Discharge channel extraction method and system based on edge detection and texture feature fusion

Abstract

The invention discloses a discharge channel extraction method and system based on edge detection and texture feature fusion, and belongs to the technical field of power equipment fault diagnosis and digital image processing, and the method comprises the steps: obtaining an ultraviolet weak light image of power equipment, and carrying out the preprocessing; multi-scale edge detection is executed based on the preprocessed image, edge information is fused through a multi-scale voting mechanism, and a candidate edge graph is generated by combining direction consistency connection fracture edges; extracting texture features of the pre-processed image, wherein the texture features comprise a rotation invariant local binary pattern uniformity feature and a multi-direction gray level co-occurrence matrix feature; according to a dynamic distribution weight coefficient of a discharge type, carrying out weighted fusion on the edge intensity of the candidate edge graph and the texture features, and generating a discharge channel confidence score graph; and carrying out binarization segmentation and contour optimization processing on the confidence score graph of the discharge channel, and extracting morphology quantization parameters of the discharge channel.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment fault diagnosis and digital image processing. More specifically, it relates to a method and system for intelligent extraction of discharge channels based on multi-scale edge detection and dynamic texture feature fusion. The method and system are particularly suitable for accurately extracting the contours, morphology and dynamic evolution characteristics of discharge channels caused by external insulation defects in high-voltage equipment from ultraviolet (UVA) weak-light imaging data, providing technical support for intelligent diagnosis of discharge faults. Background Art

[0002] Partial discharge (PD) caused by defects in the external insulation of high-voltage electrical equipment is a significant threat to the safe operation of power grids. Ultraviolet imaging technology captures the weak photon radiation generated by this discharge, providing a visual means for assessing insulation status. Currently, the industry relies primarily on manual interpretation of UV images, which suffers from low efficiency and poor consistency. While previous studies have attempted to improve detection accuracy through multi-scale filtering or feature fusion, balancing noise suppression with detail preservation and achieving accurate channel extraction in low-light environments remains a technical challenge that urgently needs to be overcome.

[0003] The existing technology has the following problems:

[0004] ① In ultraviolet low-light imaging, the discharge channel signal is weak (the signal-to-noise ratio is often less than 10dB). Traditional edge detection methods are susceptible to random noise interference, resulting in broken channel contours or an increase in false edges. Single-scale edge detection is difficult to balance the integrity of details and the main channel.

[0005] Second, the texture characteristics of discharge channels have not been fully explored. Existing methods rely on grayscale threshold segmentation, resulting in over 30% missed detection rates in low-contrast areas (such as the discharge starting point). The grayscale co-occurrence matrix features commonly used in the literature are not dynamically integrated with edge responses, leading to channel misjudgment in complex backgrounds.

[0006] ③ The expansion path of the discharge channel in the time dimension lacks a quantitative description. Traditional methods only output static contours, which makes it difficult to support timing analysis of the discharge stage.

[0007] Prior art document 1 (CN113191313A) discloses a method for video stream discharge identification in the field of power equipment monitoring. This method narrows the monitoring range and reduces the amount of image processing by determining the foreground mask of the current image in real-time video data, binarizing the foreground mask to locate highlighted pixels, and generating discharge area location information based on coordinate clustering. Its shortcomings are: 1. Relying on a single binary highlighted pixel to determine the discharge area, resulting in a high rate of missed detection of discharge channels in low-contrast or complex backgrounds; 2. Using simple threshold segmentation and foreground mask processing, it is susceptible to random noise interference, resulting in an increase in pseudo-edges or broken contours; 3. Locating the discharge area solely through coordinate clustering cannot support the study of the dynamic evolution of discharges.

[0008] Prior art document 2 (CN110533064A) discloses a partial discharge pattern recognition method for power equipment fault diagnosis. This method preprocesses PRPS / PRPD patterns, extracts statistical and morphological features, and then uses a neural network to classify and output defect types. Its shortcomings include: 1. Relying on static feature extraction, it fails to capture detailed information about discharge channels in low-light environments; 2. The preprocessing is not optimized for the high noise characteristics of ultraviolet (UV) low-light images, making feature extraction susceptible to interference; and 3. It only outputs defect type classification results, lacking support for time-series research on discharge development paths. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this paper proposes an intelligent discharge channel extraction method based on edge detection and texture feature fusion to address the noise sensitivity, contour fracture, and morphology distortion problems of discharge channel extraction in ultraviolet weak light images. Through dynamic weight allocation, multimodal feature fusion, and morphology optimization constraints, this method achieves high-precision and robust extraction of discharge channels and provides a data foundation for quantitative analysis of discharge development paths. Specifically, this method includes the following:

[0010] The present invention adopts the following technical solutions.

[0011] A first aspect of the present invention provides a discharge channel extraction method based on edge detection and texture feature fusion, comprising the following steps:

[0012] Acquire ultraviolet weak light images of power equipment and perform preprocessing;

[0013] Perform multi-scale edge detection based on the preprocessed image, fuse edge information through a multi-scale voting mechanism, and combine directional consistency to connect broken edges to generate a candidate edge map;

[0014] Extracting texture features of the preprocessed image, including rotation-invariant local binary pattern uniformity features and multi-directional gray-level co-occurrence matrix features;

[0015] Dynamically assigning weight coefficients according to discharge types, weightedly fusing the edge strength of the candidate edge map with the texture features to generate a discharge channel confidence score map;

[0016] The discharge channel confidence score map is subjected to binary segmentation and contour optimization processing, and the morphology of the discharge channel is extracted, and a parameterized discharge channel contour and a quantitative report are output.

[0017] Optionally, the preprocessing of the ultraviolet weak light image includes:

[0018] The non-local means algorithm is used to denoise the original image, where:

[0019] Use integral image method to accelerate similarity calculation of pixel blocks of preset size;

[0020] The weights are calculated based on the exponential decay function of the Euclidean distance of the pixel blocks, and the denoised image is subjected to the limited contrast histogram equalization process with the set attenuation factor.

[0021] Set blocks of preset size according to the average width of the discharge channel.

[0022] Optionally, generating a candidate edge graph includes:

[0023] The gradient magnitude and gradient direction of the preprocessed image are calculated under three preset Gaussian kernel scales;

[0024] Performing non-maximum suppression on the gradient field at each scale, and dynamically setting a dual threshold for pixel segmentation based on the maximum gradient amplitude at the current scale, wherein the dual threshold is set based on the maximum gradient amplitude according to a preset ratio;

[0025] For each pixel, if it is judged as an edge pixel at at least two different scales, it is marked as a candidate edge;

[0026] Traversing adjacent broken edge segments in the candidate edge map, connecting them if the pixel distance between the broken endpoints is less than a preset distance threshold and the gradient direction difference between the two endpoints is less than a preset angle threshold;

[0027] Output a continuous candidate edge map.

[0028] Optionally, extracting texture features of the preprocessed image includes:

[0029] calculating a rotationally invariant local binary pattern feature of the preprocessed image;

[0030] Counting the proportion of uniform pattern pixels in the rotation-invariant local binary pattern feature as a texture uniformity feature;

[0031] Calculating the gray level co-occurrence matrix of the pre-processed image in a plurality of preset directions;

[0032] Calculating the contrast value and entropy value of the gray level co-occurrence matrix in each preset direction;

[0033] The average value of the contrast values ​​in the four preset directions is taken as the comprehensive contrast index;

[0034] Take the average of the entropy values ​​in the four preset directions as the comprehensive entropy index;

[0035] The comprehensive contrast index and the comprehensive entropy index together constitute the multi-directional gray-level co-occurrence matrix feature.

[0036] Optionally, generating a discharge channel confidence score map includes:

[0037] Automatically identifying the discharge type based on the global gradient amplitude statistical characteristics and global gray-level co-occurrence matrix entropy characteristics of the preprocessed image;

[0038] According to the identified discharge type, assigning corresponding first weight coefficient α, second weight coefficient β and third weight coefficient γ to the edge strength, texture uniformity feature and texture complexity index based on a preset weight assignment rule;

[0039] For each pixel position in the candidate edge map, perform the following operations:

[0040] Multiply the normalized edge intensity by α, the normalized texture uniformity eigenvalue by β, and the normalized texture complexity index by γ;

[0041] The confidence score of the pixel position is obtained by summing the three product results;

[0042] Output the discharge channel confidence score map consisting of the confidence scores of all pixel positions.

[0043] Optionally, the automatically identifying the discharge type includes:

[0044] Calculate the mean μ and standard deviation σ of the gradient amplitude of the entire image;

[0045] If the gradient amplitude of a certain area is greater than the linear combination value of μ and σ and the texture uniformity characteristic value of the area is less than the preset threshold, it is determined to be corona discharge;

[0046] Calculate the set percentile value of the gray-level co-occurrence matrix entropy of the entire image as the entropy threshold;

[0047] If the entropy value of the region is greater than the entropy threshold, it is marked as a surface creepage candidate region;

[0048] Calculate the minimum aspect ratio of the bounding rectangle of the candidate area;

[0049] When the aspect ratio is greater than the set ratio threshold, it is determined to be surface creepage, otherwise it is determined to be spark discharge.

[0050] Optionally, the performing binarization segmentation and contour optimization processing on the discharge channel confidence score map includes:

[0051] Adopting an adaptive threshold segmentation algorithm to perform binarization processing on the confidence score map to generate an initial discharge channel binary mask;

[0052] performing a morphological closing operation on the initial binary mask to fill internal holes and smooth the contour, wherein a structure element adopts a kernel of a preset shape and size;

[0053] Calculate the curvature value of each point in the contour after the closing operation, and eliminate the mutation points whose absolute value of curvature is greater than the set curvature threshold;

[0054] The optimized contour is skeletonized to obtain the discharge channel centerline, and the channel length is calculated based on the number of centerline pixels.

[0055] The fractal dimension of the discharge channel is calculated using the box counting method;

[0056] The number of branches is counted through contour topology analysis.

[0057] Optionally, the calculating the fractal dimension of the discharge channel by using a box counting method includes:

[0058] Use multiple square boxes with set side lengths to cover the discharge channel binary image;

[0059] Record the minimum number of boxes required to completely cover the channel;

[0060] The fractal dimension value is calculated based on the logarithmic relationship between the box size and the corresponding number of boxes.

[0061] Optionally, the output parameterized discharge channel profile and quantitative report includes:

[0062] The morphological parameters include at least channel length, fractal dimension and number of branches;

[0063] The parameterized discharge channel profile is a vector profile that has been optimized through curvature screening.

[0064] A second aspect of the present invention provides a discharge channel extraction system based on edge detection and texture feature fusion, based on the discharge channel extraction method based on edge detection and texture feature fusion described in the first aspect of the present invention, the system comprises:

[0065] The ultraviolet imaging module, consisting of a UVA-band enhanced CCD camera and an optical filter, is used to capture ultraviolet weak-light images of power equipment;

[0066] Edge-texture fusion processor, integrating FPGA hardware acceleration unit and broken edge connection unit, is used to perform image preprocessing, multi-scale edge detection, broken connection, texture feature extraction and dynamic weighted fusion;

[0067] Dynamic weight configuration module, used to identify discharge types based on gradient and texture features and dynamically assign fusion weight coefficients;

[0068] The contour optimization output module is used to perform binary segmentation on the fusion results, perform curvature optimization contour processing, and quantify the output of morphological parameters.

[0069] Compared with the prior art, the beneficial effects of the present invention include at least:

[0070] 1. Achieve high-precision discharge channel extraction. Through the dynamic fusion of multi-scale edge detection and texture features, the detection capability of low-contrast areas is significantly improved, and the missed detection rate is reduced from 33% to 5.2%;

[0071] 2. Enhanced noise immunity: Using an improved non-local means denoising algorithm combined with CLAHE contrast enhancement, the false edges are reduced by 72% in ultraviolet low-light images with a signal-to-noise ratio (SNR) of 8.7dB.

[0072] 3. Dynamic quantitative analysis of discharge morphology is achieved by filling holes through closing operations, optimizing contours through curvature screening, and quantifying channel length, fractal dimension, and number of branches, supporting temporal evolution analysis of discharge paths.

[0073] 4. Improve the integrity of the discharge profile. By connecting the fracture edges through a multi-scale voting mechanism and directional consistency, the profile continuity index (CI) is increased from 0.76 to 0.93, enhancing the integrity of the discharge channel. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a flow chart of an overall method provided according to an embodiment of the present invention;

[0075] Figure 2 is a flow chart of multi-scale edge detection and fusion provided according to an embodiment of the present invention;

[0076] Figure 3 is a schematic diagram of texture feature fusion provided according to an embodiment of the present invention;

[0077] Figure 4 is a flow chart of contour optimization and morphology quantification provided according to an embodiment of the present invention;

[0078] Figure 5 It is a block diagram of the system composition provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0079] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0080] In order to more clearly introduce the outstanding essential features of the present invention and the significant progress it brings to the prior art, an application example of implementing the present invention is introduced below.

[0081] The following describes an embodiment of the present invention in detail with reference to the accompanying drawings. The application example specifically includes:

[0082] like Figure 1-4 As shown, the present invention provides a discharge channel extraction method based on multi-scale edge detection and dynamic texture feature fusion in embodiment 1, comprising the following steps:

[0083] Step 1: Obtain ultraviolet weak light images of power equipment and preprocess them based on the improved non-local means algorithm (integral acceleration) and limited contrast histogram equalization (CLAHE block enhancement).

[0084] Preferably, the step 1 comprises:

[0085] Step 1.1, obtaining the original ultraviolet weak light image of the power equipment;

[0086] Step 1.2: Use the improved fast non-local means algorithm to denoise the original ultraviolet low-light image;

[0087] Further preferably, the adopting of the improved fast non-local means algorithm includes:

[0088] Define the denoising formula:

[0089]

[0090] Where Z(x) is the normalization coefficient used to ensure the stability and physical meaning of the pixel value after denoising. The weight w(x,y) is calculated based on the similarity of the pixel blocks, and an exponential decay function is introduced to optimize the computational efficiency:

[0091]

[0092] Where P(x) represents a 7×7 block centered on pixel x. It should be noted that the size of the pixel block determines the denoising effect and computational time. The parameter selection of the 7×7 pixel block significantly reduces the amount of similarity calculation while maintaining the denoising effect, ensuring the real-time performance of the algorithm. For scenes with low real-time requirements, the size can be reduced to enhance the denoising effect. h is the attenuation factor (the empirical value is set to 10). The integral image method is used to reduce the computational complexity from O(N 2 ) is reduced to O(N), ensuring real-time performance.

[0093] In step 1.3, the contrast of the denoised ultraviolet low-light image is enhanced by contrast-limited histogram equalization (CLAHE block enhancement).

[0094] Further preferably, the step 1.3 includes:

[0095] Based on the average discharge channel width of 20 pixels, a 32×32 block size was set. A contrast limit of 0.03 was applied to suppress noise amplification, and block histogram equalization was performed to ensure that local contrast was improved while avoiding blocking effects.

[0096] Aiming at the problem that ultraviolet weak light image feature extraction is interfered by noise in the existing technology, the present invention effectively suppresses noise interference and improves image contrast by improving the non-local means denoising algorithm and the contrast-limited block histogram equalization (CLAHE block enhancement) preprocessing.

[0097] In step 2, based on the preprocessed image in step 1, a gradient field is constructed using a multi-scale Gaussian kernel and edge fusion is performed to generate a continuous and complete candidate edge map.

[0098] Preferably, the step 2 comprises:

[0099] Step 2.1, based on the improved Canny algorithm, calculate the gradient amplitude G at multiple Gaussian kernel scales σ (x,y) and direction θ σ (x,y);

[0100] Preferably, the step 2.1 determines the Gaussian kernel scale parameter by the following steps:

[0101] The resolution of the UV imaging system is R (unit: μm / pixel), the average width of the discharge channel is W (unit: μm), and the corresponding pixel width is W1:

[0102]

[0103] The Gaussian kernel standard deviation σ needs to match the half-width height of the discharge channel:

[0104]

[0105] If the physical width is directly matched according to the multi-scale characteristics of the discharge channel, the scales of the three groups of Gaussian kernels are:

[0106] Main channel detection: σ1 = 0.425 × 20 ≈ 8.5 pixel (matching 20 pixel width);

[0107] Medium branch detection: σ2 = 0.425 × 10 ≈ 4.25 pixels (matching 10 pixel width);

[0108] Fine bifurcation detection: σ3 = 0.425 × 5 ≈ 2.13 pixels (matching 5-pixel width).

[0109] However, the Gaussian kernel scale calculated based on the physical width of the discharge channel and the imaging resolution corresponds to a Gaussian kernel radius of about 3σ pixels, which results in the calculation of each pixel involving a 3σ×3σ neighborhood, and the computational complexity is O(3σ 2 ), which is difficult to implement in real-time processing;

[0110] Therefore, combined with computational efficiency optimization, we finally selected three scales of σ = 1.0, 1.5, and 2.0 to cover different levels of features through multi-scale fusion:

[0111] σ = 1.0 is used to capture high-frequency details (such as bifurcation tips, noise edges); σ = 1.5 is used to balance details and noise resistance (medium branches); σ = 2.0 is used to suppress noise and preserve the main channel outline.

[0112] Further preferably, the step 2.1 includes:

[0113] The gradient magnitude Gσ(x,y) and direction θσ(x,y) are calculated at three Gaussian kernel scales (σ=1.0, 1.5, 2.0):

[0114] G x,σ (x,y)=I(x,y)*G x (σ)

[0115] G y,σ (x,y)=I(x,y)*G y (σ)

[0116]

[0117] Where I(x,y) is the pixel value of the image at (x,y); G x (σ) is the first-order derivative of the Gaussian kernel in the x direction; G y (σ) is the first-order derivative of the Gaussian kernel in the y direction; G y,σ (x, y) is the vertical gradient value of the image at position (x, y) under scale σ; G x,σ(x,y) is the horizontal gradient value of the image at position (x,y) under scale σ.

[0118] The present invention introduces a multi-Gaussian kernel gradient field and solves the edge breakage problem in weak light through a multi-scale voting mechanism and direction consistency connection.

[0119] Step 2.2: Based on the gradient magnitude calculated in step 2.1, perform non-maximum suppression (NMS) on the gradient field at each scale, retain local maximum pixels along the gradient direction, dynamically set dual thresholds, and segment the pixels based on the dual thresholds.

[0120] Further preferably, the step 2.2 includes:

[0121] Non-maximum suppression (NMS) is used to refine the edge, and the amplitude of neighboring pixels is compared along the gradient direction, and only the local maximum pixels are retained;

[0122] Dynamically set dual thresholds:

[0123] T low =0.1·G max

[0124] T high =0.3·G max

[0125] Among them G max is the maximum gradient amplitude at the current scale;

[0126] Pixel segmentation based on dual thresholds, including:

[0127] Gradient value G σ (x,y)≥T high The pixels are directly marked as strong edges;

[0128] Gradient value T low ≤G σ (x,y) <T high Pixels are only retained if they are connected to a strong edge;

[0129] Gradient value G σ (x,y) <T low Pixels with are considered as noise and removed.

[0130] In step 2.3, edge fusion and direction consistency connection are performed through a multi-scale voting mechanism to generate a continuous and complete candidate edge map.

[0131] Further preferably, the step 2.3 includes:

[0132] (1) Multi-scale voting fusion

[0133] For each pixel, if it is judged as an edge (strong or weak edge) at at least two different scales, it is marked as a candidate edge;

[0134] (2) Fracture edge connection

[0135] Traverse the adjacent broken edge segments in the candidate edge graph, and perform local connection based on direction consistency if the following conditions are met. The connection condition is that the pixel distance between the broken endpoints is less than the preset distance threshold D max And the gradient direction difference between the two end points is less than the preset angle threshold θ, specifically:

[0136] |θ σ (x1,y1)-θ σ (x2,y2)|<θ and the distance <D max

[0137] Output a continuous and complete candidate edge map for subsequent texture feature fusion.

[0138] Preferably, the spatial resolution of the UV imaging system determines the actual physical size corresponding to the pixel, and the preset distance threshold is dynamically set according to the imaging resolution R and the local noise level of the image:

[0139]

[0140] Among them L max is the maximum physical break distance allowed, σ 噪声 is the noise standard deviation, k is the empirical coefficient (1 to 2);

[0141] Since photon noise in ultraviolet imaging can cause gradient direction estimation errors, assuming the system angle error is Δθ, the preset system angle threshold θ must satisfy:

[0142] θ≥2Δθ

[0143] The experimental results show that the gradient direction error of UV imaging Δθ≤7°, so θ=15° is set.

[0144] Step 3: Extract texture features of the preprocessed image, including rotation-invariant local binary pattern uniformity features and multi-directional grayscale co-occurrence matrix features; dynamically assign weight coefficients according to discharge types, perform weighted fusion on the edge strength of the candidate edge map and the texture features, and generate a discharge channel confidence score map.

[0145] Preferably, the step 3 includes:

[0146] Step 3.1: For the preprocessed image obtained in step 1, calculate the uniformity feature of the rotationally invariant local binary pattern and count the percentage of uniform pattern pixels, wherein the uniformity percentage of the discharge channel region is less than a first threshold and the uniformity percentage of the background region is greater than a second threshold;

[0147] Further preferably, the step 3.1 includes:

[0148] Local Binary Pattern (LBP) uniformity feature: Calculate the rotation-invariant uniform LBP pattern and calculate the uniform pattern ratio (Uniform Ratio) in the image:

[0149]

[0150] The discharge channel area has a high texture complexity, so its Uratio is less than 0.7, while the background area has a Uratio greater than 0.85.

[0151] Step 3.2, performing multi-directional gray-level co-occurrence matrix (GLCM) feature extraction on the pre-processed image of step 1, including calculating the contrast and entropy values ​​of the gray-level co-occurrence matrix in four preset directions, taking the average of the contrast and entropy values ​​in each direction as the texture complexity index, and the gray-level co-occurrence matrix has a preset number of gray levels;

[0152] Further preferably, the step 3.2 includes:

[0153] Gray-level co-occurrence matrix (GLCM) features: Calculate the contrast C(x,y) and entropy E(x,y) in four directions (0°, 45°, 90°, 135°), and take the mean as the texture complexity index:

[0154]

[0155] Among them, p(i,j) is the normalized gray-level co-occurrence matrix element, which represents the probability that the gray values ​​of two adjacent pixels are i and j respectively under a specific spatial relationship.

[0156] In step 3.3, the edge intensity of each pixel in the candidate edge map of step 2 is weightedly fused with the LBP uniformity feature of step 3.1 and the GLCM contrast and entropy value of step 3.2 to generate a pixel-level confidence score and obtain a discharge channel confidence score map.

[0157] Further preferably, the step 3.3 includes:

[0158] Construct pixel-level confidence score S(x,y), dynamically fusing edge strength and texture features:

[0159]

[0160] S(x,y)=α·G σ (x,y)+β·U(x,y)+γ·C(x,y)

[0161] Where U(x,y) is the texture uniformity, N represents the number of gray levels in the gray-level co-occurrence matrix (GLCM);

[0162] The weight coefficients α, β, and γ are adaptively adjusted according to different discharge types:

[0163] Calculate the mean μ and standard deviation σ of the gradient amplitude of the whole image, set the threshold T1 = μ + 2σ, if the gradient amplitude of a certain area G σ (x,y)>T and the U of the region ratio <0.7, it is judged as corona discharge;

[0164] Calculate 90% of the GLCM entropy value E of the entire image and set the threshold T2 = E × 0.9. If the entropy value E of a certain area is greater than T2, it is preliminarily determined to be a candidate area for surface creepage.

[0165] The aspect ratio of the candidate area is calculated by sliding the window. If the aspect ratio is greater than 3:1, it is considered as surface creepage, otherwise it is considered as spark discharge.

[0166] Weight coefficient settings for different discharge types:

[0167] Corona discharge: α = 0.6, β = 0.2, γ = 0.2, focusing on edge strength;

[0168] Surface creepage: α = 0.3, β = 0.5, γ = 0.2, focusing on texture uniformity;

[0169] Spark discharge: α = 0.4, β = 0.3, γ = 0.3, balanced mode.

[0170] It is worth noting that, in order to address the problem of insufficient capture of discharge channel details in weak light environments in the prior art, the present invention significantly enhances the integrity of the discharge channel contour by integrating multi-scale edge detection and gray-level co-occurrence matrix texture complexity index.

[0171] Step 4: Binarize and segment the discharge channel confidence score map, and output the parameterized discharge channel contour and quantitative report through contour optimization and morphology parameter extraction.

[0172] Preferably, in step 4, performing contour optimization includes:

[0173] A 3×3 elliptical kernel is used to close the binary mask to fill holes and smooth the contours.

[0174] Calculate the curvature k of the contour point:

[0175]

[0176] The curvature mutation points (|k|>0.25) are eliminated to retain the naturally curved discharge path.

[0177] Preferably, in step 4, extracting morphological parameters includes:

[0178] Channel length: Pixel counting based on the centerline of the skeletonized discharge channel;

[0179] Fractal dimension: calculated using the box counting method to quantify the channel tortuosity;

[0180]

[0181] Where N(ε) is the minimum number of boxes required to completely cover the discharge channel under a certain box size ε, and ε is the side length (pixels) of the square box used to cover the pattern;

[0182] Number of branches: obtained through contour topology analysis.

[0183] In discharge channel analysis: the larger the fractal dimension D value, the more tortuous the channel path; the D value increases dynamically with the development of discharge, reflecting the fault evolution trend; if there is a sudden change in the D value, it indicates that there may be a risk of insulation breakdown.

[0184] In response to the problem of lack of quantification of discharge channel morphology parameters in the existing technology, the present invention realizes the quantitative characterization of the discharge channel morphology and supports the timing analysis of the discharge development path through fractal dimension calculation and branch number topological analysis, and quantifies the tortuosity and bifurcation evolution characteristics of the discharge channel.

[0185] The technical effects of the method provided by the present invention include:

[0186] 1. Combining multi-scale edge detection with texture features, dynamic weight allocation enhances robustness in low-light environments. This improves anti-interference capabilities, reducing the number of false edges in Example 3 from 458 to 128, and the missed detection rate in low-contrast areas from 33% to 5.2%.

[0187] 2. A multi-Gaussian kernel is used to construct a gradient field, and fracture edges are connected through a multi-scale voting mechanism and directional consistency, significantly improving contour continuity. This enhances the integrity of the discharge contour. In Example 3, the contour continuity index is increased from 0.76 to 0.93 (number of effective connected edges / total number of edges).

[0188] 3. Introducing closing operations and curvature screening to optimize the contour, and quantifying parameters such as channel length, fractal dimension, and number of branches, supporting discharge timing analysis and discharge channel evolution analysis, and using quantitative parameters such as fractal dimension and number of branches to capture the bifurcation and expansion process of the discharge channel.

[0189] like Figure 5 As shown, the present invention provides a discharge channel extraction system based on edge detection and texture feature fusion in embodiment 2. Based on the discharge channel extraction method based on edge detection and texture feature fusion described in embodiment 1, the system includes:

[0190] UV imaging module, edge-texture fusion processor, dynamic weight configuration module and contour optimization output module, wherein:

[0191] The ultraviolet imaging module includes:

[0192] UVA-band enhanced CCD camera (response wavelength 320-400 nm), resolution 1280 × 1024, frame rate 30 fps;

[0193] The optical filter suppresses visible light interference, with a transmittance of >90% (center wavelength 340nm, bandwidth 10nm).

[0194] The edge-texture fusion processor includes:

[0195] FPGA hardware acceleration unit for parallel multi-scale Gaussian gradient calculation and integrated LBP / GLCM texture feature parallel calculation unit;

[0196] The broken edge connection unit is used to perform "AND OR" logic voting on the three-scale edge detection results, output the candidate edge map, and max and the angle threshold θ = 15° connects the fracture edges.

[0197] The dynamic weight configuration module includes:

[0198] According to the calculation results of the gradient amplitude mean μ, standard deviation σ, and GLCM entropy E of the entire image, the discharge type is automatically identified, and the weight coefficients α, β, and γ are adjusted based on the discharge type.

[0199] The contour optimization output module includes:

[0200] Morphological processor: 3×3 ellipse kernel closing operation (iteration 2 times), filling holes;

[0201] Curvature screening unit: calculates the curvature k of the contour points and eliminates mutation points with |k|>0.25;

[0202] Shape quantization engine: skeletonized centerline pixel count → channel length;

[0203] Box counting method (box size ε = 2 / 4 / 8 / 16 pixels) → fractal dimension D;

[0204] Contour topology analysis → number of branches;

[0205] Finally, a vectorized discharge channel profile is generated; a morphology parameter report is output and supported for comparison with historical data.

[0206] In Example 3, the present invention provides an application example of a discharge channel extraction method based on edge detection and texture feature fusion. Based on the discharge channel extraction method based on edge detection and texture feature fusion described in Example 1, the application example includes:

[0207] 1. Test object: artificial defects on the surface of 220kV composite insulators (length 3cm, depth 0.5mm)

[0208] 2. Environmental parameters: relative humidity 65% ​​RH, temperature 28°C, observation distance 15m

[0209] 3. Implementation steps:

[0210] (1) Data collection and preprocessing:

[0211] The discharge sequence images were collected (lasting 5 seconds, a total of 150 frames), and the 80th frame typical image was selected. The original image signal-to-noise ratio (SNR) was 8.7 dB.

[0212] (2) Improved non-local mean denoising:

[0213] The search window Ω is set to 21×21, the similar block P(x) is set to 7×7, the attenuation factor is set to 10, and the integral image method is used to accelerate the weight calculation, reducing the time consumption from 1.2s of the traditional algorithm to 0.3s.

[0214] (3) CLAHE enhancement:

[0215] Using 32×32 blocks and a cropping limit of 0.03, the contrast of the output image is improved by 2.6 times.

[0216] (4) Multi-scale edge detection:

[0217] Construct a three-scale gradient field with Gaussian kernel σ = 1.0 / 1.5 / 2.0, and set the dual threshold adaptively:

[0218] When σ=1.0, G max =189→T low =18.9, T high =56.7

[0219] When σ=1.5, G max =153→T low =15.3, T high =45.9

[0220] When σ=2.0, G max =132→T low =13.2, T high=39.6

[0221] After multi-scale voting fusion, 12,385 edge pixels were detected, and the fracture of the connection was achieved through direction consistency (imaging resolution R = 10 μm / pixel, which is the maximum allowable physical fracture distance L max =50μm, connection distance threshold D max =5 pixels, angle difference <15°).

[0222] (5) Texture feature fusion:

[0223] Calculate the LBP uniformity feature: Uratio = 0.62 in the discharge area and Uratio = 0.91 in the background area.

[0224] GLCM feature extraction (grayscale N = 16): discharge area contrast C = 4.8, entropy E = 2.1; background area C = 1.2, E = 0.7.

[0225] (6) Dynamic weight configuration:

[0226] It is a surface creepage type, and α=0.3, β=0.5, and γ=0.2 are set.

[0227] (7) Contour optimization and quantization:

[0228] Morphological closing operation: Use a 3×3 ellipse kernel for two iterations to fill five internal holes and smooth out jagged contours.

[0229] Curvature screening: 23 mutation points with |k|>0.25 were eliminated, and the natural curved path was retained.

[0230] The output results of the morphological parameters are shown in Table 1:

[0231] Table 1 Morphological parameter output results

[0232] parameter Measurements Channel length / pixel 356 Fractal dimension D 1.68 Number of branches 5

[0233] 4. Implementation effect:

[0234] Compared with the traditional Canny algorithm: false edges are reduced by 72% (from 458 to 128); the detection rate of low-contrast areas is improved by 41% (the missed detection rate is reduced from 33% to 5.2%); the contour continuity index (CI) is improved from 0.76 to 0.93 (CI = number of effective connected edges / total number of edges)

[0235] 5. Timing analysis: Processing 150 consecutive frames took 4.8 seconds (32ms per frame under FPGA acceleration), successfully capturing the bifurcation and expansion process of the discharge channel (the number of branches increased from 3 to 5).

[0236] The present disclosure may be a system, method and / or computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0237] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A discharge channel extraction method based on edge detection and texture feature fusion, characterized in that: The steps include: Acquire ultraviolet weak light images of power equipment and perform preprocessing; Perform multi-scale edge detection based on the preprocessed image, fuse edge information through a multi-scale voting mechanism, and combine directional consistency to connect broken edges to generate a candidate edge map; Extracting texture features of the preprocessed image, including rotation-invariant local binary pattern uniformity features and multi-directional gray-level co-occurrence matrix features; Dynamically assigning weight coefficients according to discharge types, weightedly fusing the edge strength of the candidate edge map with the texture features to generate a discharge channel confidence score map; The discharge channel confidence score map is subjected to binary segmentation and contour optimization processing, and the morphology of the discharge channel is extracted, and a parameterized discharge channel contour and a quantitative report are output.

2. The method for extracting discharge channels based on edge detection and texture feature fusion according to claim 1, characterized in that: The preprocessing of the ultraviolet weak light image includes: The non-local means algorithm is used to denoise the original image, where: Use integral image method to accelerate similarity calculation of pixel blocks of preset size; The weights are calculated based on the exponential decay function of the Euclidean distance of the pixel blocks, and the denoised image is subjected to the limited contrast histogram equalization process with the set attenuation factor. Set blocks of preset size according to the average width of the discharge channel.

3. The method for extracting discharge channels based on edge detection and texture feature fusion according to claim 1, characterized in that: Generating a candidate edge graph comprises: The gradient magnitude and gradient direction of the preprocessed image are calculated under three preset Gaussian kernel scales; Performing non-maximum suppression on the gradient field at each scale, and dynamically setting a dual threshold for pixel segmentation based on the maximum gradient amplitude at the current scale, wherein the dual threshold is set based on the maximum gradient amplitude according to a preset ratio; For each pixel, if it is judged as an edge pixel at at least two different scales, it is marked as a candidate edge; Traversing adjacent broken edge segments in the candidate edge map, connecting them if the pixel distance between the broken endpoints is less than a preset distance threshold and the gradient direction difference between the two endpoints is less than a preset angle threshold; Output a continuous candidate edge map.

4. The method for extracting discharge channels based on edge detection and texture feature fusion according to claim 1, characterized in that: The extracting of texture features of the preprocessed image comprises: calculating a rotationally invariant local binary pattern feature of the preprocessed image; Counting the proportion of uniform pattern pixels in the rotation-invariant local binary pattern feature as a texture uniformity feature; Calculating the gray level co-occurrence matrix of the pre-processed image in a plurality of preset directions; Calculating the contrast value and entropy value of the gray level co-occurrence matrix in each preset direction; Take the average of the contrast values ​​in the four preset directions as the comprehensive contrast index; Take the average of the entropy values ​​in the four preset directions as the comprehensive entropy index; The comprehensive contrast index and the comprehensive entropy index together constitute the multi-directional gray-level co-occurrence matrix feature.

5. The method for extracting discharge channels based on edge detection and texture feature fusion according to claim 1, characterized in that: Generating a discharge channel confidence score map includes: Automatically identifying the discharge type based on the global gradient amplitude statistical characteristics and global gray-level co-occurrence matrix entropy characteristics of the preprocessed image; According to the identified discharge type, assigning corresponding first weight coefficient α, second weight coefficient β and third weight coefficient γ to the edge strength, texture uniformity feature and texture complexity index based on a preset weight assignment rule; For each pixel position in the candidate edge map, perform the following operations: Multiply the normalized edge intensity by α, the normalized texture uniformity eigenvalue by β, and the normalized texture complexity index by γ; The confidence score of the pixel position is obtained by summing the three product results; Output the discharge channel confidence score map consisting of the confidence scores of all pixel positions.

6. The method for extracting discharge channels based on edge detection and texture feature fusion according to claim 5, characterized in that: The automatic identification of discharge types includes: Calculate the mean μ and standard deviation σ of the gradient amplitude of the entire image; If the gradient amplitude of a certain area is greater than the linear combination value of μ and σ and the texture uniformity characteristic value of the area is less than the preset threshold, it is determined to be corona discharge; Calculate the set percentile value of the gray-level co-occurrence matrix entropy of the entire image as the entropy threshold; If the entropy value of the region is greater than the entropy threshold, it is marked as a surface creepage candidate region; Calculate the minimum aspect ratio of the bounding rectangle of the candidate area; When the aspect ratio is greater than the set ratio threshold, it is determined to be surface creepage, otherwise it is determined to be spark discharge.

7. The method for extracting discharge channels based on edge detection and texture feature fusion according to claim 1, characterized in that: The binary segmentation and contour optimization processing of the discharge channel confidence score map includes: Adopting an adaptive threshold segmentation algorithm to perform binarization processing on the confidence score map to generate an initial discharge channel binary mask; performing a morphological closing operation on the initial binary mask to fill internal holes and smooth the contour, wherein a structure element adopts a kernel of a preset shape and size; Calculate the curvature value of each point in the contour after the closing operation, and eliminate the mutation points whose absolute value of curvature is greater than the set curvature threshold; The optimized contour is skeletonized to obtain the center line of the discharge channel, and the channel length is calculated based on the number of center line pixels. The fractal dimension of the discharge channel is calculated using the box counting method; The number of branches is counted through contour topology analysis.

8. The method for extracting discharge channels based on edge detection and texture feature fusion according to claim 7, characterized in that: The method of calculating the fractal dimension of the discharge channel by using the box counting method includes: Use multiple square boxes with set side lengths to cover the discharge channel binary image; Record the minimum number of boxes required to completely cover the channel; The fractal dimension value is calculated based on the logarithmic relationship between the box size and the corresponding number of boxes.

9. The method for extracting discharge channels based on edge detection and texture feature fusion according to claim 8, characterized in that: The output parameterized discharge channel profile and quantitative report include: The morphological parameters include at least channel length, fractal dimension and number of branches; The parameterized discharge channel profile is a vector profile that has been optimized through curvature screening.

10. A discharge channel extraction system based on edge detection and texture feature fusion, based on the discharge channel extraction method based on edge detection and texture feature fusion according to any one of claims 1 to 9, characterized in that: The system includes: The ultraviolet imaging module, consisting of a UVA-band enhanced CCD camera and an optical filter, is used to capture ultraviolet weak-light images of power equipment; Edge-texture fusion processor, integrating FPGA hardware acceleration unit and broken edge connection unit, is used to perform image preprocessing, multi-scale edge detection, broken connection, texture feature extraction and dynamic weighted fusion; Dynamic weight configuration module, used to identify discharge types based on gradient and texture features and dynamically assign fusion weight coefficients; The contour optimization output module is used to perform binary segmentation on the fusion results, perform curvature optimization contour processing, and quantify the output of morphological parameters.

Images