial screening connected region reaches the response intensity threshold.

6. The method for identifying transmission line defects based on multi-frame verification according to claim 5, characterized in that: In S3, after obtaining the suspected defect region, the region center point, the region boundary point set, and the arc length normalized detection window to which the suspected defect region belongs are determined. Project the center point of the region along the normal direction of the conductor centerline to the conductor centerline to obtain the center projection point, and read the arc length normalized coordinates of the conductor corresponding to the center projection point from the conductor body calibration reference to obtain the arc length normalized position. Calculate the directed normal distance between the center point of the region and the center projection point to obtain the normal offset; Project the set of boundary points of the region onto the center line of the conductor to obtain the set of boundary projection points, and read the normalized coordinates of the arc length of the conductor corresponding to the set of boundary projection points from the conductor body verification reference. The minimum arc length normalized coordinate of the conductor in the boundary projection point set is taken as the starting arc length normalized position, the maximum arc length normalized coordinate of the conductor in the boundary projection point set is taken as the ending arc length normalized position, and the interval between the starting arc length normalized position and the ending arc length normalized position is taken as the attachment range. According to the frame sequence identifier, suspected defect areas that meet the following conditions are associated and determined as the same suspected defect target: the difference between the arc length normalization position is not greater than the preset arc length association threshold, the difference between the normal offsets is not greater than the preset normal association threshold, and the overlap ratio of the attachment range is not less than the preset range association threshold. The arc length normalization position, the normal offset, and the attachment range corresponding to the same suspected defect target are then arranged in time sequence to generate the suspected defect body position sequence.

7. The method for identifying transmission line defects based on multi-frame verification according to claim 1, characterized in that: In S4, the arc length normalized position, normal offset and attachment range corresponding to M consecutive frames are read from the suspected defect satellite position sequence, where M is an integer greater than or equal to 3; Calculate the arc length normalized position difference between adjacent frames, and take the maximum value among multiple arc length normalized position differences as the arc length position change; Calculate the difference in normal offset between adjacent frames, and take the maximum value among multiple normal offset differences as the change in normal offset; Calculate the attachment range overlap ratio between adjacent frames, and take the minimum value among multiple attachment range overlap ratios as the attachment range retention amount; The arc length position change, the normal offset change, and the attachment range retention amount are normalized respectively to obtain the arc length stable component, the normal stable component, and the range stable component. The arc length stable component, the normal stable component, and the range stable component are weighted and fused to obtain the position stability gate value; When the position stability gating value reaches the position stability threshold, a position stability gating result that satisfies the position stability condition is generated; When the position stability gating value does not reach the position stability threshold, a position stability gating result that does not meet the position stability condition is generated.

8. The method for identifying transmission line defects based on multi-frame verification according to claim 7, characterized in that: In S4, the suspected defect regions corresponding to consecutive M frames are read from the suspected defect satellite position sequence, and the region center point of the suspected defect region in each frame is obtained. The displacement of the center point between the center points of the region in adjacent frames is calculated according to the frame order to obtain the suspected defect displacement sequence. The suspected defect motion quantity is generated based on the suspected defect displacement sequence, and the suspected defect motion quantity includes the suspected defect motion direction, the suspected defect motion amplitude, and the suspected defect motion velocity. Read the arc length normalized position corresponding to each frame from the suspected defect satellite position sequence, and interpolate the local motion direction, local motion amplitude, and local motion velocity of the conductor based on the adjacent arc length normalized coordinate points corresponding to the arc length normalized position from the conductor satellite verification benchmark to obtain the local motion amount of the conductor corresponding to the arc length normalized position. The local motion amount of the conductor includes the local motion direction, local motion amplitude, and local motion velocity of the conductor. Calculate the angle between the direction of movement of the suspected defect and the direction of local movement of the conductor, and obtain the directional homogeneous components based on the angle. Calculate the relative amplitude difference between the suspected defect motion amplitude and the local motion amplitude of the conductor, and obtain the amplitude homogeneous components based on the relative amplitude difference; Calculate the relative velocity difference between the suspected defect's movement velocity and the local movement velocity of the conductor, and obtain the velocity homogeneous components based on the relative velocity difference; The relative motion residual is calculated based on the direction homogeneous component, the amplitude homogeneous component, and the velocity homogeneous component, and a motion homogeneous gating value is generated based on the relative motion residual; When the motion homology gate value reaches the motion homology threshold, a motion homology gate result that satisfies the motion homology condition is generated; When the motion homology gate value does not reach the motion homology threshold, a motion homology gate result that does not meet the motion homology condition is generated.

9. The method for identifying transmission line defects based on multi-frame verification according to claim 8, characterized in that: In S5, the position stability gating value in the position stability gating result is read, and the motion homogeneity gating value in the motion homogeneity gating result is also read; The dual-gated fusion process obtains a dual-gated weighted value by weighting and fusing the position stability gating value and the motion homogeneous gating value. The minimum value between the position stability gate value and the motion homogeneity gate value is read to obtain the dual-gated constraint value; The body homogeneity confirmation value is obtained by weighting the dual-gated weighted value, the dual-gated constraint value, the confirmation fusion coefficient, and the constraint fusion coefficient. The sum of the confirmation fusion coefficient and the constraint fusion coefficient is 1, and the dual-gated constraint value is the minimum value between the position stability gating value and the motion homogeneity gating value. The satellite homology confirmation value is compared with the defect confirmation threshold and the defect exclusion threshold, and dual-gated fusion processing is performed, wherein the defect confirmation threshold is greater than the defect exclusion threshold. When the homology confirmation value of the satellite reaches the defect confirmation threshold, the defect confirmation result is output; When the homology confirmation value is lower than the defect elimination threshold, the defect elimination result is output; When the homology confirmation value of the satellite reaches the defect exclusion threshold but is lower than the defect confirmation threshold, the result to be reviewed is output. The defect identification results include the same visible conductor segment, arc length normalization position, normal offset, attachment range, position stability gate value, motion homology gate value, and body homology confirmation value corresponding to the suspected defect area.

10. A transmission line defect identification system based on multi-frame verification, based on the transmission line defect identification method based on multi-frame verification according to any one of claims 1-9, characterized in that: It includes an image locking module, a body reference module, a preliminary screening and positioning module, a dual gating module, and a confirmation output module.

Images