Visual defect feature extraction and recognition method for impact-resistant power safety warning facility

Abstract

This invention discloses a method for extracting and identifying visual defect features of impact-resistant power safety warning facilities, belonging to the field of intelligent power facility inspection technology. The method includes acquiring image sequences of facilities and environmental state parameters in an outdoor environment; deconstructing the images into physical and apparent states to obtain deconstructed images of components such as the outer shell, reflective coating, and printed markings; performing pixel-level difference analysis between these images and a standard template to generate an anomaly difference map. A facility health degradation model is constructed based on the correlation between environmental parameters and the difference map. Under the guidance of this model, iterative feature focusing calculations are performed on the difference map to extract multi-scale visual defect feature vectors corresponding to impact damage patterns. Based on these feature vectors, defect matching decisions are made, outputting specific defect types and location information. This invention achieves accurate identification of impact damage and health status assessment of warning facilities in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection technology for power facilities, specifically a method for extracting and identifying visual defect features of impact-resistant power safety warning facilities. Background Technology

[0002] Currently, image recognition technology based on fixed algorithms is commonly used for defect detection of outdoor power safety warning facilities. This type of technology typically performs feature analysis on the acquired images or compares them with standard templates to identify surface contamination, damage, or missing parts. This method treats visual defects as isolated, static appearances.

[0003] Existing technical solutions have shortcomings. Their detection process is disconnected from the complex outdoor environment and long-term environmental stresses experienced by the facilities, making it impossible to distinguish between gradual defects caused by natural aging and sudden damage caused by accidental impacts. Furthermore, conventional feature extraction methods are general and fixed, lacking targeted modeling for the specific physical structure and impact damage evolution patterns of power safety warning facilities. This results in low feature discrimination, making it difficult to accurately identify the root causes and types of defects, especially impact damage, a critical failure mode.

[0004] A technique is needed to correlate the cumulative effects of environmental factors with visual defects in order to quantify the health degradation process of facilities. Furthermore, an adaptive method for extracting defect features that aligns with the physical mechanisms of impact is required to achieve accurate identification of impact damage patterns. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] Therefore, this invention proposes a method for extracting and identifying visual defect features of impact-resistant power safety warning facilities, including:

[0007] Collect image sequences of electrical safety warning facilities in outdoor environments and obtain corresponding environmental status parameters;

[0008] Each frame of the image sequence is deconstructed in multiple dimensions of physical and apparent states to obtain component deconstruction images including the shell structure area, reflective coating area, and printed marking area.

[0009] The component deconstruction image is compared with a preset standard template image by pixel-level difference operation to generate an abnormal difference map reflecting the component's geometric deformation and color attenuation;

[0010] Based on the correlation mapping between the environmental state parameters and the abnormal difference map, a facility health degradation model is constructed.

[0011] Guided by the facility health degradation model, iterative feature focusing calculations are performed on the abnormal difference map to extract multi-scale visual defect feature vectors corresponding to the impact damage pattern.

[0012] Based on the multi-scale visual defect feature vector, a defect type matching decision is performed, and the specific defect type and location information of the power safety warning facility are output.

[0013] Furthermore, the step of performing multi-dimensional deconstruction of the physical and apparent states of each frame in the image sequence to obtain component deconstruction images including the shell structure region, the reflective coating region, and the printed marking region includes:

[0014] An edge-aware segmentation algorithm is used to initially divide the image, resulting in a mask that includes the main structure and the background.

[0015] Within the facility body area defined by the mask, a group of pixels with high reflectivity is separated according to a color space clustering algorithm, and the group of pixels with high reflectivity constitutes the reflective coating area.

[0016] Within the facility body area defined by the mask, a set of pixels with a continuous flat surface is defined according to the texture consistency analysis algorithm, and the set of pixels with a continuous flat surface constitutes the outer shell structure area.

[0017] Within the facility's main area defined by the mask, optical character recognition pre-detection technology is used to locate pixel blocks containing specific combinations of characters or symbols, and these pixel blocks containing specific combinations of characters or symbols constitute the printed marking area.

[0018] The reflective coating area, the outer shell structure area, and the printed marking area are extracted and stored as independent image layers, which together form the component deconstruction image.

[0019] Further, the step of performing pixel-level difference operations between the component deconstruction image and a preset standard template image to generate an abnormal difference map reflecting the component's geometric deformation and color attenuation includes:

[0020] Spatial registration is performed between each independent image layer in the component deconstruction image and the corresponding standard layer in the preset standard template image;

[0021] For each registered layer pair, calculate the absolute difference in the three channels of brightness, chroma and saturation pixel by pixel;

[0022] Pixels whose absolute value difference exceeds a preset perception threshold are marked as abnormal pixels, and their deviation from the standard value is recorded.

[0023] For the layer corresponding to the shell structure region, the curvature change of the local region is calculated in addition, and the region whose deviation from the standard curvature model exceeds the tolerance is marked as a geometric deformation anomaly region.

[0024] By integrating abnormal pixel markers and geometric deformation abnormal area information from all layers, an abnormality map is generated that encodes the type and degree of deviation using different color channels or grayscale levels.

[0025] Furthermore, the step of constructing a facility health degradation model based on the correlation mapping between the environmental state parameters and the anomaly difference map includes:

[0026] The environmental status parameters include facility service life, historical extreme weather event records, and surrounding vibration source distribution data;

[0027] Analyze the statistical distribution characteristics of various abnormal pixels in the abnormal difference map, including the density, spatial clustering and average deviation intensity of abnormal pixels;

[0028] Establish a correlation model between each parameter in the environmental state parameters and the statistical distribution characteristics of the anomaly difference map;

[0029] Based on the aforementioned correlation model, regression analysis was used to quantify the contribution weights of different combinations of environmental state parameters to various abnormal growths.

[0030] Using the contribution weights, a mathematical model is constructed that can predict the potential development rate of various anomalies in the anomaly difference map based on the input environmental state parameters. The mathematical model is defined as the facility health degradation model.

[0031] Furthermore, guided by the facility health degradation model, the iterative feature focusing calculation is performed on the abnormal difference map to extract multi-scale visual defect feature vectors corresponding to the impact damage pattern, including:

[0032] An initial analysis window of interest is set on the anomaly difference map, centered on the potential abnormal development area predicted by the facility health degradation model.

[0033] In each iteration, the directional gradient histogram and local binary pattern features of the abnormal pixels within the current analysis window of interest are calculated.

[0034] The calculated features are compared with a preset feature library of typical impact damage patterns. If the similarity is lower than the threshold, the position and size of the analysis window of interest are adjusted according to the gradient information.

[0035] Repeat the feature calculation, comparison and window adjustment steps until a stable window with a similarity to a certain type of impact damage pattern feature reaches a preset threshold is found;

[0036] From the abnormal difference map data within the stable window, normalized multi-dimensional statistical features are extracted, including but not limited to texture roughness, edge break length, and color patch size distribution. The normalized multi-dimensional statistical features are then combined into the multi-scale visual defect feature vector.

[0037] Furthermore, the multi-scale visual defect feature vector includes a macroscopic deformation feature sub-vector and a microscopic texture feature sub-vector. The macroscopic deformation feature sub-vector describes the distortion or depression of the overall contour of the component, and the microscopic texture feature sub-vector describes the fine patterns of coating peeling and wear marks.

[0038] Furthermore, the step of performing defect type matching decisions based on the multi-scale visual defect feature vector and outputting the specific defect type and location information of the power safety warning facility includes:

[0039] The multi-scale visual defect feature vectors are input into a pre-trained classification decision network;

[0040] The classification decision network contains multiple fully connected layers, which are used to map the input feature vector to a high-dimensional space for nonlinear transformation.

[0041] In the output layer of the classification decision network, a normalized exponential function is used to calculate the probability that the multi-scale visual defect feature vector belongs to each preset defect category. The preset defect categories include structural cracks, coating peeling, blurry markings, and composite damage.

[0042] Select the preset defect category with the highest probability as the final identified defect type;

[0043] Simultaneously, the coordinate position of the stable window that generates the multi-scale visual defect feature vector in the original image sequence is reverse-mapped to the physical location coordinates on the power safety warning facility to generate the positioning information;

[0044] The defect type is associated with the location information and output.

[0045] Furthermore, after acquiring image sequences of electrical safety warning facilities in the outdoor environment and obtaining corresponding environmental state parameters, the method further includes a step of preprocessing and enhancing the image sequences:

[0046] Temporal filtering is performed on the image sequence to eliminate random noise;

[0047] Histogram equalization algorithm is used to adjust the contrast of each frame of the image to enhance the distinction between the reflective coating area and the background.

[0048] Photometric compensation is applied to the shadow areas caused by uneven lighting, so that the brightness distribution of the entire facility surface tends to be uniform.

[0049] Furthermore, the histogram equalization algorithm is used to adjust the contrast of each frame of the image to enhance the distinction between the reflective coating area and the background, specifically as follows:

[0050] Calculate the global brightness histogram of the input image;

[0051] Construct the cumulative distribution function of the global brightness histogram;

[0052] The cumulative distribution function is used to remap the brightness value of each pixel in the input image to generate a contrast-enhanced image, making the pixel brightness distribution in the reflective coating area more concentrated.

[0053] Furthermore, after outputting the specific defect type and location information of the power safety warning facility, a step of quantifying the severity of the defect is also included:

[0054] Based on the values ​​of the macroscopic deformation feature sub-vectors in the multi-scale visual defect feature vector, the overall offset of the structure from the standard shape is calculated.

[0055] Based on the values ​​of the micro-texture feature sub-vectors in the multi-scale visual defect feature vector, the ratio of the defect area to the total apparent area of ​​the facility is calculated.

[0056] By combining the overall offset with the proportion, a comprehensive defect severity level is obtained through a predefined quantitative scoring table. The defect severity level is stored together with the defect type and location information.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] By mapping temporal environmental state parameters with pixel-level anomaly difference maps, a facility health degradation model is constructed that reflects the quantitative relationship between environmental stress and component performance degradation. This model overcomes the limitations of traditional methods that only focus on defect manifestations, enabling the system to understand the environmental causes and degradation stages of defects. Therefore, when analyzing visual data, it can prioritize and focus on component areas most likely to fail under specific environmental histories, elevating defect detection from static identification to dynamic health status assessment.

[0059] Using the output of a facility health degradation model as prior knowledge, the analysis process of anomaly difference maps is dynamically guided by iterative feature-focusing computation. This process adaptively adjusts the scale and spatial attention of feature extraction based on the degradation weaknesses predicted by the model, and performs multiple rounds of feature screening and enhancement around the physical patterns of impact damage. The resulting multi-scale visual defect feature vectors have clear physical meaning and highly correspond to specific impact damage mechanisms, improving the accuracy and reliability of subsequent defect type matching decisions, particularly in distinguishing between impact damage and natural aging. Attached Figure Description

[0060] Figure 1 This is a step diagram of the visual defect feature extraction and identification method for impact-resistant power safety warning facilities described in this invention;

[0061] Figure 2 A flowchart for obtaining component deconstruction images;

[0062] Figure 3 A flowchart for generating anomaly difference maps;

[0063] Figure 4 A diagram illustrating the co-evolutionary trend of multi-stage facility health degradation characteristics;

[0064] Figure 5 A comparative chart showing the development trend of defects in power safety warning facilities and the predicted operation and maintenance effects. Detailed Implementation

[0065] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0066] See Figure 1A continuous image sequence of power safety warning facilities is acquired using image acquisition equipment, while environmental state parameters related to the facility are simultaneously collected from an environmental monitoring system or historical database. Each frame in the image sequence undergoes multi-dimensional deconstruction processing to extract component deconstruction images of the outer shell structure area, reflective coating area, and printed marking area. These component deconstruction images are then subjected to pixel-level difference operations with pre-stored standard template images to generate an anomaly difference map that simultaneously reflects geometric deformation and color decay. Utilizing the correlation between environmental state parameters and statistical features extracted from the anomaly difference map, a facility health degradation model is established to predict changes in facility status. Guided by this model, iterative feature focusing calculations are performed on the anomaly difference map to extract multi-scale visual defect feature vectors associated with specific impact damage patterns. Based on these feature vectors, a classification decision process is used to match and determine defect types, ultimately outputting the specific defect types present on the power safety warning facility and their precise location information.

[0067] See Figure 2 In one embodiment of the present invention, in an example scenario, an image acquisition device captures an image containing a triangular power safety warning sign. The sign has a metal casing, a red and white reflective coating, and a black printed sign that reads "High Voltage Danger." The image resolution is 1920x1080 pixels, and it is stored in RGB format. In a specific implementation, an edge-aware segmentation algorithm is used to initially segment the input image. The edge-aware segmentation algorithm detects gradient edges in the image based on the Canny operator and combines a region growing algorithm to merge similar pixels, ultimately generating a binary mask image. In the mask image, white pixels correspond to the main body area of ​​the facility, and black pixels correspond to the background area. For example, in the original image, the outline of the main body of the warning sign is completely extracted, while the trees and sky in the background are excluded by the mask. Data comparison shows that the mask image reduces the number of pixels of the main body of the facility from 2,073,600 pixels in the full image to approximately 50,000 pixels.

[0068] In some embodiments, within the facility's main area defined by the mask, pixel groups belonging to high-brightness reflectivity are separated according to a color space clustering algorithm. The color space clustering algorithm converts the image from the RGB color space to the HSV color space and performs K-means clustering on the saturation and lightness channels of the HSV color space. The color space clustering algorithm calculates the distance between pixels and cluster centers using the following formula:

[0069]

[0070] in: Indicates distance, This represents the saturation value of the i-th pixel. This represents the brightness value of the i-th pixel. The saturation value representing the cluster centers. This represents the brightness value of the cluster center. The number of clusters is set to 3, corresponding to the bright reflective area, the non-reflective area, and the transition area, respectively. In the example scene, after processing by the color space clustering algorithm, the pixel group with bright reflective characteristics is separated to form the reflective coating area. Data comparison shows that the reflective coating area contains approximately 20,000 pixels, mainly distributed in the red stripe part of the warning sign.

[0071] In some embodiments, within the facility's main body area defined by the mask, a set of pixels with continuous, flat surfaces is defined according to a texture consistency analysis algorithm. The texture consistency analysis algorithm calculates the gray-level co-occurrence matrix within a local window of the image and extracts contrast and homogeneity features. Areas with feature values ​​below a threshold are determined to have continuous, flat surfaces. In an example scenario, the texture consistency analysis algorithm scans the facility's main body area and identifies the set of pixels corresponding to the metal shell. This set of pixels constitutes the shell structure area. Data comparison shows that the shell structure area contains approximately 25,000 pixels, corresponding to the triangular metal frame of the warning sign.

[0072] Optionally, within the facility's main area defined by the mask, optical character recognition (OCR) pre-detection technology is used to locate pixel blocks containing specific combinations of text or symbols. OCR pre-detection first uses connected component analysis to detect potential character regions, and then matches them with a preset font template. In the example scenario, OCR pre-detection technology located the pixel block containing the text "High Voltage Danger," which constitutes the printed label area. Data comparison shows that the printed label area contains approximately 5000 pixels, accurately defining the text boundaries.

[0073] It's understandable that the reflective coating area, the outer shell structure area, and the printed marking area are extracted and stored as independent image layers. Each image layer has the same size and coordinate system as the original image, but only the pixel values ​​of the corresponding area are retained, while other areas are filled with zero values. In practice, the pixels of each area are extracted using functions from image processing libraries such as OpenCV, and three independent single-channel or multi-channel image arrays are created to jointly form the component deconstruction image. In the example scenario, the component deconstruction image is composed of three layers superimposed: the reflective coating area layer is highlighted in the red channel, the outer shell structure area layer is displayed in the blue channel, and the printed marking area layer is displayed in the green channel. Data comparison shows that the total data volume of the component deconstruction image is approximately 1.5 times that of the original image because each layer is stored independently.

[0074] In practical implementation, the parameters of the edge-aware segmentation algorithm, color space clustering algorithm, texture consistency analysis algorithm, and optical character recognition pre-detection technology can be adjusted according to the actual facility type. For example, for circular warning facilities, the gradient threshold of the edge-aware segmentation algorithm can be appropriately reduced to capture the curved contour. Optionally, the number of clusters in the color space clustering algorithm can be dynamically set based on the number of reflective coating colors to handle multi-color warning facilities. It can be understood that the generation of component deconstruction images provides structured input for subsequent difference operations, ensuring that each component region is processed independently.

[0075] See Figure 3 In one embodiment of the present invention, the component deconstruction image comes from an outdoor triangular power safety warning sign image that has been deconstructed in multiple dimensions. The preset standard template image is an ideal image of the same model warning sign that is free of defects and stored in the database. The standard template image also includes the corresponding standard shell structure layer, standard reflective coating layer and standard printed marking layer, and has the same resolution and color space as the component deconstruction image.

[0076] In some embodiments, each independent image layer in the component deconstruction image is spatially registered with the corresponding standard layer in a preset standard template image. Spatial registration employs an affine transformation method based on feature points. In the example scenario, for the shell structure region layer, the three vertices of the warning sign are selected as feature points and matched with their corresponding points in the standard shell structure layer. A transformation matrix is ​​calculated, and this matrix aligns the shell structure region layer in the component deconstruction image with the standard shell structure layer. Data comparison shows that the average positional error of corresponding pixels in the two layers after registration is less than 0.5 pixels.

[0077] In practice, for each registered layer pair, the absolute difference in the three channels of luminance, chroma, and saturation is calculated pixel-by-pixel. The calculation for each pixel follows the formula:

[0078]

[0079] in: This represents the overall difference value of a single pixel. This represents the brightness value of a pixel in the current component deconstruction layer. This represents the brightness value of the corresponding pixel in the layer corresponding to the standard template. This represents the chromaticity value of a pixel in the current component deconstruction layer. This represents the chromaticity value of the corresponding pixel in the layer corresponding to the standard template. This represents the saturation value of the pixels in the current component deconstruction layer. This represents the saturation value of the corresponding pixel in the layer corresponding to the standard template. In the example scene, a difference matrix with the same size as the layer is calculated for the reflective coating area layer pair, and the data comparison shows the defect-free areas. The values ​​were generally below 5, with one area suspected of fading. The value is between 15 and 30.

[0080] In some embodiments, pixels with an absolute difference exceeding a preset perception threshold are marked as abnormal pixels, and their deviation from the standard value is recorded. The preset perception threshold is set differently for different channels: the threshold for the luminance channel is set to 10, the threshold for the chroma channel is set to 15, and the threshold for the saturation channel is set to 10. In the example scenario, after the calculation is completed, the difference matrix is ​​scanned, and... Pixels with values ​​greater than a preset perception threshold of 20 are marked as abnormal pixels, and the coordinates of each abnormal pixel are stored in a separate abnormal record array. , , Compared with standard value , , The deviation was measured, and the data comparison showed that a total of 152 abnormal pixels were marked on the reflective coating area layer.

[0081] Optionally, for the layer corresponding to the shell structure region, the curvature change of the local area is additionally calculated, and areas whose deviation from the standard curvature model exceeds the tolerance are marked as geometric deformation anomaly regions. The curvature change is estimated by calculating the second derivative of the image surface. In the example scene, a 9x9 pixel sliding window moves on the shell structure region layer, the average curvature of the center point of each window is calculated, and compared with the theoretical curvature value of the corresponding position in the standard curvature model. The preset tolerance is 0.05 of the absolute difference in curvature. The data comparison shows that a continuous area with a curvature deviation of 0.08 is detected in the lower left corner of the warning sign, and this area is marked as a geometric deformation anomaly region.

[0082] It is understandable that the process integrates anomalous pixel markers and geometric deformation anomaly information from all layers to generate an anomaly difference map that encodes the type and degree of deviation using different color channels or grayscale levels. In practice, a three-channel color image of the same size as the original image is created as the anomaly difference map. In the example scenario, pixels with abnormal brightness marked on the reflective coating layer are represented by the red channel (R=255), pixels with abnormal chroma are represented by the green channel (G=255), and pixels with abnormal saturation are represented by the blue channel (B=255). For pixels with multiple anomalies, the channels are overlaid. Geometric deformation anomaly areas marked on the shell structure layer are outlined with a bright yellow (R=255, G=255, B=0) contour. The final generated anomaly difference map visually displays the distribution of color attenuation and geometric deformation. Data comparison shows that non-zero pixels in the anomaly difference map account for approximately 2.1% of the total pixels in the image.

[0083] Optionally, the preset perception threshold and curvature tolerance can be dynamically configured according to the material and process standards of different models of warning facilities. It is understood that the accuracy of spatial registration directly affects the accuracy of subsequent differential operations; therefore, robust feature point detection and matching algorithms are required. In specific implementations, for printed marking area layers, spatial registration requires special attention to the fine alignment of character areas, and a non-rigid registration method based on mutual information may be used to improve accuracy.

[0084] In one embodiment of the present invention, a facility health degradation model is constructed based on the correlation mapping between environmental state parameters and anomaly difference maps. Under the guidance of the facility health degradation model, iterative feature focusing calculations are performed to extract multi-scale visual defect feature vectors. The environmental state parameters include facility service life, historical extreme weather event records, and surrounding vibration source distribution data. In an example scenario, a columnar power safety warning facility located at the entrance of an outdoor substation is analyzed. The facility has been in service for 1200 days. Historical extreme weather event records show that it has encountered a total of 8 heavy rainfall events and 2 hail events. The surrounding vibration source distribution data indicates that it is adjacent to a heavy vehicle traffic road with a daily average vibration monitoring index of 5.7. The collected anomaly difference map shows that there are scattered color abnormal pixels and a local deformation area.

[0085] In some embodiments, the statistical distribution characteristics of various abnormal pixels in the abnormal difference map are analyzed, including the density, spatial clustering, and average deviation intensity of abnormal pixels. The abnormal pixel density is calculated as the ratio of the total number of abnormal pixels to the total number of pixels in the facility area. The spatial clustering is measured by the Moran index to measure the spatial autocorrelation of abnormal pixels. The average deviation intensity is calculated as the mean of the absolute values ​​of the deviations of all abnormal pixels. In the example scenario, for color attenuation anomalies, the abnormal pixel density is 0.018, the Moran index of spatial clustering is 0.32, and the average deviation intensity is 22.5. For geometric deformation anomalies, the abnormal pixel density is 0.003, the spatial clustering is a single connected domain, and the average deviation intensity is a curvature deviation of 0.12.

[0086] In practical implementation, a correlation model is established between each parameter in the environmental state parameters and the statistical distribution characteristics of the anomaly difference map. Multivariate statistical analysis methods are used to calculate the partial correlation coefficient between facility service duration and anomaly pixel density, the Spearman rank correlation coefficient between the cumulative number of historical extreme weather events and the average deviation intensity, and the Pearson correlation coefficient between the surrounding vibration source intensity index and spatial clustering. In the example scenario, the partial correlation coefficient between facility service duration and color attenuation anomaly density is 0.58, the Spearman correlation coefficient between the cumulative number of heavy rainfall events and the average deviation intensity of color attenuation is 0.61, and the Pearson correlation coefficient between the road vibration source intensity index and the spatial clustering of geometric deformation anomalies is 0.52.

[0087] Optionally, based on a correlation model, a multiple linear regression analysis method is used to quantify the contribution weights of different combinations of environmental state parameters to various abnormal growths. The contribution weights are determined by standardized regression coefficients. In the example scenario, the regression analysis of the contribution weights of abnormal growth in color decay shows that the contribution weight of facility service time is 0.40, the contribution weight of the number of heavy rainfall events is 0.38, the contribution weight of the number of hail events is 0.15, and the contribution weight of road vibration source intensity index is 0.07.

[0088] It is understandable that, by utilizing contribution weights, a mathematical model can be constructed that can predict the potential development rate of various anomalies in the anomaly difference map based on the input environmental state parameters. This mathematical model is defined as a facility health degradation model, which is expressed by the following linear combination formula:

[0089]

[0090] in: This represents the predicted monthly potential growth rate of the k-th type of anomaly. This represents the contribution weight coefficient of the q-th environmental state parameter corresponding to the k-th type of anomaly. This represents the normalized value of the q-th environmental state parameter. This represents the total number of environmental state parameters. In the example scenario, for abnormal color decay, substituting the contribution weights and normalized environmental parameter values ​​yields... The predicted value is a monthly density increase of 0.008.

[0091] In the specific implementation, an initial analysis of interest (ROI) window is set on the anomaly difference map, centered on the potential abnormal development area predicted by the facility health degradation model. The initial window shape is rectangular, with a size of 80 pixels by 80 pixels. The center point coordinates are determined by the grid position with the highest predicted development rate. In the example scenario, the facility health degradation model predicts that the color decay anomaly has the highest development rate in the right-central region of the warning facility, and the initial ROI window is centered at coordinates (320, 240). In each iteration, the directional gradient histogram and local binary pattern features of the abnormal pixels within the current ROI window are calculated. The directional gradient histogram feature extraction is set with a cell unit size of 8x8 pixels, a block size of 2x2 cells, and 9 directional intervals. The local binary pattern features use a circular neighborhood with a radius of 3 pixels and 24 sampling points. In the example scenario, the first iteration calculates a directional gradient histogram feature vector with a dimension of 81 and a local binary pattern feature vector with a dimension of 26.

[0092] The calculated features are compared with a preset feature library of typical impact damage patterns. The similarity comparison uses the reciprocal of the Euclidean distance as a metric. If the similarity is lower than the preset threshold of 0.8, the position and size of the window of interest are adjusted according to the main direction of the histogram of the gradient magnitude within the window. When adjusting the position, the center of the window is moved by 10 pixels along the main direction. When adjusting the size, the length and width of the window are increased by 15 pixels each. In the example scenario, the similarity between the features calculated in the first iteration and the "coating peeling" pattern in the feature library is 0.72, which is lower than the threshold. The main gradient direction points to the lower right, the center of the window is adjusted to (330, 250), and the size is adjusted to 95x95 pixels. Repeat the feature calculation, comparison, and window adjustment steps until a stable window with a feature similarity to a certain type of impact damage pattern that reaches a preset threshold is found. The preset similarity threshold is 0.85. In the example scenario, after four iterations, the window center is located at (345, 260) with a size of 125x125 pixels. At this point, the similarity between the feature and the "coating peeling" pattern is 0.87, reaching the threshold. The iteration stops, and this window is determined to be a stable window. From the abnormal difference map data within the stable window, normalized multi-dimensional statistical features are extracted, including texture roughness, edge break length, and color patch size distribution. Texture roughness is obtained by calculating the standard deviation of pixel gray level difference. Edge break length is calculated by extracting the cumulative sum of Euclidean distances between the endpoints of the abnormal region after skeletonization. Color patch size distribution is obtained by counting the number of connected components in different area intervals and normalizing them to the [0,1] interval. In the example scene, the extracted texture roughness value is 0.08, the edge break length is 42 pixels, and the color patch size distribution vector is [0.1,0.3,0.4,0.2]. These normalized multi-dimensional statistical features are sequentially concatenated to form a multi-scale visual defect feature vector with a total length of 7.

[0093] In one embodiment of the present invention, a defect type matching decision is performed based on a multi-scale visual defect feature vector, and the specific defect type and location information are output. The multi-scale visual defect feature vector includes a macroscopic deformation feature sub-vector and a microscopic texture feature sub-vector. The macroscopic deformation feature sub-vector describes the distortion or indentation of the overall contour of the component, and the microscopic texture feature sub-vector describes the fine patterns of coating peeling and marking wear. In an example scenario, an outdoor triangular power safety warning sign is analyzed. The multi-scale visual defect feature vector extracted from the stable window of the abnormal difference map is [0.12, 0.05, -0.08, 0.21, 0.34, 0.02, 0.15]. The first three elements constitute the macroscopic deformation feature sub-vector, which describes the degree of contour indentation and angular offset of a corner area of ​​the warning sign. The last four elements constitute the microscopic texture feature sub-vector, which describes the graininess and local color consistency of the surface coating in that area.

[0094] In some embodiments, a multi-scale visual defect feature vector is input into a pre-trained classification decision network. The classification decision network employs a feedforward neural network structure, with an input layer having 7 neurons corresponding to the length of the feature vector. The network contains two hidden layers: the first hidden layer has 64 neurons, and the second hidden layer has 32 neurons. All hidden layers use the ReLU activation function. In the example scenario, the feature vector [0.12, 0.05, -0.08, 0.21, 0.34, 0.02, 0.15] is input into the classification decision network, and the data undergoes forward propagation computation within the network.

[0095] In practical implementation, the classification decision network contains multiple fully connected layers to map the input feature vectors to a high-dimensional space for nonlinear transformation. The output of each fully connected layer is calculated by a linear weighted sum and an activation function. In the example scenario, the multi-scale visual defect feature vector is first multiplied by the weight matrix of the first hidden layer and a bias is added. Then, a 64-dimensional intermediate feature representation is generated through the ReLU function. This intermediate feature is then transformed similarly through the second hidden layer to generate a 32-dimensional high-level abstract feature representation.

[0096] In the output layer of the classification decision network, a normalized exponential function is used to calculate the probability that the multi-scale visual defect feature vector belongs to each preset defect category. The preset defect categories include structural cracks, coating peeling, blurred markings, and combined damage. The normalized exponential function is calculated using the following formula:

[0097]

[0098] in: This represents the probability that the feature vector belongs to the y-th preset defect category. This represents the original score of the output layer corresponding to the y-th category. This represents the original score of the output layer corresponding to the k-th category. This represents the total number of preset defect categories, where K=4. In the example scenario, the original score vector of the output layer is [1.2, 3.5, 0.8, 1.0]. After calculation using the normalized exponential function, the corresponding probability distribution is obtained. For specific values, please refer to Table 1.

[0099] Table 1: Probability Distribution of Defect Categories

[0100] Preset Defect Categories Raw scores of the output layer Calculate probability Structural cracks 1.2 0.18 Coating peeling 3.5 0.68 blurry logo 0.8 0.10 Complex damage 1.0 0.04

[0101] The system selects the preset defect category with the highest probability as the final identified defect type. In the example scenario, according to the defect category probability distribution table, the "coating peeling" category has the highest probability of 0.68. Therefore, the system determines the final identified defect type as "coating peeling".

[0102] In some embodiments, the coordinates of the stable window generating multi-scale visual defect feature vectors in the original image sequence are back-mapped to the physical location coordinates on the power safety warning facility, generating positioning information. The mapping process is based on pre-calibrated camera intrinsic and extrinsic parameters and the approximate size of the facility in the world coordinate system. In the example scenario, the pixel coordinates of the stable window generating features in the image are (300, 200) at the top left corner and (425, 325) at the bottom right corner. Through coordinate transformation matrix calculation, the position of this window on the physical surface of the power safety warning sign is obtained as a rectangular area with the bottom left corner of the sign as the origin, ranging from 0.15 meters to 0.28 meters in the X direction and from 0.40 meters to 0.55 meters in the Y direction.

[0103] See Figure 4 In the analysis of facility health degradation, the characterization of multi-stage defect development trends relies on the synergistic evolution of three core characteristics: macroscopic deformation deviation, microscopic damage proportion, and vibration source influence index. Specifically, the facility damage stages are divided into five levels: initial stage, minor damage, moderate damage, severe damage, and critical state. Macroscopic deformation deviation reflects the degree of geometric offset of the overall facility structure; microscopic damage represents the cumulative proportion of local defects such as coating peeling and marking wear; and the vibration source influence index quantifies the driving effect of surrounding environmental vibration on defect development. From the curve evolution pattern, all three show a monotonically increasing trend with the progression of the damage stage. The growth rate of microscopic damage proportion and vibration source influence index is more significant, approaching 0.6 in the critical state stage, while the growth of macroscopic deformation deviation is relatively slow, approximately 0.38 in the critical state stage. This reveals that microscopic damage is more sensitive to environmental vibration and has a significant lag correlation with macroscopic deformation.

[0104] In one embodiment of the present invention, the image sequence is preprocessed and enhanced after acquisition, and the severity of defects is quantified after the output defect information is processed. In an example scenario, an image sequence containing an outdoor cylindrical power safety warning facility was acquired. The facility has been in service for approximately 900 days. The image sequence contains 30 frames of RGB images with a resolution of 1280x720. Random noise was introduced into the images during transmission and acquisition. Furthermore, due to uneven lighting caused by cloudy weather during shooting, there were obvious shadows on the surface of the facility, and the contrast between the reflective coating area and the dark background was insufficient.

[0105] In practical implementation, after acquiring image sequences of power safety warning facilities in an outdoor environment and obtaining corresponding environmental state parameters, the process includes preprocessing and enhancing the image sequences. Temporal filtering is performed on the image sequences to eliminate random noise. Temporal filtering uses the median value of the grayscale values ​​of the same pixel location across multiple consecutive frames in the image sequence. In the example scenario, for a continuous 5-frame image sequence, the median value of the R, G, and B channels of each pixel across the 5 frames is taken as the filtered output value. Data comparison shows that the noise variance of a single frame image after filtering decreases from the estimated 4.7 to 1.2.

[0106] A histogram equalization algorithm is used to adjust the contrast of each frame of the image to enhance the distinction between the reflective coating area and the background. Specifically, the histogram equalization algorithm adjusts the contrast of each frame of the image to enhance the distinction between the reflective coating area and the background. This involves calculating the global luminance histogram of the input image. The global luminance histogram is obtained by converting the RGB image to a grayscale image and counting the pixel frequency of each grayscale level (0-255). A cumulative distribution function for the global luminance histogram is constructed, and this cumulative distribution function is used to remap the luminance value of each pixel in the input image, generating a contrast-enhanced image that makes the pixel luminance distribution in the reflective coating area more concentrated. The remapping follows the formula:

[0107]

[0108] in: This represents the new pixel brightness value at coordinates (x, y) in the output image. This represents the original brightness value of the pixel at coordinates (x, y) in the input image. This indicates the possible number of gray levels in the image (usually 256). This represents the probability of gray level i appearing in the original image. This indicates a rounding down operation. In the example scene, the pixel brightness of the reflective coating area in the original image is mainly distributed between 120 and 150, while the background brightness is distributed between 30 and 60. After applying the histogram equalization algorithm, the pixel brightness of the reflective coating area is remapped to the higher range of 180 to 220, while the background brightness is compressed to the lower range of 10 to 40. The distinction between the two is significantly increased. Data comparison shows that the normalized contrast ratio between the target area and the background increases from 0.35 to 0.62.

[0109] In some embodiments, photometric compensation is performed on shadowed areas caused by uneven illumination to make the brightness distribution of the entire facility surface more uniform. The photometric compensation adopts a single-scale retinal enhancement method based on Retinex theory, estimating the illumination components and performing division operations. In the example scenario, a large area of ​​shadow was identified on the left side of the facility image. The estimated average brightness of this area was only 65% ​​of that of the non-shadowed area on the right. After photometric compensation, the average brightness of the shadowed area on the left increased to 92% of that of the area on the right, and the standard deviation of the brightness distribution of the entire image decreased from 48.3 before compensation to 19.5.

[0110] In practical implementation, after outputting the specific defect type and location information of the power safety warning facility, a step of quantifying the defect severity is also included. Based on the values ​​of the macroscopic deformation feature sub-vectors in the multi-scale visual defect feature vector, the overall offset of the structure from the standard shape is calculated. In the example scenario, the macroscopic deformation feature sub-vectors are [0.18, -0.05, 0.12], representing deformation in the X direction, deformation in the Y direction, and angular deflection, respectively. The overall offset O is obtained by calculating the weighted sum of the absolute values ​​of each element of this sub-vector, as shown in the formula: .

[0111] Based on the values ​​of the micro-texture feature sub-vectors in the multi-scale visual defect feature vector, the ratio of the defect area to the total apparent area of ​​the facility is calculated. The micro-texture feature sub-vectors contain normalization parameters representing the defect area. In the example scenario, the micro-texture feature sub-vectors are [0.02, 0.15, 0.08, 0.21], where the fourth element, 0.21, directly corresponds to the normalized value of the defect area ratio R. Through inverse normalization calculation, the actual defect area ratio is obtained. (The maximum possible percentage is 5%).

[0112] In some embodiments, a comprehensive defect severity level is obtained by combining the overall offset and the proportion through a predefined quantitative scoring table. The defect severity level is stored together with the defect type and location information. Optionally, the predefined quantitative scoring table is a two-dimensional lookup table. The row index corresponds to the discretized interval of the overall offset O, and the column index corresponds to the discretized interval of the defect area proportion R. Each cell stores an integer level from 1 to 5, where 1 represents slight and 5 represents severe. In the example scenario, O=0.265 and R=1.05% are calculated. Querying the quantitative scoring table, the O value falls into the interval [0.2, 0.3), and the R value falls into the interval [1.0%, 2.0%). The corresponding cross cell stores a level of 3. It can be understood that the quantitative scoring table is based on historical operation and maintenance data and expert experience. In specific implementation, the final "coating peeling" defect type, location information "X: [0.15, 0.28]m, Y: [0.40, 0.55]m" and defect severity level "3" are associated and stored in the database. Optionally, the weights of different deformation components in the calculation of the overall offset O can be adjusted according to the facility type. It can be understood that the severity level of defects provides a quantitative basis for prioritizing maintenance operations.

[0113] See Figure 5 The graph, plotted with time on the horizontal axis and defect severity level on the vertical axis, visually illustrates the evolution patterns under two scenarios. The defect development curve without maintenance (red) starts from the current level 3.0 and shows a continuous upward trend over time, reaching 4.0 after 12 months, reflecting the accelerated deterioration of defect severity under natural evolution. The effect curve after maintenance intervention (green) starts from the same initial level and gradually decreases to 2.0 after 12 months, verifying the inhibitory effect of maintenance measures on defect development. The current level is fixed at 3.0 by a gray dashed line, serving as the benchmark for the two evolution curves. This quantitative comparison provides a visual basis for verifying the facility health degradation model and evaluating the effectiveness of maintenance strategies. The defect growth rate in the no-maintenance scenario (0.083 levels / month) and the defect mitigation rate in the maintenance scenario (0.083 levels / month) exhibit symmetrical evolution characteristics, which can be used for marginal benefit analysis of subsequent maintenance resource investment.

[0114] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for extracting and recognizing visual defects of an impact-resistant electric power safety warning facility, characterized in that, The method includes: Collect image sequences of electrical safety warning facilities in outdoor environments and obtain corresponding environmental status parameters; Each frame of the image sequence is deconstructed in multiple dimensions of physical and apparent states to obtain component deconstruction images including the shell structure area, reflective coating area, and printed marking area. The component deconstruction image is compared with a preset standard template image by pixel-level difference operation to generate an abnormal difference map reflecting the component's geometric deformation and color attenuation; Based on the correlation mapping between the environmental state parameters and the abnormal difference map, a facility health degradation model is constructed. Guided by the facility health degradation model, iterative feature focusing calculations are performed on the abnormal difference map to extract multi-scale visual defect feature vectors corresponding to the impact damage pattern. Based on the multi-scale visual defect feature vector, a defect type matching decision is performed, and the specific defect type and location information of the power safety warning facility are output.

2. The visual imperfection feature extraction and recognition method of the impact-resistant type electric power safety warning facility according to claim 1, characterized in that, The step of performing multi-dimensional deconstruction of the physical and apparent states of each frame in the image sequence to obtain component deconstruction images including the shell structure area, reflective coating area, and printed marking area includes: An edge-aware segmentation algorithm is used to initially divide the image, resulting in a mask that includes the main structure and the background. Within the facility body area defined by the mask, a group of pixels with high reflectivity is separated according to a color space clustering algorithm, and the group of pixels with high reflectivity constitutes the reflective coating area. Within the facility body area defined by the mask, a set of pixels with a continuous flat surface is defined according to the texture consistency analysis algorithm, and the set of pixels with a continuous flat surface constitutes the outer shell structure area. Within the facility's main area defined by the mask, optical character recognition pre-detection technology is used to locate pixel blocks containing specific combinations of characters or symbols, and these pixel blocks containing specific combinations of characters or symbols constitute the printed marking area. The reflective coating area, the outer shell structure area, and the printed marking area are extracted and stored as independent image layers, which together form the component deconstruction image.

3. The method for extracting and identifying visual defect features of impact-resistant power safety warning facilities according to claim 2, characterized in that, The step of performing pixel-level difference operations between the component deconstruction image and a preset standard template image to generate an anomaly difference map reflecting the component's geometric deformation and color attenuation includes: Spatial registration is performed between each independent image layer in the component deconstruction image and the corresponding standard layer in the preset standard template image; For each registered layer pair, calculate the absolute difference in the three channels of brightness, chroma and saturation pixel by pixel; Pixels whose absolute value difference exceeds a preset perception threshold are marked as abnormal pixels, and their deviation from the standard value is recorded. For the layer corresponding to the shell structure region, the curvature change of the local region is calculated in addition, and the region whose deviation from the standard curvature model exceeds the tolerance is marked as a geometric deformation anomaly region. By integrating abnormal pixel markers and geometric deformation abnormal area information from all layers, an abnormality map is generated that encodes the type and degree of deviation using different color channels or grayscale levels.

4. The method for extracting and identifying visual defect features of impact-resistant power safety warning facilities according to claim 1, characterized in that, The step of constructing a facility health degradation model based on the correlation mapping between the environmental state parameters and the anomaly difference map includes: The environmental status parameters include facility service life, historical extreme weather event records, and surrounding vibration source distribution data; Analyze the statistical distribution characteristics of various abnormal pixels in the abnormal difference map, including the density, spatial clustering and average deviation intensity of abnormal pixels; Establish a correlation model between each parameter in the environmental state parameters and the statistical distribution characteristics of the anomaly difference map; Based on the aforementioned correlation model, regression analysis was used to quantify the contribution weights of different combinations of environmental state parameters to various abnormal growths. Using the contribution weights, a mathematical model is constructed that can predict the potential development rate of various anomalies in the anomaly difference map based on the input environmental state parameters. The mathematical model is defined as the facility health degradation model.

5. The method for extracting and identifying visual defect features of impact-resistant power safety warning facilities according to claim 4, characterized in that, Guided by the facility health degradation model, the abnormal difference map is subjected to iterative feature focusing calculation to extract multi-scale visual defect feature vectors corresponding to the impact damage pattern, including: An initial analysis window of interest is set on the anomaly difference map, centered on the potential abnormal development area predicted by the facility health degradation model. In each iteration, the directional gradient histogram and local binary pattern features of the abnormal pixels within the current analysis window of interest are calculated; The calculated features are compared with a preset feature library of typical impact damage patterns. If the similarity is lower than the threshold, the position and size of the analysis window of interest are adjusted according to the gradient information. Repeat the feature calculation, comparison and window adjustment steps until a stable window with a similarity to a certain type of impact damage pattern feature reaches a preset threshold is found; From the abnormal difference map data within the stable window, normalized multi-dimensional statistical features are extracted, including but not limited to texture roughness, edge break length, and color patch size distribution. The normalized multi-dimensional statistical features are then combined into the multi-scale visual defect feature vector.

6. The method for extracting and identifying visual defect features of impact-resistant power safety warning facilities according to claim 5, characterized in that, The multi-scale visual defect feature vector includes macroscopic deformation feature vectors and microscopic texture feature vectors. The macroscopic deformation feature vectors describe the distortion or depression of the overall contour of the component, and the microscopic texture feature vectors describe the fine patterns of coating peeling and wear marks.

7. The method for extracting and identifying visual defect features of impact-resistant power safety warning facilities according to claim 6, characterized in that, The process of performing defect type matching decisions based on the multi-scale visual defect feature vector, and outputting the specific defect type and location information of the power safety warning facility, includes: The multi-scale visual defect feature vectors are input into a pre-trained classification decision network; The classification decision network contains multiple fully connected layers, which are used to map the input feature vector to a high-dimensional space for nonlinear transformation. In the output layer of the classification decision network, a normalized exponential function is used to calculate the probability that the multi-scale visual defect feature vector belongs to each preset defect category. The preset defect categories include structural cracks, coating peeling, blurry markings, and composite damage. Select the preset defect category with the highest probability as the final identified defect type; Simultaneously, the coordinate position of the stable window that generates the multi-scale visual defect feature vector in the original image sequence is reverse-mapped to the physical location coordinates on the power safety warning facility to generate the positioning information; The defect type is associated with the location information and output.

8. The method for extracting and identifying visual defect features of impact-resistant power safety warning facilities according to claim 1, characterized in that, After acquiring image sequences of electrical safety warning facilities in an outdoor environment and obtaining corresponding environmental state parameters, the method further includes preprocessing and enhancing the image sequences: Temporal filtering is performed on the image sequence to eliminate random noise; Histogram equalization algorithm is used to adjust the contrast of each frame of the image to enhance the distinction between the reflective coating area and the background. Photometric compensation is applied to the shadow areas caused by uneven lighting, so that the brightness distribution of the entire facility surface tends to be uniform.

9. The method for extracting and identifying visual defect features of impact-resistant power safety warning facilities according to claim 8, characterized in that, The histogram equalization algorithm is used to adjust the contrast of each frame of the image to enhance the distinction between the reflective coating area and the background. Specifically: Calculate the global brightness histogram of the input image; Construct the cumulative distribution function of the global brightness histogram; The cumulative distribution function is used to remap the brightness value of each pixel in the input image to generate a contrast-enhanced image, making the pixel brightness distribution in the reflective coating area more concentrated.

10. The method for extracting and identifying visual defect features of impact-resistant power safety warning facilities according to claim 1, characterized in that, After outputting the specific defect type and location information of the power safety warning facility, the method further includes a step of quantifying the severity of the defect: Based on the values ​​of the macroscopic deformation feature sub-vectors in the multi-scale visual defect feature vector, the overall offset of the structure from the standard shape is calculated. Based on the values ​​of the micro-texture feature sub-vectors in the multi-scale visual defect feature vector, the ratio of the defect area to the total apparent area of ​​the facility is calculated. By combining the overall offset with the proportion, a comprehensive defect severity level is obtained through a predefined quantitative scoring table. The defect severity level is stored together with the defect type and location information.

Images