Power transmission line image intelligent recognition online monitoring system

Abstract

The present application relates to a kind of transmission line image intelligent recognition online monitoring system, and the accurate identification of wire and abnormal detection are realized by processing to continuous image.Access image stream and construct the cache data containing wire track and direction information;Image is normalized and noise is suppressed in brightness, and candidate area is extracted according to track information;Sub-pixel level skeleton information of wire is extracted, and continuous structure in image is enhanced processing;Skeleton information and enhanced result are coupled and fused, and expression result containing wire track and continuity is generated;Abnormal disturbance in wire is identified based on fusion result, and key area is incrementally updated, and adaptive calculation to subsequent image is realized.The present application has the characteristics such as track guide processing, structure enhancement fusion and abnormal dynamic update, realizes the continuous identification and abnormal monitoring of transmission line.

Description

Technical Field

[0001] This invention relates to the field of power system operation monitoring technology, and in particular to an online monitoring system for intelligent image recognition of transmission lines. Background Technology

[0002] With the development of intelligent inspection of transmission lines and power Internet of Things technology, image-based online monitoring of transmission line status has become an important means to ensure the safe operation of the power grid. Currently, inspection systems typically acquire line images through camera equipment and use image processing and pattern recognition methods to extract conductors, analyze structures, and detect defects.

[0003] Existing image recognition technologies for power transmission lines still have significant limitations under complex backgrounds, varying lighting conditions, and multi-interference structures. On the one hand, traditional edge detection or threshold segmentation methods are insufficient in maintaining the continuity of the slender conductor structure, easily affected by tower structure, vegetation texture, and intersecting lines, leading to discontinuous conductor skeleton extraction or false detections, making it difficult to achieve stable trajectory representation. On the other hand, existing methods lack multi-feature collaborative fusion mechanisms, making it difficult to establish effective constraints between directional consistency and structural response, resulting in a disconnect between conductor enhancement results and skeleton information, affecting overall recognition accuracy. Furthermore, for complex conditions such as conductor breakage, offset, and local anomalies, existing technologies often employ global recalculation, which is computationally inefficient and lacks dynamic updating capabilities for key areas, failing to meet the real-time and adaptive requirements of online monitoring.

[0004] Therefore, how to provide an intelligent online monitoring system for transmission line image recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent online monitoring system for transmission line image recognition. The system utilizes image standardization processing, subpixel skeleton extraction, continuous structure enhancement, coupled fusion analysis, and abnormal disturbance recognition and incremental update techniques to describe in detail the process of identifying, representing, and monitoring transmission lines in continuous images.

[0006] An online monitoring system for intelligent image recognition of power transmission lines according to an embodiment of the present invention includes the following steps:

[0007] The trajectory caching module receives a continuous image stream to generate a sequence of timestamped image frames, and stores the trajectory of the conductor centerline, local direction, and continuity parameters of the previous moment to form a trajectory cache;

[0008] The standardization module performs illumination normalization and noise suppression on the image frame, and uses the trajectory buffer to determine the strip candidate region along the direction of the conductor, generating a standardized local image.

[0009] The subpixel skeleton extraction module performs improved Steger line extraction on a standardized local image, calculates the line response and orientation field, and filters skeleton points according to the rule that the orientation difference is less than a preset threshold and the response intensity meets a preset threshold, and outputs a subpixel skeleton set.

[0010] The continuous structure enhancement module performs B-COSFIRE strip structure response calculation on a standardized local image, and generates a set of directional responses and a structure enhancement response map based on a preset set of directions and a set of scales.

[0011] The coupling and fusion module uses the center coordinates and direction of the skeleton set as anchor points, selects the maximum response in the same direction from the set of directional responses to establish a mapping, and performs connectivity verification and chain completion according to the rules of spacing and direction consistency between adjacent skeleton segments, outputting the fusion expression result of the conductor containing trajectory, topology and continuity parameters.

[0012] The identification module calculates response attenuation, interruption gap, curvature abrupt change and lateral offset indices on the wire fusion expression result to generate defect markers, and performs local incremental updates on the strip candidate region and the two-channel calculation range of the next frame based on the defect markers and trajectory buffer.

[0013] Optionally, forming a trajectory cache includes:

[0014] Access the video stream or image stream and complete the decoding and frame extraction. Write a unified clock timestamp for each frame and generate a frame sequence number. Establish a frame index table that increments by time.

[0015] Read the centerline trajectory from the previous moment's conductor fusion expression result, resample the trajectory points according to the distance along the line to obtain a point series with equal step size, and remove duplicate points and outliers to form a normalized trajectory point series;

[0016] The tangential direction is calculated point by point for the normalized trajectory point sequence and neighborhood smoothing is performed. The tangential direction is obtained by the direction of the line connecting adjacent points, and the endpoint direction is completed by the forward or backward neighborhood to form a direction sequence.

[0017] Calculate the distance between adjacent trajectory point pairs and generate a connectivity record. The connectivity record contains the distance value and the breakpoint identifier. The breakpoint identifier is determined by whether the distance exceeds the break threshold. Write the breakpoint number and the left and right segment numbers for each breakpoint.

[0018] For each trajectory segment, a segment-level continuity parameter is written. The segment-level continuity parameter is obtained by combining the number of breakpoints within the segment, the segment length, and the average spacing. The segment boundary index generated by the next frame's strip candidate region is written at the beginning and end of the trajectory segment.

[0019] The timestamp, frame number, normalized trajectory point list, direction sequence, connectivity record, segment-level continuity parameter, breakpoint number, validity marker and version number are encapsulated into a trajectory cache unit, and a validity window is set; when the next frame arrives, if the number of points covered by the new fusion expression result reaches the minimum number of points and the segment-level continuity parameter meets the update conditions, the old unit is replaced with the new trajectory cache unit; otherwise, the old unit is kept and only the frame index table is updated.

[0020] Optionally, generating a normalized local image includes:

[0021] Read the image frame with the corresponding frame number in the current frame index table and perform grayscale processing, keeping the pixel coordinate system consistent with the trajectory buffer, to obtain the grayscale frame;

[0022] In the grayscale frame, the brightness reference is calculated using the strip candidate region of the trajectory buffer as the statistical range, and brightness offset correction and contrast scaling are performed on the pixels within the statistical range so that the corrected grayscale distribution meets the preset target range, and an illumination normalized frame is generated.

[0023] Noise suppression processing is performed on the strip candidate region in the illumination normalized frame. The noise suppression processing adopts edge-preserving filtering and performs neighborhood update on the pixels in the region using a preset window to obtain the noise-suppressed frame.

[0024] Read the normalized trajectory point list, direction sequence, breakpoint number and segment boundary index from the trajectory cache unit, generate the centerline chain according to the point list order, and expand the preset half width along its normal at each centerline point to obtain a strip slice. When the breakpoint number is encountered, the strip slice is segmented and the end shrinkage is performed at the segment boundary index.

[0025] Each strip slice is merged according to its intra-frame coordinates to generate a strip candidate mask. The candidate mask is then closed and filled to form a single connected region. At the same time, the correspondence between the mask and the center line chain is recorded.

[0026] A strip candidate mask is used to crop and retain pixels of the noise suppression frame, outputting a standardized local image. The standardized local image is then bound and stored with the frame number, timestamp, candidate mask identifier, and centerline chain identifier.

[0027] Optionally, generating a subpixel skeleton set includes:

[0028] Read the standardized local image, calculate the Gaussian derivative at a fixed smoothing scale σ to obtain I_x, I_y, I_xx, I_xy, I_yy, and construct a Hessian matrix H(x,y)=[■(I_xx (x,y)&I_xy (x,y)@I_xy (x,y)&I_yy (x,y))] at each pixel position, and construct the gradient vector ∇I(x,y)=[I_x (x,y),I_y (x,y)]^⊤;

[0029] For each pixel, the Hessian matrix is ​​decomposed to obtain eigenvalues ​​λ_1, λ_2 and unit eigenvectors v_1, v_2. v_1 is selected as the normal n and v_2 is selected as the tangent t according to |λ_1|≥|λ_2|. The candidate point record is written with |λ_1| as the line response value R.

[0030] An initial candidate point set is generated based on the line type determination criteria. The line type determination criteria include: R is not lower than the preset response threshold, |λ_2| is not higher than the preset upper limit, and the signs of λ_1 and λ_2 satisfy the preset line type consistency condition. Pixels that do not meet the line type determination criteria are not included in the candidate point set.

[0031] For each candidate point, perform sub-pixel localization along the normal direction, calculate the normal displacement Δ, and update the center coordinates to the sub-pixel center coordinates. The displacement satisfies:

[0032] Δ=-(n^⊤ ∇I) / (n^⊤ Hn);

[0033] The sub-pixel center coordinates, tangent t, and response value R are written into the sub-pixel record.

[0034] Non-maximum constraints are applied along the normal direction to eliminate duplicate responses: Based on the sub-pixel center coordinates, samples are taken on both sides of the normal direction at fixed sampling intervals. The response values ​​at the center position and the two sides are compared, and only sub-pixel points whose response values ​​at the center position are not less than the response values ​​at the two sides are recorded.

[0035] Based on centerline recalibration and interference removal of normal double boundary symbol pairs, for each sub-pixel record that passes the non-maximum constraint, a one-dimensional profile is established along the normal direction and the first derivative projection in the normal direction is calculated:

[0036] g(s) = ∇I(p+sn)⋅n;

[0037] Where p is the subpixel center coordinate and s is the normal offset; on the profile, search for points on one side where |g(s)| achieves a local maximum and has a positive sign as boundary points, and search for points on the other side where |g(s)| achieves a local maximum and has a negative sign as boundary points, to obtain the two boundary offsets s_L and s_R; when the two boundaries satisfy g(s_L)⋅g(s_R)<0 and |s_L+s_R| does not exceed the preset symmetry threshold, move the subpixel center coordinates along the normal to the midpoint position p←p+(s_L+s_R) / 2 n of the two boundaries and retain the point; when no pair of boundary points with opposite signs are found or |s_L+s_R| exceeds the preset symmetry threshold, delete the subpixel point record;

[0038] In the recalibrated subpixel records, a trajectory cache orientation consistency constraint is introduced: for each subpixel, the nearest neighbor trajectory point is matched in the trajectory cache and the trajectory orientation t_ref is read, Δθ=arccos(|t⋅t_ref |) is calculated, and only subpixel records with Δθ less than the preset orientation difference threshold are retained;

[0039] Based on the single-chain skeleton construction process of mutual best inheritance, each retained sub-pixel is used as a node to establish a candidate inheritance relationship: the forward half-plane of the node's tangent t is used as the search area, and the node with the largest response value R within the fixed search radius is selected as the forward inheritance node q_f; the reverse verification node q_b of its forward inheritance node is also selected in the forward half-plane of q_f's tangent. When the reverse verification node is the same node as the current node, mutual best inheritance is confirmed and a directed link is established; otherwise, no link is established. The mutual best inheritance determination is repeated for all nodes to obtain N non-branching links.

[0040] Output skeleton segments by link: Sort the nodes of each link according to the link order and write the skeleton segment number and the sequence number within the segment; delete the entire link whose length is less than the preset minimum number of points; summarize all nodes in the retained links to form a sub-pixel skeleton set.

[0041] Optionally, generating directional response sets and structural enhancement response maps includes:

[0042] Read standardized local images, establish a strip primitive response library within a preset scale set, perform differential Gaussian convolution on the images at each scale and perform half-wave rectification to obtain strip primitive response maps, and perform local nonmaximum suppression on the strip primitive response maps to form a peak candidate table;

[0043] For each direction θ_k in the preset direction set, a B-COSFIRE operator template is generated. The operator template sets a radial distance sequence ρ_j with the operator center as the origin and generates an angle search window for each ρ_j. The angle search window is centered on θ_k and has a limited angle span. Peak candidates that fall within the radius interval [ρ_j-δ, ρ_j+δ] and whose angles fall within the angle search window are retrieved from the peak candidate table. The peak candidate ranked first in peak intensity from high to low is selected as the support point of ρ_j, and the (ρ_j, ϕ_j) and scale index of the support point are written into the operator template. When no peak candidate is found for ρ_j, ρ_j is marked as missing and not written into the support point.

[0044] For each operator center in each direction θ_k, support point response sampling is performed. For each support point in the operator template, two sampling steps are completed on the bar primitive response map corresponding to its scale index: the first step is to establish a fixed radius neighborhood centered on the theoretical coordinates of the support point and search for the maximum response in the neighborhood as the support point sample value; the second step is to perform allowable displacement diffusion of the support point sample value in space. The diffusion adopts a fixed diffusion rule with the radial distance from the support point to the center as input, and the diffused value is written into the support point response table.

[0045] The support point responses under the same direction θ_k are combined according to the B-COSFIRE combination rule to generate directional response values. The directional response values ​​are then combined with the support point responses in the form of geometric mean. The combination expression is as follows:

[0046] R_(θ_k ) (x,y)=(∏_(i=1)^N▒ s_i (x,y))^(1 / N);

[0047] Where s_i (x,y) is the support point response of the i-th support point at the operator center (x,y), and N is the number of support points participating in the combination; the set of directional responses is obtained by repeating the calculation for all directions;

[0048] The template generation and response gating process for bidirectional pairwise locking of the main axis involves generating an additional symmetrical radial distance -ρ_j in the reverse direction of the main axis for each radial distance ρ_j, and performing the same peak retrieval and support point selection. The support points in the forward direction and the support points in the reverse direction of the main axis are written into the same pair of numbers. When a support point is missing on either side of the same pair of numbers, the pair of numbers is directly deleted from the operator template. In the direction response calculation, the smaller value of the responses of the support points on both sides of the same pair of numbers is taken as the effective response of the pair of numbers, and the effective response replaces the original responses of the support points on both sides to participate in the geometric mean combination.

[0049] The cross-structure shielding process for tangential disproving sampling involves generating a tangential disproving sampling segment at a reference position corresponding to each radial distance ρ_j. The tangential disproving sampling segment is expanded along the tangential direction of direction θ_k and has a fixed length. Peak candidates falling into the tangential disproving sampling segment are retrieved from the peak candidate table, and their maximum peak intensity is taken as the disproving intensity. When the disproving intensity of any ρ_j reaches a preset disproving threshold, the directional response value of the operator center in direction θ_k is set to zero, and the directional output is terminated.

[0050] At each pixel location, the maximum value of the directional response set is taken to form a structural enhancement response map, and the direction index of the maximum value is recorded to form an enhancement direction map. On the enhancement direction map, only pixels with the same direction index and structural enhancement response not lower than the preset enhancement threshold are connected by 8 neighborhoods to form a connected link. Connected links with a link length lower than the preset minimum link length are deleted, and the structural enhancement response map and directional response set after link filtering are output.

[0051] Optionally, the generated wire fusion expression results include:

[0052] Read the subpixel skeleton set and form an ordered skeleton segment sequence according to the skeleton segment number. At the same time, read the directional response set and the structural enhancement response map to establish a fusion processing unit with the skeleton point as the core. Extract the center coordinates and tangential direction of each skeleton point as anchor point parameters.

[0053] For each skeleton point, a directional search sector is constructed in the structural enhancement response map. The directional search sector has the skeleton point as the vertex, the tangential direction as the axis of symmetry, and a fixed angle range as the fan-shaped angle. Within the sector, sampling is performed step by step along the tangential principal axis. The pixel with the largest structural enhancement response value is selected as the coupled response pixel, and the pixel coordinates and enhancement direction index are recorded to complete the directional maximum mapping from the skeleton point to the enhancement response.

[0054] For each skeleton segment, a coupling strength sequence is constructed in sequence according to the serial number. The coupling strength sequence is obtained by combining the skeleton point and line response values ​​with the corresponding coupling response values ​​according to a fixed weight. The sequence is scanned in sequence, and segments that are continuously below the coupling threshold are marked as candidate fracture segments.

[0055] The trajectory extrapolation-enhanced connectivity verification process is performed on adjacent skeleton segments: a linear extrapolation segment of fixed length is generated at the end of each skeleton segment based on the tangential direction of the end point. Within the range of the linear extrapolation segment, it is searched to see if the starting point of another skeleton segment falls within a preset distance interval. When a matching starting point exists, the continuity of the enhanced response in the area covered by the extrapolation segment is further verified in the structural enhanced response map. Only when the enhanced response on the extrapolation path is continuous and not lower than the enhancement threshold is it determined that the two skeleton segments are connected.

[0056] For skeleton segments determined to be connectable, an enhanced guided chain-complementation process is executed: with the extrapolated segment as the main axis, sampling points are generated along the main axis direction in the structural enhancement response map at a fixed step size, and the pixel with the largest enhancement response is selected as the chain-complementation node within the limited width on both sides of the normal of each sampling point. The chain-complementation nodes are inserted sequentially between the two skeleton segments to form a fusion trajectory.

[0057] The fused trajectory is subjected to topology reconstruction, the trajectory number and segment number are renumbered, and a connectivity table is generated. At the same time, trajectory-level continuity parameters are calculated, including total trajectory length, supplementary chain length ratio and number of breaks. The trajectory point sequence, topology number and continuity parameters are encapsulated into a conductor fusion expression result for output.

[0058] Optionally, generating defect markers and performing local incremental updates includes:

[0059] Read the current frame conductor fusion expression result, obtain the fusion trajectory point sequence, skeleton segment number, supplementary chain segment mark, coupling strength sequence and enhancement direction index, and establish a continuous detection window within the segment according to the trajectory number;

[0060] Four types of disturbance candidate segments are formed on the trajectory sequence: segments that continuously decrease and whose decrease reaches the attenuation threshold are recorded as response attenuation candidates by comparing adjacent windows of the coupling strength sequence with a fixed window; segments that trigger a break are recorded as interruption gap candidates by comparing the distance between adjacent trajectory points with the break threshold, and the "repaired chain" and "unrepaired chain" states are distinguished by the supplementary chain segment; discrete curvature is calculated by three-point neighborhood and compared with the baseline curvature within the segment, and segments that exceed the abrupt change threshold and whose duration reaches the minimum length are recorded as curvature abrupt change candidates; the current trajectory is aligned to the trajectory cache trajectory by arc length, and the normal distance is calculated at the alignment point, and segments whose normal distance continuously exceeds the offset threshold are recorded as lateral offset candidates;

[0061] The defect identification process of intra-segment consensus voting uses the skeleton segment as the voting unit. It maps four types of candidate segments to the corresponding skeleton segments and counts the number of hits. Only when at least two types of candidates in the same skeleton segment hit simultaneously and one of them must be "unpaired chain interruption gap" or "lateral offset" is output as a defect segment. The output defect segment is written with the defect type set, start and end number, segment length, enhancement direction index and chain completion status to form the defect marking result.

[0062] Based on the defect marking results, the incremental candidate region for the next frame is generated: taking the minimum bounding rectangle of the defect segment in the image coordinates as the basis, the region is expanded by a fixed length margin along the trajectory tangent and by a fixed half width along the normal to obtain the defect update region; at the same time, a confidence recovery region is generated for the segment that has not triggered a defect but whose coupling strength is lower than the coupling threshold. The confidence recovery region is generated according to the same expansion rule.

[0063] The defect update region is set to high priority and written into the two-channel calculation range of the next frame, while the confidence recovery region is set to low priority and written only into the calculation range of the structure enhancement channel. When the high-priority region exists, the next frame only performs a full-process recalculation of the sub-pixel skeleton channel and continuous structure enhancement channel in the high-priority region, and updates the trajectory cache with the recalculation result. When only the low-priority region exists, the next frame keeps the skeleton channel output unchanged while updating the structure enhancement response in the low-priority region and triggering the coupling fusion verification.

[0064] The beneficial effects of this invention are:

[0065] This invention introduces an improved Steger line extraction method, which adds normal dual-boundary constraints and direction consistency screening mechanisms on the basis of sub-pixel positioning. This enables accurate correction of the conductor centerline and elimination of pseudo-line responses, effectively overcoming the problem of traditional line extraction methods being easily affected by unilateral edges, structural noise and texture interference, and improving the accuracy and stability of skeleton extraction.

[0066] This invention performs B-COSFIRE strip structure response calculations and combines bidirectional pairwise constraints and tangential disproving suppression mechanisms along the principal axis to ensure that the structure enhancement process is triggered only when bidirectional consistent strip features are present. At the same time, it effectively suppresses cross structures and non-target textures, solving the problem of insufficient ability to distinguish complex backgrounds in existing structure enhancement methods, thereby improving the accuracy and robustness of the wire structure response. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is a flowchart of an online monitoring system for intelligent image recognition of power transmission lines proposed in this invention;

[0069] Figure 2 This is a schematic diagram of the improved Steger line subpixel skeleton extraction process proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] refer to Figures 1-2 An online monitoring system for intelligent image recognition of power transmission lines includes the following steps:

[0072] The trajectory caching module receives a continuous image stream to generate a sequence of timestamped image frames, and stores the trajectory of the conductor centerline, local direction, and continuity parameters of the previous moment to form a trajectory cache;

[0073] The standardization module performs illumination normalization and noise suppression on the image frame, and uses the trajectory buffer to determine the strip candidate region along the direction of the conductor, generating a standardized local image.

[0074] The subpixel skeleton extraction module performs improved Steger line extraction on a standardized local image, calculates the line response and orientation field, and filters skeleton points according to the rule that the orientation difference is less than a preset threshold and the response intensity meets a preset threshold, and outputs a subpixel skeleton set.

[0075] The continuous structure enhancement module performs B-COSFIRE strip structure response calculation on a standardized local image, and generates a set of directional responses and a structure enhancement response map based on a preset set of directions and a set of scales.

[0076] The coupling and fusion module uses the center coordinates and direction of the skeleton set as anchor points, selects the maximum response in the same direction from the set of directional responses to establish a mapping, and performs connectivity verification and chain completion according to the rules of spacing and direction consistency between adjacent skeleton segments, outputting the fusion expression result of the conductor containing trajectory, topology and continuity parameters.

[0077] The identification module calculates response attenuation, interruption gap, curvature abrupt change and lateral offset indices on the wire fusion expression result to generate defect markers, and performs local incremental updates on the strip candidate region and the two-channel calculation range of the next frame based on the defect markers and trajectory buffer.

[0078] In this embodiment, forming a trajectory cache includes:

[0079] Access the video stream or image stream and complete the decoding and frame extraction. Write a unified clock timestamp for each frame and generate a frame sequence number. Establish a frame index table that increments by time.

[0080] Read the centerline trajectory from the previous moment's conductor fusion expression result, resample the trajectory points according to the distance along the line to obtain a point series with equal step size, and remove duplicate points and outliers to form a normalized trajectory point series;

[0081] The tangential direction is calculated point by point for the normalized trajectory point sequence and neighborhood smoothing is performed. The tangential direction is obtained by the direction of the line connecting adjacent points, and the endpoint direction is completed by the forward or backward neighborhood to form a direction sequence.

[0082] Calculate the distance between adjacent trajectory point pairs and generate a connectivity record. The connectivity record contains the distance value and the breakpoint identifier. The breakpoint identifier is determined by whether the distance exceeds the break threshold. Write the breakpoint number and the left and right segment numbers for each breakpoint.

[0083] For each trajectory segment, a segment-level continuity parameter is written. The segment-level continuity parameter is obtained by combining the number of breakpoints within the segment, the segment length, and the average spacing. Segment boundary indices that can be used to generate strip candidate regions in the next frame are written at the beginning and end of the trajectory segment.

[0084] The timestamp, frame number, normalized trajectory point list, direction sequence, connectivity record, segment-level continuity parameter, breakpoint number, validity marker and version number are encapsulated into a trajectory cache unit, and a validity window is set; when the next frame arrives, if the number of points covered by the new fusion expression result reaches the minimum number of points and the segment-level continuity parameter meets the update conditions, the old unit is replaced with the new trajectory cache unit; otherwise, the old unit is kept and only the frame index table is updated.

[0085] In this embodiment, generating a standardized local image includes:

[0086] Read the image frame with the corresponding frame number in the current frame index table and perform grayscale processing, keeping the pixel coordinate system consistent with the trajectory buffer, to obtain the grayscale frame;

[0087] In the grayscale frame, the brightness reference is calculated using the strip candidate region of the trajectory buffer as the statistical range, and brightness offset correction and contrast scaling are performed on the pixels within the statistical range so that the corrected grayscale distribution meets the preset target range, and an illumination normalized frame is generated.

[0088] Noise suppression processing is performed on the strip candidate region in the illumination normalized frame. The noise suppression processing adopts edge-preserving filtering and performs neighborhood update on the pixels in the region using a preset window to obtain the noise-suppressed frame.

[0089] Read the normalized trajectory point list, direction sequence, breakpoint number and segment boundary index from the trajectory cache unit, generate the centerline chain according to the point list order, and expand the preset half width along its normal at each centerline point to obtain a strip slice. When the breakpoint number is encountered, the strip slice is segmented and the end shrinkage is performed at the segment boundary index.

[0090] Each strip slice is merged according to its intra-frame coordinates to generate a strip candidate mask. The candidate mask is then closed and filled to form a single connected region. At the same time, the correspondence between the mask and the center line chain is recorded.

[0091] A strip candidate mask is used to crop and preserve pixels of the noise suppression frame, outputting a normalized local image. The normalized local image is then bound and stored with the frame number, timestamp, candidate mask identifier, and centerline chain identifier for subsequent joint calculation of subpixel skeleton channel and continuous structure enhancement channel.

[0092] In this embodiment, generating the sub-pixel skeleton set includes:

[0093] Read the standardized local image, calculate the Gaussian derivative at a fixed smoothing scale σ to obtain I_x, I_y, I_xx, I_xy, I_yy, and construct a Hessian matrix H(x,y)=[■(I_xx (x,y)&I_xy (x,y)@I_xy (x,y)&I_yy (x,y))] at each pixel position;

[0094] Simultaneously construct the gradient vector ∇I(x,y)=[I_x (x,y),I_y (x,y)]^⊤;

[0095] For each pixel, the Hessian matrix is ​​decomposed to obtain eigenvalues ​​λ_1, λ_2 and unit eigenvectors v_1, v_2. v_1 is selected as the normal n and v_2 is selected as the tangent t according to |λ_1|≥|λ_2|. The candidate point record is written with |λ_1| as the line response value R.

[0096] An initial candidate point set is generated based on the line type determination criteria. The line type determination criteria include: R is not lower than the preset response threshold, |λ_2| is not higher than the preset upper limit, and the signs of λ_1 and λ_2 satisfy the preset line type consistency condition. Pixels that do not meet the line type determination criteria are not included in the candidate point set.

[0097] For each candidate point, perform sub-pixel localization along the normal direction, calculate the normal displacement Δ, and update the center coordinates to the sub-pixel center coordinates. The displacement satisfies:

[0098] Δ=-(n^⊤ ∇I) / (n^⊤ Hn);

[0099] The sub-pixel center coordinates, tangent t, and response value R are written into the sub-pixel record.

[0100] Non-maximum constraints are applied along the normal direction to eliminate duplicate responses: Based on the sub-pixel center coordinates, samples are taken on both sides of the normal direction at fixed sampling intervals. The response values ​​at the center position and the two sides are compared, and only sub-pixel points whose response values ​​at the center position are not less than the response values ​​at the two sides are recorded.

[0101] Based on centerline recalibration and interference removal of normal double boundary symbol pairs, a one-dimensional profile is established along the normal direction for each sub-pixel record that passes the non-maximum constraint, and the first derivative projection of the normal direction is calculated.

[0102] g(s) = ∇I(p+sn)⋅n;

[0103] Where p is the sub-pixel center coordinate and s is the normal offset; on the cross-section, points satisfying |g(s)| with a local maximum and a positive sign are searched as boundary points on one side, and points satisfying |g(s)| with a local maximum and a negative sign are searched as boundary points on the other side, resulting in two boundary offsets s_L and s_R; when the two boundaries satisfy g(s_L)⋅g(s_R)<0 and |s_L+s_R| does not exceed the preset symmetry threshold, the sub-pixel center coordinates are moved along the normal to the midpoint position p←p+(s_L+s_R) / 2 n of the two boundaries and the point is retained; when no pair of boundary points with opposite signs are found or |s_L+s_R| exceeds the preset symmetry threshold, the sub-pixel point record is deleted, thereby completing the centerline recalibration and suppressing the pseudo-line response caused by unilateral edges, straight edges of tower materials, and vegetation lines;

[0104] In the corrected sub-pixel records, a trajectory cache orientation consistency constraint is introduced: for each sub-pixel, the nearest neighbor trajectory point is matched in the trajectory cache and the trajectory orientation t_ref is read, Δθ=arccos(|t⋅t_ref |) is calculated, and only sub-pixel records with Δθ less than the preset orientation difference threshold are retained;

[0105] Based on the single-chain skeleton construction process of mutual best inheritance, each retained sub-pixel is used as a node to establish a candidate inheritance relationship: the forward half-plane of the node's tangent t is used as the search area, and the node with the largest response value R within the fixed search radius is selected as the forward inheritance node q_f; the reverse verification node q_b of its forward inheritance node is also selected in the forward half-plane of q_f's tangent. When the reverse verification node is the same node as the current node, mutual best inheritance is confirmed and a directed link is established; otherwise, no link is established. The mutual best inheritance determination is repeated for all nodes to obtain several non-branching links.

[0106] Output skeleton segments on a per-link basis: Sort the nodes of each link according to the link order and write the skeleton segment number and the sequence number within the segment; delete the entire link whose length is less than the preset minimum number of points; summarize all nodes in the retained links to form a sub-pixel skeleton set, which includes the sub-pixel center coordinates, tangential direction, response value, skeleton segment number and sequence number within the segment.

[0107] In this embodiment, generating the directional response set and the structural enhancement response map includes:

[0108] Read standardized local images, establish a strip primitive response library within a preset scale set, perform differential Gaussian convolution on the images at each scale and perform half-wave rectification to obtain strip primitive response maps, and perform local non-maximum suppression on the strip primitive response maps to form a peak candidate table. The peak candidate table records at least the peak coordinates, peak intensity and peak scale index.

[0109] For each direction θ_k in the preset direction set, a B-COSFIRE operator template is generated. The operator template sets a radial distance sequence ρ_j with the operator center as the origin and generates an angle search window for each ρ_j. The angle search window is centered on θ_k and has a limited angle span. Peak candidates that fall within the radius interval [ρ_j-δ, ρ_j+δ] and whose angles fall within the angle search window are retrieved from the peak candidate table. The peak candidate ranked first in peak intensity from high to low is selected as the support point of ρ_j, and the (ρ_j, ϕ_j) and scale index of the support point are written into the operator template. When no peak candidate is found for ρ_j, ρ_j is marked as missing and not written into the support point.

[0110] For each operator center in each direction θ_k, support point response sampling is performed. For each support point in the operator template, two sampling steps are completed on the bar primitive response map corresponding to its scale index: the first step is to establish a fixed radius neighborhood centered on the theoretical coordinates of the support point and search for the maximum response in the neighborhood as the support point sample value; the second step is to perform allowable displacement diffusion of the support point sample value in space. The diffusion adopts a fixed diffusion rule with the radial distance from the support point to the center as input, and the diffused value is written into the support point response table.

[0111] The support point responses under the same direction θ_k are combined according to the B-COSFIRE combination rule to generate directional response values. The directional response values ​​are then combined with the support point responses in the form of geometric mean. The combination expression is as follows:

[0112] R_(θ_k ) (x,y)=(∏_(i=1)^N▒ s_i (x,y))^(1 / N);

[0113] Where s_i (x,y) is the support point response of the i-th support point at the operator center (x,y), and N is the number of support points participating in the combination; the set of directional responses is obtained by repeating the calculation for all directions;

[0114] The template generation and response gating process for bidirectional pairwise locking of the main axis involves generating an additional symmetrical radial distance -ρ_j in the reverse direction of the main axis for each radial distance ρ_j, and performing the same peak retrieval and support point selection. The support points in the forward direction and the support points in the reverse direction of the main axis are written into the same pair of numbers. When a support point is missing on either side of the same pair of numbers, the pair of numbers is directly deleted from the operator template. When calculating the directional response, the smaller value of the responses of the support points on both sides of the same pair of numbers is taken as the effective response of the pair of numbers, and the effective response replaces the original responses of the support points on both sides to participate in the geometric mean combination, so that the directional response is only triggered when strip-shaped evidence is formed in both directions of the main axis.

[0115] The cross-structure shielding process for tangential disproving sampling involves generating a tangential disproving sampling segment at a reference position corresponding to each radial distance ρ_j. The tangential disproving sampling segment is expanded along the tangential direction of direction θ_k and has a fixed length. Peak candidates falling into the tangential disproving sampling segment are retrieved from the peak candidate table, and their maximum peak intensity is taken as the disproving intensity. When the disproving intensity of any ρ_j reaches a preset disproving threshold, the directional response value of the operator center in direction θ_k is set to zero and the directional output is terminated, so that the peak values ​​generated in the tangential direction by the crossarm edge, grid stripes, and cross textures form a disproving suppression of the directional response.

[0116] At each pixel location, the maximum value of the directional response set is taken to form a structural enhancement response map, and the direction index of the maximum value is recorded to form an enhancement direction map. On the enhancement direction map, only pixels with the same direction index and structural enhancement response not lower than the preset enhancement threshold are connected by 8 neighborhoods to form a connected link. Connected links with a link length lower than the preset minimum link length are deleted, and the structural enhancement response map and directional response set after link filtering are output.

[0117] In this embodiment, the generated wire fusion expression result includes:

[0118] Read the subpixel skeleton set and form an ordered skeleton segment sequence according to the skeleton segment number. At the same time, read the directional response set and the structural enhancement response map to establish a fusion processing unit with the skeleton point as the core. Extract the center coordinates and tangential direction of each skeleton point as anchor point parameters.

[0119] For each skeleton point, a directional search sector is constructed in the structural enhancement response map. The directional search sector has the skeleton point as the vertex, the tangential direction as the axis of symmetry, and a fixed angle range as the fan-shaped angle. Within the sector, sampling is performed step by step along the tangential principal axis. The pixel with the largest structural enhancement response value is selected as the coupled response pixel, and the pixel coordinates and enhancement direction index are recorded to complete the directional maximum mapping from the skeleton point to the enhancement response.

[0120] For each skeleton segment, a coupling strength sequence is constructed in sequence according to the serial number. The coupling strength sequence is obtained by combining the skeleton point and line response values ​​with the corresponding coupling response values ​​according to a fixed weight. The sequence is scanned in sequence, and segments that are continuously below the coupling threshold are marked as candidate fracture segments.

[0121] The trajectory extrapolation-enhanced connectivity verification process is performed on adjacent skeleton segments: a linear extrapolation segment of fixed length is generated at the end of each skeleton segment based on the tangential direction of the end point. Within the range of the linear extrapolation segment, it is searched to see if the starting point of another skeleton segment falls within a preset distance interval. When a matching starting point exists, the continuity of the enhanced response in the area covered by the extrapolation segment is further verified in the structural enhanced response map. Only when the enhanced response on the extrapolation path is continuous and not lower than the enhancement threshold is it determined that the two skeleton segments are connected.

[0122] For skeleton segments determined to be connectable, an enhanced guided chain-complementation process is executed: with the extrapolated segment as the main axis, sampling points are generated along the main axis direction in the structural enhancement response map at a fixed step size, and the pixel with the largest enhancement response is selected as the chain-complementation node within the limited width on both sides of the normal of each sampling point. The chain-complementation nodes are inserted sequentially between the two skeleton segments to form a fusion trajectory.

[0123] The fused trajectory is subjected to topological reconstruction, the trajectory number and segment number are renumbered, and a connectivity table is generated. At the same time, trajectory-level continuity parameters are calculated, including total trajectory length, supplementary chain length ratio and number of breaks. The trajectory point sequence, topological number and continuity parameters are encapsulated into a traverse fusion expression output, so that the fusion result has both geometric continuity and structural enhancement consistency constraints.

[0124] In this embodiment, generating defect markers and performing local incremental updates includes:

[0125] Read the current frame conductor fusion expression result, obtain the fusion trajectory point sequence, skeleton segment number, supplementary chain segment mark, coupling strength sequence and enhancement direction index, and establish a continuous detection window within the segment according to the trajectory number;

[0126] Four types of disturbance candidate segments are formed on the trajectory sequence: segments that continuously decrease and whose decrease reaches the attenuation threshold are recorded as response attenuation candidates by comparing adjacent windows of the coupling strength sequence with a fixed window; segments that trigger a break are recorded as interruption gap candidates by comparing the distance between adjacent trajectory points with the break threshold, and the "repaired chain" and "unrepaired chain" states are distinguished by the supplementary chain segment; discrete curvature is calculated by three-point neighborhood and compared with the baseline curvature within the segment, and segments that exceed the abrupt change threshold and whose duration reaches the minimum length are recorded as curvature abrupt change candidates; the current trajectory is aligned to the trajectory cache trajectory by arc length, and the normal distance is calculated at the alignment point, and segments whose normal distance continuously exceeds the offset threshold are recorded as lateral offset candidates;

[0127] The defect identification process of intra-segment consensus voting uses the skeleton segment as the voting unit. It maps four types of candidate segments to the corresponding skeleton segments and counts the number of hits. Only when at least two types of candidates in the same skeleton segment hit simultaneously and one of them must be "unpaired chain interruption gap" or "lateral offset" is output as a defect segment. The output defect segment is written with the defect type set, start and end number, segment length, enhancement direction index and chain completion status to form the defect marking result.

[0128] Based on the defect marking results, the incremental candidate region for the next frame is generated: taking the minimum bounding rectangle of the defect segment in the image coordinates as the basis, the region is expanded by a fixed length margin along the trajectory tangent and by a fixed half width along the normal to obtain the defect update region; at the same time, a confidence recovery region is generated for the segment that has not triggered a defect but whose coupling strength is lower than the coupling threshold. The confidence recovery region is generated according to the same expansion rule.

[0129] The two-layer incremental recalculation scheduling process prioritizes the defect update region and writes it into the two-channel calculation range of the next frame, while prioritizing the confidence recovery region and writing it only into the calculation range of the structural enhancement channel. When the high-priority region exists, the next frame performs full-process recalculation of the sub-pixel skeleton channel and continuous structural enhancement channel only in the high-priority region, and updates the trajectory cache with the recalculation results. When only the low-priority region exists, the next frame keeps the skeleton channel output unchanged, updates the structural enhancement response in the low-priority region, and triggers coupling fusion verification.

[0130] In this embodiment, triggering the coupling fusion verification includes: when only a secondary priority region is detected, reading the trajectory point sequence, segment number, and continuity parameter of the previous time step conductor fusion expression result from the trajectory cache, and recalculating the secondary priority region in the current frame structure enhancement channel to obtain a new structure enhancement response map, direction response set, and enhancement direction index; using the trajectory points in the trajectory cache as anchor points, constructing a directional search sector at each anchor point position according to its tangential direction, retrieving candidate directions consistent with the anchor point tangential direction from the recalculated direction response set, and selecting the pixel with the largest structure enhancement response within the sector to establish a mapping from anchor point to enhancement response, obtaining an updated coupling response sequence; and performing segment-by-segment coupling strength sequence with the updated coupling response sequence and the coupling strength sequence stored in the original fusion expression result. For alignment comparison, continuous segments with positioning differences exceeding a preset change threshold are designated as verification segments. Connectivity verification is performed on the verification segments. Specifically, linear extrapolation segments are generated tangentially at both ends of the verification segments, and the enhanced response continuity of the extrapolation path is verified on the structural enhanced response map. If the continuity is satisfied, the original topology connection and segment number remain unchanged, and only the coupling response field, continuity parameters, and confidence flags within the verification segments are updated. If the continuity is not satisfied, enhanced guided supplementary chains are performed within the boundaries of the verification segments. Samples are taken along the extrapolation principal axis at a fixed step size, and the largest enhanced response pixel is selected as the supplementary chain node within the normal defined width. The supplementary chain node is inserted to form a new fusion trajectory. Subsequently, topology reconstruction is performed on the affected trajectory segments, and the trajectory point sequence, connection table, and continuity parameters in the trajectory cache are updated.

[0131] Example:

[0132] To verify the feasibility of this invention, it was applied to an online monitoring scenario of a power transmission line characterized by mountains, forest belts, road crossings, and densely distributed towers. In this scenario, the monitoring image simultaneously contains various slender structures such as conductors, tower crossarms, insulator strings, ground wires, tree branches, shrub edges, road guardrails, and distant building outlines. Conductors in the image exhibit narrow width, sensitivity to brightness changes, and susceptibility to local occlusion. Under conditions of backlighting, sidelighting, fog, compression noise, and slight equipment vibration, problems such as unclear conductor edges, local grayscale collapse, line segment interruptions, and enhanced pseudo-line responses easily occur. Traditional methods typically perform threshold segmentation, edge detection, or ordinary line structure extraction directly on the entire image frame. While these methods can achieve some recognition results in relatively simple backgrounds, in complex conditions like this, they often misidentify tower edges or vegetation lines as conductors. Furthermore, problems such as local occlusion of conductors, uneven brightness, and insufficient continuity in areas with background cross-textures further lead to unstable subsequent anomaly recognition, excessive online computational burden, and difficulty in meeting the needs of long-term online monitoring.

[0133] In this embodiment, an online monitoring system for intelligent image recognition of transmission lines includes a trajectory caching module, a standardization module, a subpixel skeleton extraction module, a continuous structure enhancement module, a coupling fusion module, and a recognition module. The monitoring end acquires the transmission line video stream using continuous image acquisition, with an image resolution of 1920 x 1080 pixels and an input frame rate of 25 frames per second. The trajectory caching module decodes, extracts, and binds time stamps to the input image stream, forming a continuous image frame sequence, and establishes an incremental frame index for each frame. The module writes the centerline trajectory, trajectory point tangential direction, adjacent point spacing, breakpoint identifier, and segment continuity parameters from the conductor fusion expression result of the previous moment into the trajectory caching unit. The trajectory points are standardized using a constant-step resampling method, with a resampling interval of three pixels. The direction of adjacent points is smoothly obtained by connecting previous and subsequent neighboring areas, and the breakpoint threshold is set to eight pixels. When a new frame arrives, if the current frame fusion result covers at least eighty points and the trajectory segment continuity parameters meet the update conditions, the current frame cache replaces the previous frame cache; otherwise, the original cache is maintained, and only the frame index is updated. With the above settings, the system introduces the geometric information of the conductor that has been confirmed in the previous frame into the processing of the current frame, so that the conductor search in the current frame no longer depends on the unconstrained scanning of the entire image, but performs a directional search along the vicinity of the historical trajectory.

[0134] After reading the current frame image, the standardization module converts the color image to a grayscale image while maintaining the pixel coordinate system consistent with the trajectory buffer. Using the candidate strip region determined by the trajectory buffer as the statistical range, brightness shift correction and contrast scaling are performed on the grayscale distribution within the candidate region. The candidate region is obtained by extending the centerline chain along both sides of the normal direction, with an expansion half-width of 24 pixels. When there are breakpoints in the trajectory, the strip region is segmented and shrunk according to the breakpoint positions to avoid crossing invalid areas. After brightness normalization, edge-preserving filtering is performed within the candidate region, with a filtering window set to 5x5 pixels to suppress compression noise and local random noise while preserving the transition features of the conductor edges. Then, the system crops the image based on the candidate region mask, retaining only the standardized local image related to the conductor, and saves the local image, corresponding frame number, timestamp, and centerline chain identifier together for subsequent dual-channel joint calculations. Through this processing, large areas unrelated to the conductor in the complex background are eliminated, and the subsequent calculation range is limited to the local strip space distributed along the conductor.

[0135] The subpixel skeleton extraction module performs improved Steger line extraction on a standardized local image. The module first calculates the second derivative and gradient information of the image at a fixed smoothing scale, constructs a Hessian matrix at each pixel location, obtains the normal and tangential directions through eigenvalue decomposition, and takes the principal eigenvalue as the line response value. In this embodiment, the smoothing scale is set to 1.2, and the response threshold is set to 0.18. The module generates an initial candidate point set based on the response intensity and line type consistency, and then performs subpixel localization along the normal direction, calculating the subpixel center coordinates. To avoid misidentifying a single-sided edge as the center line of a conductor, this embodiment adds a normal double-boundary sign pair correction process after subpixel localization. Specifically, a one-dimensional profile is established along the normal direction for each candidate point, and positive and negative peak boundary points of the first derivative projection are searched on both sides of the normal direction. If both boundaries exist simultaneously and the boundary distance difference does not exceed a preset symmetry threshold, the candidate point is moved to the midpoint of the two boundaries as the re-correction center; if only a single boundary exists, or the distance between the two boundaries is significantly asymmetrical, the candidate point is deleted. The symmetry threshold is set to one pixel. After recalibration, the module performs directional consistency filtering based on the direction information of the nearest neighbor trajectory points in the trajectory cache, retaining only sub-pixel points with a directional difference of no more than 15 degrees. Subsequently, the module uses each retained point as a node and establishes a single-chain skeleton segment according to the best-inheritance rule. The search radius is set to six pixels, and a link is established only if the optimal successor node selected for a node in the forward half-plane of the tangential direction still points to the node itself after reverse verification. Skeleton segments with a chain length of less than twelve points are deleted entirely. After this processing, the system outputs a set of sub-pixel skeletons containing sub-pixel center coordinates, tangential direction, response value, skeleton segment number, and segment sequence number. These technical features enable the conductor skeleton to not only have sub-pixel-level positioning accuracy but also effectively avoid pseudo-skeleton bifurcations caused by straight edges of tower materials, vegetation edges, and random textures.

[0136] The continuous structure enhancement module performs B-COSFIRE strip structure response calculations on the same standardized local image. First, the module establishes a strip primitive response library for a preset scale set. The scale set includes four scales: two, three, four, and five pixels. The direction set is divided into twelve directions, ranging from 0 to 180 degrees, each in 15-degree increments. For each scale, the system performs differential Gaussian convolution and half-wave rectification to obtain the strip primitive response map, and extracts local peak candidates from the response map. Then, a B-COSFIRE operator template is constructed for each direction, setting multiple radial distance positions around the operator center. In this embodiment, the radial distances are six pixels, ten pixels, fourteen pixels, and eighteen pixels, respectively. The module selects the candidate point with the strongest peak within the angle search window at each radial position as the support point and records the corresponding coordinates and scale index. To ensure the enhancement results better align with the bidirectional extension characteristics of the conductor's main axis, this embodiment introduces a bidirectional pairing locking mechanism during template generation. For each positive radial position, a reverse symmetrical position is simultaneously established. Only when both positive and negative positions simultaneously acquire valid support points does the pair of support points participate in subsequent combinations; if either side is missing, the entire pair of support points is deleted. During the response calculation phase, the smaller response value is taken as the valid response for the same pair of support points to ensure that the directional response is supported bidirectionally by the main axis. Furthermore, to suppress false enhancements caused by crossarm edges, grid textures, and intersecting background lines, this embodiment constructs a tangential disproving sampling segment at each radial reference position. If a peak value exceeding the disproving threshold is found within the segment range, the current directional response is directly set to zero. The disproving threshold is set to 0.22. After completing all directional calculations, the module takes the maximum directional response at each pixel position to form a structural enhancement response map and outputs the direction index map corresponding to the maximum response. Short connected chains with a length less than ten pixels are then deleted. With this technical solution, the enhancement result obtained by the system is no longer a uniform amplification of all slender textures in the ordinary sense, but only retains the strip structure response that simultaneously satisfies the bidirectional consistency of the principal axis and the absence of obvious disproving interference in the tangential direction.

[0137] The coupling fusion module reads the subpixel skeleton set, directional response set, and structural enhancement response map, and establishes a fusion processing unit using the center coordinates and tangential direction of the skeleton point as fusion anchor points. For each skeleton point, the system constructs a directional search sector in front of its tangential direction, with a sector radius of 12 pixels and a subtended angle of 30 degrees. Within the sector, it finds the pixel with the largest structural enhancement response as the coupling response point and records the coordinates and enhancement direction index of the point. Then, the system constructs a coupling strength sequence according to the skeleton segment order. The coupling strength is formed by weighting the skeleton line response value and the enhancement response value, with weight ratios set to 0.6 and 0.4. When the coupling strength is continuously lower than the threshold, the corresponding segment is marked as a candidate break segment. For adjacent skeleton segments, the system performs linear extrapolation in the tangential direction of the end point, with an extrapolation length of 15 pixels. If the starting point of another skeleton segment falls within a preset distance window, and the structural enhancement response remains continuous on the extrapolation path, it is determined that the two skeleton segments are connected, and supplementary chain nodes are generated by sampling near the extrapolation main axis at a fixed step size. The nodes of the supplementary chain still need to meet the enhanced response continuity condition, and finally together with the original skeleton segment, they form a fused trajectory. The output results include the trajectory point sequence, trajectory number, segment number, supplementary chain ratio, number of breaks, and continuity parameters. The technical feature of the module is that it does not simply rely on the geometric position of the skeleton points for splicing, but uses directional response and structural continuity to jointly judge the rationality of the supplementary chain, so that the final traverse expression result has both centerline accuracy and overall trajectory continuity.

[0138] The identification module calculates four types of indicators based on the fused trajectory: response decay, interruption gap, curvature abrupt change, and lateral offset. For response decay, the system uses a sliding window of nine points on the coupling strength sequence to compare adjacent windows. If the continuous decrease exceeds 25%, it is marked as a decay candidate. For interruption gap, if the distance between adjacent trajectory points exceeds the break threshold, an interruption gap candidate is formed, and the supplementary chain mark is used to distinguish between supplemented and unsupplemented states. For curvature abrupt change, the system uses a three-point neighborhood discrete curvature calculation method. If the local curvature continuously exceeds twice the baseline curvature within the segment, a curvature abrupt change candidate is generated. For lateral offset, the system aligns the current trajectory with the historical trajectories in the trajectory cache by arc length and calculates the normal distance. If the normal distance continuously exceeds five pixels, a lateral offset candidate is generated. Then, the module performs consensus voting on the four types of candidates on a skeleton segment basis. Only when at least two types of candidates in the same skeleton segment are simultaneously matched, and it includes an unsupplemented interruption gap or lateral offset, is a defect mark output. Based on the defect marking results, the system generates high-priority defect update regions; for segments that do not trigger defects but have low coupling strength, a secondary-priority confidence recovery region is generated. In subsequent frame processing, high-priority regions undergo full-process recalculation of the skeleton channel and structural enhancement channel, while secondary-priority regions only update the structural enhancement channel and trigger coupling fusion verification. This technical feature enables the system to concentrate its main computing power on local areas where anomalies may occur during continuous monitoring, avoiding repeated calculations across entire frames.

[0139] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An online monitoring system for intelligent image recognition of power transmission lines, characterized in that, include: The trajectory caching module receives a continuous image stream to generate a sequence of timestamped image frames, and stores the trajectory of the conductor centerline, local direction, and continuity parameters of the previous moment to form a trajectory cache; The standardization module performs illumination normalization and noise suppression on the image frame, uses the trajectory buffer to determine the strip-shaped candidate region along the direction of the conductor, and generates a standardized local image. The subpixel skeleton extraction module performs improved Steger line extraction on a standardized local image, calculates the line response and orientation field, filters skeleton points, and outputs a set of subpixel skeletons. The continuous structure enhancement module performs B-COSFIRE strip structure response calculation on a standardized local image, and generates a set of directional responses and a structure enhancement response map based on a preset set of directions and a set of scales. The coupling and fusion module uses the center coordinates and direction of the skeleton set as anchor points, selects the maximum response in the same direction from the set of directional responses to establish a mapping, and performs connectivity verification and chain completion according to the rules of spacing and direction consistency between adjacent skeleton segments, outputting the fusion expression result of the conductor containing trajectory, topology and continuity parameters. The identification module calculates response attenuation, interruption gap, curvature abrupt change and lateral offset indices on the wire fusion expression result to generate defect markers, and performs local incremental updates on the strip candidate region and the two-channel calculation range of the next frame based on the defect markers and trajectory buffer.

2. The online monitoring system for intelligent image recognition of transmission lines according to claim 1, characterized in that, The formation of the trajectory cache includes: Access the video stream or image stream and complete the decoding and frame extraction. Write a unified clock timestamp for each frame and generate a frame sequence number. Establish a frame index table that increments by time. Read the centerline trajectory from the previous moment's conductor fusion expression result, resample the trajectory points according to the distance along the line to obtain a point series with equal step size, and remove duplicate points and outliers to form a normalized trajectory point series; The tangential direction is calculated point by point for the normalized trajectory point sequence and neighborhood smoothing is performed. The tangential direction is obtained by the direction of the line connecting adjacent points, and the endpoint direction is completed by the forward or backward neighborhood to form a direction sequence. Calculate the distance between adjacent trajectory point pairs and generate a connectivity record. The connectivity record contains the distance value and the breakpoint identifier. The breakpoint identifier is determined by whether the distance exceeds the break threshold. Write the breakpoint number and the left and right segment numbers for each breakpoint. For each trajectory segment, a segment-level continuity parameter is written. The segment-level continuity parameter is obtained by combining the number of breakpoints within the segment, the segment length, and the average spacing. The segment boundary index generated by the next frame's strip candidate region is written at the beginning and end of the trajectory segment. The timestamp, frame number, normalized trajectory point list, direction sequence, connectivity record, segment-level continuity parameter, breakpoint number, validity marker and version number are encapsulated into a trajectory cache unit, and an expiration window is set; when the next frame arrives, if the number of points covered by the new fusion expression result reaches the minimum number of points and the segment-level continuity parameter meets the preset update conditions, the old unit is replaced with the new trajectory cache unit; otherwise, the old unit is kept and only the frame index table is updated.

3. The online monitoring system for intelligent image recognition of transmission lines according to claim 1, characterized in that, Generating a normalized local image includes: Read the image frame with the corresponding frame number in the current frame index table and perform grayscale processing, keeping the pixel coordinate system consistent with the trajectory buffer, to obtain the grayscale frame; In the grayscale frame, the brightness reference is calculated using the strip candidate region of the trajectory buffer as the statistical range, and brightness offset correction and contrast scaling are performed on the pixels within the statistical range to generate an illumination normalized frame. Noise suppression processing is performed on the strip candidate region in the illumination normalized frame. The noise suppression processing adopts edge-preserving filtering and performs neighborhood update on the pixels in the region using a preset window to obtain the noise-suppressed frame. Read the normalized trajectory point list, direction sequence, breakpoint number and segment boundary index from the trajectory cache unit, generate the centerline chain according to the point list order, and expand the preset half width along its normal at each centerline point to obtain a strip slice. When the breakpoint number is encountered, the strip slice is segmented and the end shrinkage is performed at the segment boundary index. Each strip slice is merged according to its intra-frame coordinates to generate a strip candidate mask. The candidate mask is then closed and filled to form a single connected region. At the same time, the correspondence between the mask and the center line chain is recorded. A strip candidate mask is used to crop and retain pixels of the noise suppression frame, outputting a standardized local image. The standardized local image is then bound and stored with the frame number, timestamp, candidate mask identifier, and centerline chain identifier.

4. The online monitoring system for intelligent image recognition of transmission lines according to claim 1, characterized in that, The generated subpixel skeleton set includes: Read the standardized local image, calculate the Gaussian derivative at a fixed smoothing scale σ to obtain I_x, I_y, I_xx, I_xy, I_yy, and construct a Hessian matrix H(x,y)=[■(I_xx (x,y)&I_xy (x,y)@I_xy(x,y)&I_yy (x,y))] at each pixel position, and simultaneously construct the gradient vector ∇I(x,y)=[I_x (x,y),I_y (x,y)]^⊤; For each pixel, the Hessian matrix is ​​decomposed to obtain eigenvalues ​​λ_1, λ_2 and unit eigenvectors v_1, v_2. v_1 is selected as the normal n and v_2 is selected as the tangent t according to |λ_1|≥|λ_2|. The candidate point record is written with |λ_1| as the line response value R. An initial candidate point set is generated based on the line type determination criteria. The line type determination criteria include: R is not lower than the preset response threshold, |λ_2| is not higher than the preset upper limit, and the signs of λ_1 and λ_2 satisfy the preset line type consistency condition. Pixels that do not meet the line type determination criteria are not included in the candidate point set. For each candidate point, perform sub-pixel localization along the normal direction, calculate the normal displacement Δ, and update the center coordinates to the sub-pixel center coordinates. The displacement satisfies: Δ=-(n^⊤ ∇I) / (n^⊤ Hn); The sub-pixel center coordinates, tangent t, and response value R are written into the sub-pixel record. Non-maximum constraints are applied along the normal direction to eliminate duplicate responses: Based on the sub-pixel center coordinates, samples are taken on both sides of the normal direction at fixed sampling intervals. The response values ​​at the center position and the two sides are compared, and only sub-pixel points whose response values ​​at the center position are not less than the response values ​​at the two sides are recorded. Based on centerline recalibration and interference removal of normal double boundary symbol pairs, for each sub-pixel record that passes the non-maximum constraint, a one-dimensional profile is established along the normal direction and the first derivative projection in the normal direction is calculated: g(s) = ∇I(p+sn)⋅n; Where p is the subpixel center coordinate and s is the normal offset; on the profile, search for points on one side where |g(s)| achieves a local maximum and has a positive sign as boundary points, and search for points on the other side where |g(s)| achieves a local maximum and has a negative sign as boundary points, to obtain the two boundary offsets s_L and s_R; when the two boundaries satisfy g(s_L)⋅g(s_R)<0 and |s_L+s_R| does not exceed the preset symmetry threshold, move the subpixel center coordinates along the normal to the midpoint position p←p+(s_L+s_R) / 2 n of the two boundaries and retain the point; when no pair of boundary points with opposite signs are found or |s_L+s_R| exceeds the preset symmetry threshold, delete the subpixel point record; In the corrected sub-pixel records, a trajectory cache orientation consistency constraint is introduced: for each sub-pixel, the nearest neighbor trajectory point is matched in the trajectory cache and the trajectory orientation t_ref is read, Δθ=arccos(|t⋅t_ref |) is calculated, and only sub-pixel records with Δθ less than the preset orientation difference threshold are retained; Based on the single-chain skeleton construction process of mutual best inheritance, each retained sub-pixel is used as a node to establish a candidate inheritance relationship: the forward half-plane of the node's tangent t is used as the search area, and the node with the largest response value R within the fixed search radius is selected as the forward inheritance node q_f; the reverse verification node q_b of its forward inheritance node is also selected in the forward half-plane of q_f's tangent. When the reverse verification node is the same node as the current node, mutual best inheritance is confirmed and a directed link is established; otherwise, no link is established. The mutual best inheritance determination is repeated for all nodes to obtain N non-branching links. Output skeleton segments by link: Sort the nodes of each link according to the link order and write the skeleton segment number and the sequence number within the segment; delete the entire link whose length is less than the preset minimum number of points; summarize all nodes in the retained links to form a sub-pixel skeleton set.

5. The online monitoring system for intelligent image recognition of transmission lines according to claim 4, characterized in that, The Gaussian derivatives I_x, I_y, I_xx, I_xy, and I_yy are obtained by calculating the Gaussian derivatives under a fixed smoothing scale σ, including: Using a standardized local image as input, a two-dimensional Gaussian kernel with a smoothing scale of σ is generated in memory. The first and second partial derivatives of this Gaussian kernel are calculated along the x and y directions, resulting in five convolution kernels: a first derivative kernel in the x direction, a first derivative kernel in the y direction, a second derivative kernel in the x direction, a mixed second derivative kernel in the x and y directions, and a second derivative kernel in the y direction. The image is then subjected to boundary mirroring expansion, and the expanded image is convolved with each of the five convolution kernels in two dimensions. The convolution outputs are five derivative response maps: I_x, I_y, I_xx, I_xy, and I_yy. I_x and I_y together constitute a gradient vector field representing the direction and intensity of local gray-level changes, while I_xx, I_xy, and I_yy are assembled into a Hessian matrix at each pixel location.

6. The online monitoring system for intelligent image recognition of transmission lines according to claim 1, characterized in that, The generated directional response set and structure-enhanced response map include: Read standardized local images, establish a strip primitive response library within a preset scale set, perform differential Gaussian convolution on the images at each scale and perform half-wave rectification to obtain strip primitive response maps, and perform local nonmaximum suppression on the strip primitive response maps to form a peak candidate table; For each direction θ_k in the preset direction set, a B-COSFIRE operator template is generated. The operator template sets a radial distance sequence ρ_j with the operator center as the origin and generates an angle search window for each ρ_j. The angle search window is centered on θ_k and has a limited angle span. Peak candidates that fall within the radius interval [ρ_j-δ, ρ_j+δ] and whose angles fall within the angle search window are retrieved from the peak candidate table. The peak candidate ranked first in peak intensity from high to low is selected as the support point of ρ_j, and the (ρ_j, ϕ_j) and scale index of the support point are written into the operator template. When no peak candidate is found for ρ_j, ρ_j is marked as missing and not written into the support point. For each operator center in each direction θ_k, support point response sampling is performed. For each support point in the operator template, two sampling steps are completed on the bar primitive response map corresponding to its scale index: the first step is to establish a fixed radius neighborhood centered on the theoretical coordinates of the support point and search for the maximum response in the neighborhood as the support point sample value; the second step is to perform allowable displacement diffusion of the support point sample value in space. The diffusion adopts a fixed diffusion rule with the radial distance from the support point to the center as input, and the diffused value is written into the support point response table. The support point responses under the same direction θ_k are combined according to the B-COSFIRE combination rule to generate directional response values. The directional response values ​​are then combined with the support point responses in the form of geometric mean. The combination expression is as follows: R_(θ_k ) (x,y)=(∏_(i=1)^N▒ s_i (x,y))^(1 / N); Where s_i (x,y) is the support point response of the i-th support point at the operator center (x,y), and N is the number of support points participating in the combination; the set of directional responses is obtained by repeating the calculation for all directions; The template generation and response gating process for bidirectional pairwise locking of the main axis involves generating an additional symmetrical radial distance -ρ_j in the reverse direction of the main axis for each radial distance ρ_j, and performing the same peak retrieval and support point selection. The support points in the forward direction and the support points in the reverse direction of the main axis are written into the same pair of numbers. When a support point is missing on either side of the same pair of numbers, the pair of numbers is directly deleted from the operator template. In the direction response calculation, the smaller value of the responses of the support points on both sides of the same pair of numbers is taken as the effective response of the pair of numbers, and the effective response replaces the original responses of the support points on both sides to participate in the geometric mean combination. The cross-structure shielding process for tangential disproving sampling involves generating a tangential disproving sampling segment at a reference position corresponding to each radial distance ρ_j. The tangential disproving sampling segment is expanded along the tangential direction of direction θ_k and has a fixed length. Peak candidates falling into the tangential disproving sampling segment are retrieved from the peak candidate table, and their maximum peak intensity is taken as the disproving intensity. When the disproving intensity of any ρ_j reaches a preset disproving threshold, the directional response value of the operator center in direction θ_k is set to zero, and the directional output is terminated. At each pixel location, the maximum value of the directional response set is taken to form a structural enhancement response map, and the direction index of the maximum value is recorded to form an enhancement direction map. On the enhancement direction map, only pixels with the same direction index and structural enhancement response not lower than the preset enhancement threshold are connected by 8 neighborhoods to form a connected link. Connected links with a link length lower than the preset minimum link length are deleted, and the structural enhancement response map and directional response set after link filtering are output.

7. The online monitoring system for intelligent image recognition of transmission lines according to claim 1, characterized in that, The generated wire fusion expression results include: Read the subpixel skeleton set and form an ordered skeleton segment sequence according to the skeleton segment number. At the same time, read the directional response set and the structural enhancement response map to establish a fusion processing unit with the skeleton point as the core. Extract the center coordinates and tangential direction of each skeleton point as anchor point parameters. For each skeleton point, a directional search sector is constructed in the structural enhancement response map. The directional search sector has the skeleton point as the vertex, the tangential direction as the axis of symmetry, and a fixed angle range as the fan-shaped angle. Within the sector, sampling is performed step by step along the tangential principal axis. The pixel with the largest structural enhancement response value is selected as the coupled response pixel, and the pixel coordinates and enhancement direction index are recorded to complete the directional maximum mapping from the skeleton point to the enhancement response. For each skeleton segment, a coupling strength sequence is constructed in sequence according to the serial number. The coupling strength sequence is obtained by combining the skeleton point and line response values ​​with the corresponding coupling response values ​​according to a fixed weight. The sequence is scanned in sequence, and segments that are continuously below the coupling threshold are marked as candidate fracture segments. At the end of each skeleton segment, a linear extrapolation segment of fixed length is generated according to the tangential direction of the end point. Within the range of the linear extrapolation segment, it is searched whether the starting point of another skeleton segment falls into a preset distance interval. When a matching starting point exists, the continuity of the enhanced response in the area covered by the extrapolation segment is verified in the structural enhancement response map. Only when the enhanced response on the extrapolation path is continuous and not lower than the enhancement threshold, it is determined that the two skeleton segments are connected. Using the extrapolated segment as the main axis, sampling points are generated along the main axis direction with a fixed step size in the structural enhancement response map. Within the limited width on both sides of the normal of each sampling point, the pixel with the largest enhancement response is selected as the supplementary chain node. The supplementary chain nodes are inserted sequentially between the two skeleton segments to form a fusion trajectory. The fused trajectory is subjected to topology reconstruction, the trajectory number and segment number are renumbered, and a connectivity table is generated. At the same time, trajectory-level continuity parameters are calculated, including total trajectory length, supplementary chain length ratio and number of breaks. The trajectory point sequence, topology number and continuity parameters are encapsulated into a conductor fusion expression result for output.

8. The online monitoring system for intelligent image recognition of transmission lines according to claim 1, characterized in that, Generating defect markers and performing local incremental updates includes: Read the current frame conductor fusion expression result, obtain the fusion trajectory point sequence, skeleton segment number, supplementary chain segment mark, coupling strength sequence and enhancement direction index, and establish a continuous detection window within the segment according to the trajectory number; The coupling strength sequence is compared between adjacent windows using a fixed window. Segments that continuously decrease and whose decrease reaches the attenuation threshold are recorded as response attenuation candidates. The breakage threshold is compared with the distance between adjacent trajectory points. Segments that trigger a breakage are recorded as interruption gap candidates, and the "repaired chain" and "unrepaired chain" states are distinguished by the supplementary chain segment. The discrete curvature is calculated using a three-point neighborhood and compared with the baseline curvature within the segment. Segments that exceed the abrupt change threshold and whose duration reaches the minimum length are recorded as curvature abrupt change candidates. The current trajectory is aligned to the trajectory cache trajectory by arc length. The normal distance is calculated at the alignment point. Segments whose normal distance continuously exceeds the offset threshold are recorded as lateral offset candidates. The defect identification process of intra-segment consensus voting uses the skeleton segment as the voting unit. It maps four types of candidate segments to the corresponding skeleton segments and counts the number of hits. Only when at least two types of candidates in the same skeleton segment hit simultaneously and one of them must be "unpaired chain interruption gap" or "lateral offset" is output as a defect segment. The output defect segment is written with the defect type set, start and end number, segment length, enhancement direction index and chain completion status to form the defect marking result. Based on the minimum bounding rectangle of the defect segment in the image coordinates, the defect update region is obtained by expanding a fixed length margin along the trajectory tangent and expanding a fixed half width along the normal. At the same time, a confidence recovery region is generated for segments that have not triggered defects but whose coupling strength is lower than the coupling threshold. The confidence recovery region is generated according to the same expansion rule. The defect update region is set to high priority and written into the two-channel calculation range of the next frame, while the confidence recovery region is set to low priority and written only into the calculation range of the structure enhancement channel. When the high priority region exists, the next frame only performs full-process recalculation of the sub-pixel skeleton channel and continuous structure enhancement channel in the high priority region, and updates the trajectory cache with the recalculation result. When only the low priority region exists, the next frame keeps the skeleton channel output unchanged and updates the structure enhancement response in the low priority region and triggers the coupling fusion verification.

Images