A reference database dynamic matching method and system for conductor double-lens icing image

Abstract

The application discloses a kind of reference database dynamic matching method and system for wire double lens icing image, including the construction of visible light lens and the visible light wire ice-free reference database of power transmission wire under night vision lens, the construction of night vision wire ice-free reference database, constructs icing wire segmentation model, for the binary segmentation mask graph of output icing image;Based on mask graph, the contour pixel set of power transmission wire under icing state is constructed to calculate the centroid pixel coordinates and comprehensive inclination of power transmission wire under icing state;The determination of calling visible light wire ice-free reference database or night vision wire ice-free reference database is carried out using the centroid pixel coordinates and comprehensive inclination of power transmission wire under icing state and ice-free state.The application solves the problem that existing single reference database cannot adapt to the difference of double lens imaging, and provides technical support for improving the accuracy of wire icing detection.

Description

Technical Field

[0001] This invention relates to the field of transmission line icing monitoring technology, and in particular to a dynamic matching method and system for a benchmark database of dual-lens icing images of conductors. Background Technology

[0002] Icing on transmission lines refers to the process by which moisture in the air condenses on the surface of the conductors and gradually accumulates to form an ice layer in low-temperature and high-humidity environments. This leads to problems such as line overload, conductor galloping, and ice flashover on insulator strings, seriously threatening the safe operation of the power grid. Currently, non-contact icing monitoring methods based on machine vision are widely used in scenarios involving the perception of icing status and thickness measurement on transmission lines due to their advantages of convenient deployment, low cost, and real-time visualization monitoring. To ensure all-weather monitoring capabilities, existing mainstream icing monitoring terminals are generally equipped with dual-lens imaging devices for the same transmission line: a visible light lens and a dedicated night vision lens. The visible light lens is suitable for imaging in normal daytime lighting conditions, while the night vision lens is suitable for imaging in nighttime, low-light, and rainy / dark-light environments. The equipment adopts a time-sharing independent working mode, with only one lens activating at a time to acquire images of the transmission line, effectively covering monitoring needs at different times of the day and significantly improving the environmental adaptability of icing monitoring.

[0003] Currently, most visual icing detection algorithms rely on a baseline database of conductors in an ice-free state to perform tasks such as pixel comparison, contour correction, and thickness calculation. However, in practical applications of dual-lens imaging devices, visible light lenses and night vision lenses have inherent hardware differences in focal length, photosensitive parameters, imaging distortion, and field of view. This means that the image scale, pixel distribution, and imaging characteristics acquired by the two lenses for the same transmission line monitoring area cannot be standardized. If a single baseline database is used to compare iced conductor images acquired by both lenses, it is impossible to effectively distinguish the imaging differences between the two types of lenses, easily leading to problems such as failed pixel comparison, icing area segmentation errors, and decreased accuracy in icing thickness calculation. Summary of the Invention

[0004] Purpose of the invention: The first objective of this invention is to provide a dynamic matching method for benchmark databases of dual-lens icing images of conductors. By constructing dedicated benchmark databases for visible light lenses and night vision lenses respectively, this method solves the problem that existing single benchmark databases cannot adapt to the imaging differences of dual lenses, thus providing technical support for improving the accuracy of conductor icing detection.

[0005] A second objective of this invention is to provide a system for a dynamic matching method of a benchmark database for dual-lens icing images of conductors.

[0006] Technical Solution: To achieve the above objectives, the present invention provides a dynamic matching method for a benchmark database of dual-lens icing images of conductors, comprising the following steps:

[0007] S1. Construct a reference database for ice-free transmission lines under visible light lenses and night vision lenses respectively. The databases include the centroid pixel coordinates and comprehensive tilt angle of the transmission lines in the ice-free state.

[0008] S2. Construct an icing conductor segmentation model. The conductor segmentation model distinguishes the icing conductor region from the background region pixel by pixel based on any icing image collected from the transmission conductor, and outputs a binary segmentation mask image of the icing image.

[0009] S3. Based on the binary segmentation mask image of the icing image, construct the contour pixel set of the transmission line in the icing state, so as to calculate the centroid pixel coordinates and comprehensive tilt angle of the transmission line in the icing state.

[0010] S4. Using the centroid pixel coordinates and overall tilt angle of the transmission line in the iced and ice-free states, determine whether to call the icy reference database for visible light conductors or the icy reference database for night vision conductors, in order to further solve the icing thickness of the transmission line.

[0011] Preferably, the method for constructing the ice-free reference database for power transmission lines under visible light lenses and night vision lenses is as follows:

[0012] For the same conductor monitoring area, several ice-free images of the transmission conductor were acquired by visible light lenses and night vision lenses respectively. A set of visible light images and a set of night vision images were established, and the best ice-free images of the transmission conductors were selected by visible light and night vision.

[0013] Based on the optimal visible light image of a icy-free conductor and the optimal night vision image of a icy-free conductor, two benchmark databases for icy-free conductors are constructed. These include: a binary mask image based on the optimal icy-free conductor image; a trapezoidal contour constructed using the pixel coordinates of the conductor region; and an circumscribed trapezoidal bounding box laterally extended from the trapezoidal contour. Within the circumscribed trapezoidal bounding box, multiple standard sampling lines are constructed according to a preset step size. The coordinate sequences of the left and right endpoints of these standard sampling lines within the circumscribed trapezoidal bounding box are extracted from near to far. Center point coordinate sequence The sequence consisting of the number of pixels occupied within the trapezoidal outline. Coordinates of the four vertices and centroid of the trapezoidal profile And based on the center point coordinate sequence Calculate the overall tilt angle of the transmission line under ice-free conditions.

[0014] Preferably, the filtering includes: removing images in which the proportion of overexposed pixels and the proportion of underexposed pixels in the effective monitoring area of ​​the guide wire are greater than the set overexposure threshold and underexposure threshold, respectively; and removing images in which the variance of pixel gradient change in the effective monitoring area of ​​the guide wire is less than the sharpness threshold.

[0015] Preferably, the ice-covered conductor segmentation model is based on U-Net, HRNet, or DeepLabv3+ semantic segmentation network. It uses historically collected real-world images of the transmission line and pixel-by-pixel binarized labels that distinguish the ice-covered conductor area from the background area as training samples, and is trained using a binary classification cross-entropy loss function as an optimization constraint.

[0016] Preferably, the method for constructing the contour pixel set of the transmission line in the icing state based on the binary segmentation mask image of the icing image is as follows: extract all closed contours in the binary segmentation mask image, and perform small-area isolated noise removal and contour edge smoothing correction processing, retain the contour with the largest area and the most complete connectivity in the image as the effective contour of the transmission line corresponding to the current image, extract the coordinates of all pixel points inside the effective contour of the transmission line, and construct the contour pixel set of the line.

[0017] Preferably, the method for calculating the centroid pixel coordinates of the transmission line under icing conditions is as follows: based on the contour pixel set, establish the zero-order moment of the effective contour region of the transmission line. And the first moment of the effective contour region in the horizontal and vertical directions. , Then the centroid pixel coordinates are:

[0018] ,

[0019] ,

[0020] in, To prevent extremely small positive numbers with a denominator of zero.

[0021] Preferably, the method for calculating the comprehensive tilt angle of the transmission line under icing conditions is as follows: extract the convex hull points of the line contour, and perform closure processing through interpolation to obtain a refined contour point set; then, cluster the refined contour point set according to the longitudinal Y-coordinate to obtain the upper edge feature point set of the line. With the feature point set of the lower edge of the guide wire ; Calculate the cluster centers for the upper and lower edge point sets respectively, and obtain the upper edge center point. and the center point of the lower edge Calculate the overall tilt angle of the transmission line under icing conditions:

[0022] .

[0023] Preferably, the method for determining whether to call the visible light conductor ice-free reference database or the night vision conductor ice-free reference database using the centroid pixel coordinates and overall tilt angle of the transmission line in iced and ice-free states is as follows: calculate the difference between the overall tilt angle of the transmission line in the iced state and the tilt angle of the transmission line in the ice-free state stored in the visible light conductor ice-free reference database and the night vision conductor ice-free reference database. , ;

[0024] Calculate the centroid pixel coordinates of the transmission line under icing conditions and the horizontal difference between these coordinates and those stored in the visible light transmission line ice-free reference database and the night vision transmission line ice-free reference database under ice-free conditions. , ,

[0025] When satisfied At that time, the ice-free reference database for visible light conductors is invoked;

[0026] When satisfied At that time, the night vision guide ice-free benchmark database is invoked;

[0027] When neither of the above two conditions is met, calculate the total error between the transmission line under icing conditions and the ice-free conditions stored in the visible light transmission line ice-free reference database and the night vision transmission line ice-free reference database. , It also calls the ice-free baseline database for conductors with small total error values.

[0028] Preferably, based on the accessed visible light conductor ice-free reference database or night vision conductor ice-free reference database, the ice thickness of the transmission conductor is further calculated, including: using the coordinates of the four vertices of the trapezoidal contour stored in the database to define the reference geometric range of the conductor in the current conductor ice image to be detected; using the coordinate sequence of the left and right endpoints of the standard sampling line... Multiple sampling lines are constructed within the current image of the icing on the conductor to be detected; the center point coordinate sequence of the standard sampling lines is used. Constraints on the conductor center trajectory and verification of sampling line positions in the current image of conductor icing to be detected; utilize the pixel count sequence of standard sampling lines. The thickness of the ice accretion on the conductor is determined by comparing the actual number of pixels in each sampling line of the current image of the conductor icing to be detected.

[0029] The present invention provides a dynamic matching system for a benchmark database of dual-lens icing images of conductors, comprising the following modules:

[0030] Dual-lens reference database construction module: used to construct the ice-free reference database for power transmission lines under visible light lens and night vision lens respectively. The database includes the centroid pixel coordinates and comprehensive tilt angle of the power transmission line in the ice-free state.

[0031] Ice-covered conductor segmentation model generation module: used to construct an ice-covered conductor segmentation model. The conductor segmentation model distinguishes the ice-covered conductor region from the background region pixel by pixel based on any ice-covered image collected from the transmission conductor, and outputs a binary segmentation mask image of the ice-covered image.

[0032] The icing image conductor parameter calculation module is used to construct the contour pixel set of the transmission conductor in the icing state based on the binary segmentation mask image of the icing image, so as to calculate the centroid pixel coordinates and comprehensive tilt angle of the transmission conductor in the icing state.

[0033] Database call determination module: It is used to determine whether to call the visible light conductor ice-free reference database or the night vision conductor ice-free reference database by using the centroid pixel coordinates and comprehensive tilt angle of the transmission line in the iced and ice-free states, so as to further solve the ice thickness of the transmission line.

[0034] Beneficial Effects: This invention has the following advantages: It can establish a reference database for ice-free conductors based on the imaging differences between visible light lenses and night vision lenses. By combining the geometric features of the conductor, such as the perspective tilt angle and centroid position deviation, it performs geometric consistency matching between the real-time acquired icy conductor images and the reference data of the ice-free conductors. This allows the selection of the reference database that best matches the current icy conductor image, solving the problem of reference matching misalignment caused by differences in the hardware parameters of the two lenses. It can also reduce subsequent icing area segmentation deviation and thickness calculation errors, providing technical support for improving the accuracy of conductor icing detection. Attached Figure Description

[0035] Figure 1 Flowchart of this invention;

[0036] Figure 2 Images of ice-free conductors captured by visible light and night vision lenses in the same monitoring area;

[0037] Figure 3 For based on Figure 2 The calculated inclination angle of the conductor. Detailed Implementation

[0038] The technical solution of the present invention will be described in detail below with reference to the embodiments and accompanying drawings.

[0039] The present invention provides a dynamic matching method for a benchmark database of dual-lens icing images of conductors, such as... Figure 1 As shown, it includes the following:

[0040] 1. Construct ice-free benchmark databases for power transmission lines under the visible light lens and night vision lens of the monitoring device, respectively.

[0041] 1. For the same conductor monitoring area, acquire several ice-free images of the transmission conductors collected by visible light lenses and night vision lenses respectively, and establish a set of visible light images and a set of night vision images.

[0042] The visible light lens and night vision lens used in this invention are integrated and installed on the same ice-observation monitoring device. The two lenses are fixedly arranged side by side, both pointing at the same target power transmission line monitoring area, ensuring that the imaging field of view of the two lenses is consistent from the hardware installation level. The images collected by the ice-observation monitoring device can be transmitted to the back-end server for data processing using a network bridge data transmission method, realizing real-time monitoring of the power transmission line around the clock.

[0043] 2. Select at least one optimal ice-free image of the transmission line from the visible light image set and the night vision image set respectively, to obtain the optimal ice-free image of the transmission line under visible light and the optimal ice-free image under night vision.

[0044] To adapt to the different imaging characteristics of the two types of lenses and avoid the limitations of a single screening rule, this invention employs a three-level differentiated screening mechanism: machine-based quantitative initial screening, local feature verification, and manual fine screening, to screen the visible light image set and the night view image set. The specific steps are as follows:

[0045] 2.1 Initial Screening via Image Quantization and Brightness Equalization. This invention employs a local area grayscale statistical mechanism, extracting only the effective monitoring area of ​​the conductor from the center of the image for grayscale analysis, thus avoiding interference from invalid background pixels. The image grayscale value range is defined as follows: Statistical analysis of grayscale values ​​within the effective monitoring area Number of overexposed pixels and grayscale value The number of underexposed pixels was calculated, and the percentage of overexposed pixels was calculated separately. Underexposed pixel ratio The calculation formula is:

[0046]

[0047] In the formula: This represents the number of overexposed pixels in a local effective area. This refers to the number of underexposed pixels. This represents the total number of pixels in the locally effective area. The invention presets an overexposure threshold. Underexposure threshold If a certain image or If the image is found to have abnormal lighting or imaging failure, it will be rejected.

[0048] 2.2 Initial Screening Based on Image Sharpness Gradient Verification. The image spatial gradient variance evaluation method is introduced to calculate image sharpness by statistically analyzing the variance of pixel gradient changes in the effective monitoring area. It characterizes the sharpness of the image. When the gradient variance of an image is less than the preset sharpness threshold, the image is judged to be blurry, jittery, or out of focus, and is removed to ensure that the edges and textures of the remaining images are clear.

[0049] 2.3. Refined Manual Screening. A second, precise verification is performed on the candidate images after the initial machine screening. The screening criteria are adapted to the dual-lens baseline database requirements: First, the conductor body, hardware, and insulators are in a completely unloaded, bare wire state, free from ice, snow, and any frost or other deposits. Second, the conductor body is fully centered, without image cutoff, offset, jitter, or distortion, and without partial obstruction by birds, branches, shadows, or debris. Third, the dual-lens imaging characteristics are distinguished: visible light images show no strong light color cast, night vision images show no excessive noise or washed-out distortion, and the equipment's shooting tilt angle and field of view are regular and without abnormalities.

[0050] 3. Based on the best visible light images of ice-free conductors and the best night vision images of ice-free conductors, respectively, construct a visible light benchmark database and a night vision benchmark database for ice-free conductors.

[0051] This invention uses the Labelme open-source annotation software to annotate optimal ice-free images of power transmission lines under both visible light and night vision conditions. The annotation process includes: marking the main body of the power transmission line as the foreground region, labeled "Line"; and uniformly marking all remaining areas in the image, including mountains, sky, towers, vegetation, ground, and other irrelevant areas, as the background region, labeled "Background". This eliminates the need for further subdivision of secondary environmental targets, simplifying the annotation process and reducing redundant feature interference. During annotation, a polygonal outline method is used to fit and select points along the outline of the power transmission line. After annotation, a standard PNG format label mask image with the same resolution and size as the original image is exported and generated. This invention uses a fixed color fill rule to complete the mask rendering: the foreground region labeled "Line" is uniformly filled with pure white (pixel value 255), and the background region labeled "background" is uniformly filled with pure black (pixel value 0), forming a high-contrast, standardized binary mask image.

[0052] Based on each binary mask image, traverse all foreground region pixels and determine the top-left vertex of the trapezoid according to the extreme values ​​of the horizontal and vertical pixel coordinates. Top right vertex bottom left vertex bottom right vertex After determining the coordinates of the four feature vertices, a standard trapezoidal contour with a near-wide and far-narrow shape, adapted to the perspective characteristics of the traverse, is reconstructed by connecting these four points in an ordered manner. Simultaneously, the image moment algorithm is used to calculate the stable horizontal and vertical coordinates of the contour centroid, i.e., the centroid coordinates of the trapezoidal contour, based on the contour space moment parameters. .

[0053] Furthermore, the vertical height of the trapezoidal profile is calculated. This height is the effective pixel distance between the top and bottom edges of the trapezoid in the vertical direction of the image. The total height of the trapezoid is... Divide the data into 20 equal parts and define the step size of a single vertical translation. for:

[0054]

[0055] Using the lower base of the trapezoidal profile as the initial baseline, and following a fixed step size... The trapezoid is vertically shifted upwards 20 times sequentially, forming 20 initial sampling lines parallel to the base of the trapezoid. These 20 initial sampling lines are evenly distributed along the perspective height of the trapezoid, fully covering the effective area of ​​the guide wire. The trapezoidal outline is then optimized by expanding its entire outer edge: the lengths of the upper and lower bases are simultaneously expanded horizontally to five times their original lengths, while maintaining the inclination angles and perspective slope of the two sides of the trapezoid. This generates an extrinsic reference trapezoidal frame with the same shape as the original trapezoid but significantly expanded in size. This expansion method ensures sufficient horizontal length for all initial sampling lines, avoiding the problem of excessively short sampling lines and insufficient effective sampling area after subsequent line cutting.

[0056] The 20 initial sampling lines are truncated using the left and right side boundaries of the circumscribed trapezoidal frame, retaining only the segments within the frame, thus obtaining 20 standard sampling lines. The pixel coordinates of the left and right endpoints of each of the 20 standard sampling lines are then extracted, from far to near, to form a sequence of left and right endpoint coordinates for each standard sampling line. .

[0057] The mean coordinates of the left and right endpoints of each standard sampling line are calculated to obtain the center pixel coordinates of a single standard sampling line. The calculation formula is as follows:

[0058]

[0059] In the formula: , The first The coordinates of the left and right endpoints of the sampling line The coordinates of the midpoint of the corresponding sampling line. =1~20. Following the perspective order from far to near, obtain the center point coordinate sequence of 20 standard sampling lines. It is used for subsequent conductor center trajectory fitting, icing eccentricity correction and region alignment verification.

[0060] The pixel width at which each standard sampling line intersects the original trapezoidal contour is counted, serving as the baseline pixel count for each standard sampling line in the ice-free state. Following a perspective order from far to near, a sequence is obtained for the baseline pixel counts of the 20 standard sampling lines in each binary mask image in the ice-free state. .

[0061] Construct a visible light guideline ice-free benchmark database and a night vision guideline ice-free benchmark database, respectively, corresponding to the optimal visible light guideline ice-free image and the optimal night vision guideline ice-free image. The database includes a sequence of the number of benchmark pixels for 20 standard sampling lines in a single image under ice-free conditions. 20 standard sampling line center point coordinate sequence The coordinate sequence of the left and right endpoints of 20 standard sampling lines Coordinates of the four vertices and centroid of the trapezoidal profile And the overall tilt angle of the conductor.

[0062] like Figure 2 As shown, the images were taken by a visible light lens and a night vision lens for the same monitoring area. If the trapezoidal contour of the left image is directly used in the right image, there will be a deviation. Therefore, it is necessary to construct trapezoidal contours separately for images taken by different lenses.

[0063] 2. Construct an icing conductor segmentation model. The conductor segmentation model distinguishes the icing conductor region from the background region pixel by pixel in the icing image of the transmission conductor and outputs a binary segmentation mask image of the icing image.

[0064] The training samples are real-world images of power transmission lines historically collected by visible light lenses and night vision lenses. They include daytime strong light and normal natural light imaging samples collected by visible light lenses, as well as nighttime no light source and low light imaging samples collected by night vision lenses. At the same time, they comprehensively cover various complex weather scenarios such as sunny days, cloudy days, foggy days, rainy and snowy days, and light winds with icing, covering power transmission line images with different light intensities, different ice thicknesses, and different ice adhesion patterns.

[0065] The real-world images underwent preprocessing, including removing invalid samples that were blurry, abnormally exposed, severely distorted, or heavily occluded, retaining 5000-8000 high-quality, valid images with complete images, clear target features, and no obvious interference. The selected real-world images were then subjected to pixel-level fine-grained annotation using the Labelme open-source annotation tool. This included defining the main area of ​​the power transmission lines covered with ice or frost as the foreground target, labeled "Line"; and defining all non-power transmission line areas, such as mountains, vegetation, sky, towers, ground, birds, debris, and shadows, as the background area, labeled "Background". After annotation, a binary PNG mask label image with the exact same resolution and size as the original image was generated. Finally, the entire dataset was randomly divided according to the general optimal ratio of 7:2:1. 70% of the samples were used as the training set for iterative learning of model parameters and feature fitting; 20% of the samples were used as the validation set for hyperparameter tuning, real-time verification of model accuracy, and overfitting monitoring during training; and 10% of the samples were used as the independent test set for final performance evaluation and generalization ability detection after the model training was completed.

[0066] The segmentation model uses classic semantic segmentation networks such as U-Net, HRNet, and DeepLabv3+, which are adapted for segmenting long and thin targets, as the base model. During training, the core training objective is to distinguish between the icy guide wire region and the background region pixel by pixel. A binary cross-entropy loss function is introduced to construct the model optimization constraints. Iterative training is carried out based on the dataset. The forward propagation of the network outputs pixel-level prediction results, and the backpropagation algorithm is used to continuously correct and update the various weight parameters inside the network. The model continuously fits and learns the morphology, contour, and texture features of the icy guide wire until the segmentation accuracy on the validation set tends to be stable and the loss value converges. Finally, the model weight file with the best overall performance during training is retained to complete the construction of a dedicated semantic segmentation model for icy guide wires. This model can accurately extract the target region of icy guide wires in complex outdoor backgrounds and output a standard binary segmentation mask map, providing a reliable image foundation for subsequent sampling line layout and pixel statistics.

[0067] 3. Based on the binary segmentation mask image of the icing image, construct the contour pixel set of the transmission line in the icing state, so as to calculate the centroid pixel coordinates and comprehensive tilt angle of the transmission line in the icing state.

[0068] 1. Based on the real-time acquisition of icing images of power transmission lines by the ice-observing monitoring device, a binary segmentation mask image of the icing images of the power transmission lines is obtained using an icing-covered power line segmentation model.

[0069] 2. Based on the binary segmentation mask image of the icing image of the conductor, construct a set of pixels for the conductor outline to calculate the centroid pixel coordinates of the effective outline of the transmission conductor.

[0070] A contour-finding algorithm is used to extract all white closed contours with a pixel value of 255 from the binary segmentation mask image. Small-area isolated noise is removed, and contour edge smoothing is performed. The target contour with the largest area and most complete connectivity in the image is retained as the effective contour of the power transmission line corresponding to the current frame image. The coordinates of all foreground pixels within the effective contour of the power transmission line are extracted to construct a pixel set of the power transmission line contour. Based on this pixel set, the image moments of the effective contour region of the power transmission line are calculated to solve for the centroid pixel coordinates of the current power transmission line contour.

[0071] The specific calculation process is as follows: Let the pixel normalization value of the current binary segmentation mask be... ,in, Represents the pixel coordinates of the image. If the pixel point... The white foreground area that is within the effective outline of the transmission line is then... If pixel If it belongs to the background area, then let When the white foreground pixel value in the original binary mask image is 255, it can be normalized to 1 first.

[0072] The set of pixels for the conductor outline is as follows, where Number of pixels in the binary segmentation mask:

[0073]

[0074] The zeroth moment of the effective contour region is expressed as:

[0075] in, This represents the total number of foreground pixels within the effective contour region of the transmission line, i.e., the pixel area of ​​this contour region. Further calculations are made of the first moments of the effective contour region in the horizontal and vertical coordinate directions:

[0076]

[0077]

[0078] in, This represents the weighted sum of the foreground pixels along the horizontal axis. This represents the weighted sum of the foreground pixels along the vertical axis. Based on the zeroth and first moments described above, the centroid pixel coordinates of the effective profile of the transmission line can be expressed as:

[0079]

[0080]

[0081] in, This represents the centroid pixel coordinates of the effective contour of the transmission line in the current frame image. A very small positive number is set to prevent the denominator from being zero. Since this centroid coordinate is calculated from the spatial distribution of all foreground pixels within the effective contour of the transmission line, it can characterize the center position of the target line in the image plane under the current lens imaging conditions.

[0082] 3. Based on the binary segmentation mask image of the conductor icing image, calculate the overall tilt angle of the transmission conductor. It is used to characterize the spatial attitude changes of transmission lines under icing conditions.

[0083] From the set of wire outline pixels in the binary segmentation mask image, the convex hull point set is extracted using the convex hull algorithm, and the convex hull outline is uniformly interpolated and densified by a preset interpolation step size to obtain a dense and uniform refined outline point set. The specific process is as follows.

[0084] The set of points of the convex hull of the contour is:

[0085] in, This indicates the calculation of the convex hull. This represents the set of convex hull points generated from the set of points on the conductor profile. Denotes the first convex hull boundary. One vertex, This represents the number of vertices in the convex hull. Convex hull operations can extract the main boundary features outside the wire contour, reducing the impact of local depressions, edge burrs, and irregular contour segmentation on subsequent geometric feature extraction.

[0086] After obtaining the convex hull set, the convex hull boundary is closed. Specifically, adjacent vertices in the convex hull set are connected sequentially, and the last convex hull vertex is closed. With the first convex hull vertex They connect to form a closed convex hull profile.

[0087] Let the preset interpolation step size be... For any two adjacent convex hull vertices and Calculate the Euclidean distance between the two:

[0088]

[0089] Among them, when season This ensures that the convex hull boundary is closed at both ends.

[0090] If the distance between adjacent vertices of the convex hull Less than or equal to the preset interpolation step size If so, no new interpolation point is added on that edge; only the original convex hull vertices are retained. The number of interpolation points is then determined based on the side length and the interpolation step size.

[0091]

[0092] in, Indicates the first The number of points to be inserted on the edge of the convex hull. This indicates rounding down to the nearest integer.

[0093] Subsequently, at adjacent convex hull vertices and Linear interpolation is performed between them. The coordinates of the interpolation points are represented as follows:

[0094]

[0095] Right now: ,

[0096] in, Indicates the first The first convex edge Interpolation points.

[0097] Following the method described above, all edges of the convex hull contour are traversed sequentially. While retaining the original convex hull vertices, the interpolated points generated by each edge are added to the point set in order, finally obtaining the refined convex hull contour point set, denoted as the refined contour point set:

[0098]

[0099] in, This represents the refined contour point set after interpolation and encryption. This indicates the number of points in the interpolated point set. Compared to the original convex hull point set, the points in the refined contour point set are more densely and evenly distributed along the convex hull boundary, providing more stable contour geometric features for subsequent guideline contour direction angle extraction, geometric consistency judgment, and benchmark database matching.

[0100] Subsequently, the DBSCAN clustering algorithm was used to cluster and segment the refined contour point set according to the vertical Y-coordinate to obtain the feature point set of the guide rail edge. With the feature point set of the lower edge of the guide wire Cluster centers are calculated for the upper and lower edge point sets respectively, resulting in the upper edge center point. and the center point of the lower edge A central line is constructed by connecting two points to represent the overall tilt trend of the perspective trapezoidal traverse. Finally, the real-time composite tilt angle of the traverse is calculated based on the central line. The calculation formula is:

[0101] .

[0102] IV. Based on the composite tilt angle of the conductor in the conductor icing image and centroid coordinates The system then checks whether the wire is ice-free based on the reference database.

[0103] Using the coordinate sequences of the midpoints of the standard sampled lines stored in the visible light and night vision guideline ice-free reference databases, the reference tilt angle of the guideline in the optimal visible light and night vision guideline ice-free images was calculated, respectively. ,inclination Specifically, it involves setting the coordinates of the center points of the 20 standard sampling lines. Connect the points with a straight line, and calculate the inclination angle using the above formula, where... This is the center point of the farthest end of the conductor, corresponding to the center point of the upper edge. This is the center point of the nearest end of the conductor, corresponding to the center point of the lower edge.

[0104] Further calculations were performed on the tilt angle deviation and lateral position deviation of the conductor in the icy conductor image, the optimal visible light image of the icy-free conductor, and the optimal night vision image of the icy-free conductor:

[0105] ① Inclination deviation:

[0106]

[0107] ② Lateral position deviation:

[0108]

[0109] The smaller the tilt angle deviation, the higher the consistency between the current perspective tilt shape of the traverse and the reference imaging shape of the acquisition lens; the smaller the horizontal axis offset difference, the higher the fit between the current imaging position of the traverse and the fixed field of view of the acquisition lens.

[0110] When satisfied At that time, the ice-free reference database for visible light conductors is invoked;

[0111] When satisfied At that time, the night vision guide ice-free benchmark database is invoked;

[0112] When neither of the above two conditions is met, calculate the total error. , The benchmark database with the smaller total error is selected as the final matching result. For example, when the real-time tilt angle... center of mass Reference A , Reference B , hour, , , , Total error Finally, it was matched to the ice-free benchmark database for visible light conductors.

[0113] This invention effectively solves the problem of misjudgment caused by icing morphology, image noise, and equipment shake interference in single feature discrimination by constraining tilt angle deviation and lateral position deviation. It significantly improves the accuracy and stability of dual-lens type recognition and enables accurate and adaptive switching and calling of the benchmark database under complex working conditions in all weather.

[0114] like Figure 3 The figure shows the actual power transmission line (tilt angle 1 in the figure), and the tilt angle of the power transmission line under a visible light lens (tilt angle 2 in the figure) and a night vision lens (tilt angle 3 in the figure).

[0115] V. Based on the accessed baseline database of ice-free conductors, analyze the icing situation of transmission conductors.

[0116] The database includes a sequence of reference pixel counts for 20 standard sampling lines in an ice-free state. 20 standard sampling line center point coordinate sequence The coordinate sequence of the left and right endpoints of 20 standard sampling lines The coordinates of the four vertices of the trapezoidal profile.

[0117] Among them, the coordinates of the four vertices of the trapezoidal contour are used to define the reference geometric range of the conductor in the current conductor icing image to be detected; the coordinate sequence of the left and right endpoints of the 20 standard sampling lines are used to constrain the construction of 20 sampling lines in the current conductor icing image to be detected; the coordinates of the center point of the 20 standard sampling lines are used to constrain the center trajectory of the conductor in the current conductor icing image to be detected and to verify the position of the sampling lines; the reference pixel number sequence of the 20 standard sampling lines in the ice-free state is used to compare with the actual pixel number of each sampling line in the current conductor icing image to be detected, and further establish the pixel-physical scale conversion relationship for each sampling line, which is used for subsequent calculation of the major diameter of icing and the visible icing thickness.

Claims

1. A dynamic matching method for a benchmark database of dual-lens icing images of conductors, characterized in that, Includes the following steps: S1. Construct a reference database for ice-free transmission lines under visible light lenses and night vision lenses respectively. The databases include the centroid pixel coordinates and comprehensive tilt angle of the transmission lines in the ice-free state. S2. Construct an icing conductor segmentation model. The conductor segmentation model distinguishes the icing conductor region from the background region pixel by pixel based on any icing image collected from the transmission conductor, and outputs a binary segmentation mask image of the icing image. S3. Based on the binary segmentation mask image of the icing image, construct the contour pixel set of the transmission line in the icing state, so as to calculate the centroid pixel coordinates and comprehensive tilt angle of the transmission line in the icing state. S4. Using the centroid pixel coordinates and overall tilt angle of the transmission line in the icing and non-icing states, determine whether to call the icing reference database for visible light transmission lines or the icing reference database for night vision transmission lines, so as to further solve the icing thickness of the transmission line. Based on the binary segmentation mask image of the icing image, the method for constructing the contour pixel set of the transmission line in the icing state is as follows: extract all closed contours in the binary segmentation mask image, and perform small-area isolated noise removal and contour edge smoothing correction processing, retain the contour with the largest area and the most complete connectivity in the image as the effective contour of the transmission line corresponding to the current image, extract the coordinates of all pixel points inside the effective contour of the transmission line, and construct the contour pixel set of the line. The method for calculating the centroid pixel coordinates of the transmission line under icing conditions is as follows: based on the contour pixel set, establish the zero-order moments of the effective contour region of the transmission line. And the first moment of the effective contour region in the horizontal and vertical directions. , Then the centroid pixel coordinates are: ， ， in, To prevent extremely small positive numbers with a denominator of zero; The method for calculating the comprehensive tilt angle of the transmission line under icing conditions is as follows: extract the convex hull points of the line contour, and perform closure processing through interpolation to obtain a refined contour point set. Then, cluster the refined contour point set according to the longitudinal Y-coordinate to obtain the upper edge feature point set of the line. With the feature point set of the lower edge of the guide wire ; Calculate the cluster centers for the upper and lower edge point sets respectively, and obtain the upper edge center point. and the center point of the lower edge Calculate the overall tilt angle of the transmission line under icing conditions: 。 2. The method for dynamic matching of a reference database for dual-lens icing images of conductors according to claim 1, characterized in that, The method for constructing the ice-free benchmark database for power transmission lines under visible light and night vision lenses is as follows: For the same conductor monitoring area, several ice-free images of the transmission conductor were acquired by visible light lenses and night vision lenses respectively. A set of visible light images and a set of night vision images were established, and the best ice-free images of the transmission conductors were selected by visible light and night vision. Based on the optimal visible light image of a icy-free conductor and the optimal night vision image of a icy-free conductor, two benchmark databases for icy-free conductors are constructed. These include: a binary mask image based on the optimal icy-free conductor image; a trapezoidal contour constructed using the pixel coordinates of the conductor region; and an circumscribed trapezoidal bounding box laterally extended from the trapezoidal contour. Within the circumscribed trapezoidal bounding box, multiple standard sampling lines are constructed according to a preset step size. The coordinate sequences of the left and right endpoints of these standard sampling lines within the circumscribed trapezoidal bounding box are extracted from near to far. Center point coordinate sequence The sequence consisting of the number of pixels occupied within the trapezoidal outline. Coordinates of the four vertices and centroid of the trapezoidal profile And based on the center point coordinate sequence Calculate the overall tilt angle of the transmission line under ice-free conditions.

3. The method for dynamic matching of a reference database for dual-lens icing images of conductors according to claim 2, characterized in that, The filtering includes: removing images in which the proportion of overexposed pixels and underexposed pixels in the effective monitoring area of ​​the guide wire are greater than the set overexposure threshold and underexposure threshold, respectively; and removing images in which the variance of pixel gradient change in the effective monitoring area of ​​the guide wire is less than the sharpness threshold.

4. The method for dynamic matching of a reference database for dual-lens icing images of conductors according to claim 1, characterized in that, The ice-covered conductor segmentation model is based on U-Net, HRNet, or DeepLabv3+ semantic segmentation network. It uses historically collected real-world images of transmission lines and pixel-by-pixel binarized labels that distinguish the ice-covered conductor area from the background area as training samples, and is trained using a binary classification cross-entropy loss function as the optimization constraint.

5. The method for dynamic matching of a reference database for dual-lens icing images of conductors according to claim 2, characterized in that, The method for determining whether to call the visible light conductor ice-free reference database or the night vision conductor ice-free reference database using the centroid pixel coordinates and overall tilt angle of the transmission line in iced and ice-free states is as follows: The overall tilt angle of the transmission line in the iced state is calculated, and the difference between this tilt angle and the tilt angle of the transmission line in the ice-free state stored in the visible light conductor ice-free reference database and the night vision conductor ice-free reference database is calculated. , ; Calculate the centroid pixel coordinates of the transmission line under icing conditions and the horizontal difference between these coordinates and those stored in the visible light transmission line ice-free reference database and the night vision transmission line ice-free reference database under ice-free conditions. , , When satisfied At that time, the ice-free reference database for visible light conductors is invoked; When satisfied At that time, the night vision guide ice-free benchmark database is invoked; When neither of the above two conditions is met, calculate the total error between the transmission line under icing conditions and the ice-free conditions stored in the visible light transmission line ice-free reference database and the night vision transmission line ice-free reference database. , It also calls the ice-free baseline database for conductors with small total error values.

6. The method for dynamic matching of a reference database for dual-lens icing images of conductors according to claim 2, characterized in that, Based on the accessed visible light conductor ice-free reference database or night vision conductor ice-free reference database, the ice thickness of the transmission conductor is further calculated, including: using the coordinates of the four vertices of the trapezoidal contour stored in the database to define the reference geometric range of the conductor in the current conductor ice-covered image to be detected; and using the coordinate sequence of the left and right endpoints of the standard sampling line. Multiple sampling lines are constructed within the current image of the icing on the conductor to be detected; the center point coordinate sequence of the standard sampling lines is used. Constraints on the conductor center trajectory and verification of sampling line positions in the current image of conductor icing to be detected; utilize the pixel count sequence of standard sampling lines. The thickness of the ice accretion on the conductor is determined by comparing the actual number of pixels in each sampling line of the current image of the conductor icing to be detected.

7. A dynamic matching system for a benchmark database of dual-lens icing images of conductors, applied to the dynamic matching method for a benchmark database of dual-lens icing images of conductors as described in claim 1, characterized in that, Includes the following modules: Dual-lens reference database construction module: used to construct the ice-free reference database for power transmission lines under visible light lens and night vision lens respectively. The database includes the centroid pixel coordinates and comprehensive tilt angle of the power transmission line in the ice-free state. Ice-covered conductor segmentation model generation module: used to construct an ice-covered conductor segmentation model. The conductor segmentation model distinguishes the ice-covered conductor region from the background region pixel by pixel based on any ice-covered image collected from the transmission conductor, and outputs a binary segmentation mask image of the ice-covered image. The icing image conductor parameter calculation module is used to construct the contour pixel set of the transmission conductor in the icing state based on the binary segmentation mask image of the icing image, so as to calculate the centroid pixel coordinates and comprehensive tilt angle of the transmission conductor in the icing state. Database call determination module: It is used to determine whether to call the visible light conductor ice-free reference database or the night vision conductor ice-free reference database by using the centroid pixel coordinates and comprehensive tilt angle of the transmission line in the iced and ice-free states, so as to further solve the ice thickness of the transmission line.

Images