Intelligent detection method for angle steel of electric power iron tower

Abstract

The invention relates to the technical field of nondestructive testing and computer vision, in particular to an intelligent detection method for angle steel of an electric power iron tower, which comprises the following steps: acquiring a visible light image, an infrared thermal image and three-dimensional laser point cloud data of target angle steel; performing high-precision registration on the visible light image, the infrared thermal image and the three-dimensional laser point cloud to enable data of different modalities to form a pixel-level corresponding relation in space; carrying out cooperative processing on the registered multi-modal data; according to predefined defect discrimination logic based on a multi-modal feature coupling relationship, performing accurate identification and classification of defect types; and outputting a detection result containing defect types, levels and quantitative information. Through collaborative analysis of multi-modal data, geometric deformation, apparent anomaly and thermal anomaly of the angle steel can be comprehensively captured, and the accuracy and comprehensiveness of detection are greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of non-destructive testing and computer vision technology, and in particular to an intelligent inspection method for angle steel of power transmission towers. Background Technology

[0002] As a core infrastructure of power transmission networks, power transmission towers are exposed to complex outdoor environments for extended periods, making their main structural material, angle steel, susceptible to deformation, corrosion, and fastener loss. Current technologies primarily rely on manual climbing and inspection, using telescopes for observation or close-range measurements. This approach is inefficient, dangerous, subjective, and difficult to quantify.

[0003] In recent years, some detection methods based on unmanned aerial vehicle (UAV) imagery have emerged, but these methods generally suffer from the following inherent drawbacks: The contradiction between feature singularity and environmental interference: Existing methods mostly rely on single visible light or infrared images. Visible light images are easily affected by changes in lighting, shadows, and cluttered backgrounds; infrared images are extremely sensitive to ambient temperature, humidity, and the emissivity of the coating (paint) on the angle steel surface, leading to unstable defect feature extraction.

[0004] Missing 3D geometric information: Detection methods based on 2D images cannot accurately obtain 3D spatial deformation information such as bending and twisting of angle steel, which is a key indicator for assessing the safety of iron tower structures.

[0005] Superficial multi-source data fusion: Even with the use of multiple sensors, existing technologies often remain at the level of simple image overlay or parallel information processing, failing to achieve pixel-level, physically meaningful deep feature fusion. This results in low recognition rates and high false alarm rates for minor or early defects (such as shallow corrosion or slight warping).

[0006] Poor adaptability: Existing algorithms are not robust enough when the edges of angle steel are obscured, the surface is stained, or the shape changes under different observation angles. A lot of manual intervention is required for parameter tuning and result correction.

[0007] Therefore, there is an urgent need in this field for an intelligent detection method for angle steel of power transmission towers that can overcome the above-mentioned defects and achieve high precision, high robustness, and full automation. Summary of the Invention

[0008] To achieve the above objectives, the present invention provides an intelligent detection method for angle steel of power transmission towers, comprising the following steps: The drone platform equipped with a visible light camera, infrared thermal imager and lidar is controlled to perform contour flight of the target iron tower, and simultaneously collect visible light images, infrared thermal images and three-dimensional laser point cloud data of the target angle steel. Based on the pre-calibrated transformation relationship and the inherent structural features of angle steel extracted from the data, the visible light image, the infrared thermal image and the three-dimensional laser point cloud are registered with high precision, so that the data of different modes form a pixel-level correspondence in space; The registered multimodal data are processed collaboratively to extract a first feature map characterizing the geometric deformation of the angle steel, a second feature map characterizing the apparent anomaly, and a third feature map characterizing the thermal anomaly. The first feature map, the second feature map, and the third feature map are then fused according to a preset rule to generate a comprehensive anomaly score map. Defect candidate regions are then extracted from the comprehensive anomaly score map through adaptive threshold segmentation. For each of the defect candidate regions, the combination of its response patterns on the first feature map, the second feature map, and the third feature map is analyzed, and the defect type is accurately identified and classified according to a predefined defect discrimination logic based on multimodal feature coupling relationship. The output includes detection results containing defect type, level, and quantitative information.

[0009] Preferably, the high-precision registration of the visible light image, the infrared thermal image, and the three-dimensional laser point cloud based on the pre-calibrated transformation relationship and the inherent structural features of the angle steel extracted from the data specifically includes: Using the initial transformation matrix obtained through offline calibration between the lidar, the visible light camera, and the infrared thermal imager, preliminary spatial alignment is performed on the three-dimensional lidar point cloud, the visible light image, and the infrared thermal image; From the pre-aligned visible light image, edge pixels of the angle steel are extracted using an edge detection algorithm based on gray-level gradient, and straight line segments representing the edge of the angle steel are fitted from the edge pixels using Hough transform or probabilistic Hough line detection algorithm as two-dimensional image features. From the three-dimensional laser point cloud, point cloud clusters belonging to the same angle steel are separated using a point cloud clustering algorithm based on Euclidean distance. Then, multiple planar models are iteratively fitted from the point cloud clusters using a random sampling consensus algorithm. The intersection lines of different planes are extracted as three-dimensional line segment features, which are used as three-dimensional structural features. The three-dimensional line segment features are projected onto a two-dimensional image plane according to the initial transformation matrix to form a projection line segment set. The projection line segment set is then iteratively matched with the two-dimensional image features extracted from the visible light image using nearest-neighbor optimization. This optimization process uses the endpoints of the projection line segments as the initial point set and the nearest point on the line segment extracted from the image as the target point set. A fine rotation and translation transformation matrix is ​​iteratively optimized by minimizing the average Euclidean distance between the initial point set and the target point set. The optimization terminates when the change in the average Euclidean distance is less than a convergence threshold pre-derived based on the point cloud accuracy and image resolution. The final refined rotation and translation transformation matrix is ​​applied to the infrared thermal image and the visible light image to complete pixel-level high-precision registration with the three-dimensional laser point cloud.

[0010] Preferably, the extraction of the first feature map characterizing the geometric deformation of the angle steel specifically includes: For the registered 3D laser point cloud, take each laser point as the core and take a set of points within a certain neighborhood around it; the size of the neighborhood is adaptively determined according to the physical size of the angle steel and the point cloud density to ensure that a local area of ​​the angle steel can be covered. For each set of points extracted, an ideal plane model is fitted using the least squares method; Calculate the perpendicular distance from each point in the point set to its corresponding ideal plane model; Traverse all points in the three-dimensional laser point cloud, and map the calculated distance value of each point back to its corresponding pixel position to form a two-dimensional grayscale image, where the grayscale value of each pixel represents the distortion of the spatial point in the normal direction. This grayscale image is the first feature mapping map.

[0011] Preferably, the extraction of the second feature map representing the apparent anomaly specifically includes: On the registered visible light image, a sliding window is defined; the size of the sliding window is adjusted according to the imaging scale of the angle steel in the image to ensure that the window can contain sufficient texture information. For each position traversed by the sliding window on the image, the variance of all pixels within the window across multiple channels of the color space is calculated; the color space is chosen based on its sensitivity to apparent changes such as corrosion. The variance values ​​of multiple channels are weighted and fused to obtain a comprehensive apparent anomaly intensity value. Assign the combined apparent anomaly intensity value to the center pixel of the sliding window; After traversing the entire image, a grayscale image of the same size as the original image is generated, where the grayscale value of each pixel represents the degree of appearance abnormality of its surrounding area. This grayscale image is the second feature map.

[0012] Preferably, the extraction of the third feature map characterizing the thermal anomaly specifically includes: On the infrared thermal image that has been registered and temperature value conversion completed, the normal temperature area of ​​the angle steel surface is first determined; the normal temperature area is obtained by clustering and identifying large, continuous areas with small temperature fluctuations in the image. Calculate the statistical average of the temperature values ​​of all pixels within the normal temperature area, and use it as the reference normal temperature; Define a sliding window, and for each position that the sliding window traverses on the thermal image, calculate the average value of the temperature values ​​of all pixels within the window; Calculate the absolute difference between the average temperature of the window and the reference normal temperature; The absolute difference is assigned to the center pixel of the sliding window; After traversing the entire thermal image, a grayscale image of the same size as the thermal image is generated, where the grayscale value of each pixel represents the degree of temperature deviation of its surrounding area relative to the normal area. This grayscale image is the third feature map.

[0013] Preferably, the first feature map, the second feature map, and the third feature map are fused according to a preset rule, wherein the preset rule is: A weight coefficient is assigned to each of the first, second, and third feature maps. The weight coefficient assignment strategy is based on the dominant features exhibited by different defect types: for deformation defects, the first feature map is assigned the highest weight coefficient; for corrosion defects, the second and third feature maps are assigned relatively high weight coefficients, and the weight ratio between them is fine-tuned according to the ambient temperature. The specific values ​​of the weight coefficients are obtained by optimizing the machine learning training on a known defect sample library. The pixel grayscale values ​​of the three feature maps are multiplied by their respective weight coefficients and then linearly superimposed to obtain the comprehensive anomaly score map.

[0014] Preferably, the step of extracting defect candidate regions from the comprehensive anomaly score map through adaptive threshold segmentation specifically involves using the maximum inter-class variance method to adaptively determine the segmentation threshold. Calculate the gray-level histogram distribution of the comprehensive anomaly score map; By iterating through all possible grayscale values ​​as candidate thresholds, image pixels are divided into foreground and background categories. Calculate the mean gray value of foreground and background pixels, and their respective proportions in the image; Based on the formula for calculating the inter-class variance of foreground and background, calculate the inter-class variance value corresponding to each candidate threshold; the inter-class variance value is a function of foreground probability, background probability, and the difference between foreground mean and background mean; Select the candidate gray value that maximizes the inter-class variance as the final segmentation threshold; The final threshold is used to binarize the comprehensive anomaly score map, and connected regions are filtered. Regions with an area greater than a preset lower limit are marked as candidate defect regions. The preset lower limit is calculated based on the image resolution and the minimum physical size requirement of the defect.

[0015] Preferably, in the predefined defect discrimination logic based on multimodal feature coupling, the judgment logic for corrosion defects is as follows: First, check whether the average response intensity of the defect candidate region on the second feature map exceeds the first dynamic threshold; the first dynamic threshold is obtained by statistically analyzing the distribution of all pixel values ​​in the entire second feature map and taking the value corresponding to a certain percentile as the threshold. Simultaneously, it checks whether the average response intensity of the defect candidate region on the third feature map exceeds the second dynamic threshold; the method for determining the second dynamic threshold is the same as that for the first dynamic threshold, but based on the global statistical distribution of the third feature map; It is also necessary to check whether the average response intensity of the defect candidate region on the first feature map is lower than the third dynamic threshold; the method for determining the third dynamic threshold is similar to the above threshold. The candidate area is determined to be a corrosion defect only if all three conditions are met simultaneously.

[0016] Preferably, in the predefined defect discrimination logic based on multimodal feature coupling, the judgment logic for deformation defects is as follows: First, check whether the average response intensity of the defect candidate region on the first feature map exceeds the fourth dynamic threshold. The fourth dynamic threshold is obtained by statistically analyzing the distribution of all pixel values ​​in the entire first feature map and taking the value corresponding to a certain percentile as the threshold. The check is performed to determine whether the response distribution of the defect candidate region on the first feature map exhibits a continuous and directional gradient change. This check is accomplished by calculating the standard deviation of pixel values ​​within the region and the consistency of gradient direction. Simultaneously check whether the average response intensity of the defect candidate region on the second feature map and the third feature map is lower than the fifth dynamic threshold and the sixth dynamic threshold; The candidate region is determined to be a deformation defect only if all of the above conditions are met.

[0017] Preferably, in the output of the detection results containing defect type, level, and quantification information, the defect level is classified based on the following criteria: For corrosion defects, their level is determined by a lookup table method based on the predefined intensity range of the average response intensity value of the defect candidate region on the second feature map and the predefined area range of the pixel area of ​​the region. For deformation defects, their level is determined by a lookup table method based on the predefined depth range where the maximum distortion value of the defect candidate region is located in the first feature map and the predefined area range to which the pixel area of ​​the region belongs. The boundary values ​​of the predefined intensity range, depth range, and area range are determined by statistical analysis of the severity of known defects in historical detection data and by combining the experience and knowledge of domain experts.

[0018] The beneficial effects of this invention are: 1. This invention employs multimodal data fusion, combining visible light images, infrared thermal images, and 3D laser point clouds. This combination of multi-source data compensates for the shortcomings of a single image mode: visible light images provide rich appearance information, infrared images help detect thermal anomalies, and 3D laser point clouds accurately capture the geometric information of the angle steel. Through high-precision registration and collaborative processing, these information complement each other, effectively reducing environmental interference and improving the stability and accuracy of feature extraction.

[0019] 2. This invention employs three-dimensional laser point cloud technology. By accurately acquiring the geometric deformation of the angle steel, it can not only obtain more detailed spatial structural information but also more accurately assess the safety of the tower structure. Furthermore, by fusing three-dimensional laser point cloud data with data from other sensors, defect detection can not only rely on surface information but also assess potential structural problems through geometric deformation.

[0020] 3. This invention maximizes the advantages of each modality by performing pixel-level registration and depth feature fusion on three modalities. It utilizes temperature anomalies in infrared images to detect thermal defects, combined with precise extraction of geometric deformation from 3D point clouds, achieving a high recognition rate for early defects and significantly reducing the false alarm rate.

[0021] 4. This invention improves the algorithm's adaptability to complex scenes by employing a defect discrimination logic based on multimodal feature coupling. Even in cases of occlusion or surface contamination, geometric information can still be supplemented using 3D laser point clouds, while infrared thermal imaging can capture surface temperature anomalies, reducing misjudgments caused by image quality issues. Furthermore, the algorithm of this invention can automatically adapt to different observation angles and environmental conditions without requiring excessive manual intervention, significantly improving the system's robustness and adaptability. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a flowchart illustrating the steps of the method for determining corrosion defects in this invention. Figure 3 The flowchart illustrates the steps used to classify the defect levels in the method of this invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0025] Please see Figures 1-3 This invention provides an intelligent detection method for angle steel of power transmission towers. The method first controls a drone platform equipped with a visible light camera, infrared thermal imager, and lidar to perform contour-following flight over the target power transmission tower. In this step, the drone platform can fly precisely along a preset route and simultaneously acquire visible light images, infrared thermal images, and three-dimensional laser point cloud data of the target angle steel. This simultaneous acquisition of multimodal data ensures comprehensive information on different angles and features at the same time, providing not only the appearance data of the angle steel (visible light image) but also its thermal state (infrared thermal image) and precise spatial geometric information (three-dimensional laser point cloud). This technology solves the problem of traditional manual inspections being unable to simultaneously acquire multi-source information, improving the efficiency and comprehensiveness of data acquisition.

[0026] After data acquisition, based on pre-calibrated transformation relationships and the inherent structural characteristics of angle steel, this method performs high-precision registration of images from different modalities. Through precise geometric and spectral correction, pixel-level correspondences can be established between data from different modalities in space. The core technology of this step lies in using the spatial information of three-dimensional laser point clouds to align the image data with coordinates, enabling data from different sensors to be fused in the same coordinate system. The beneficial effect of this step is that it eliminates data deviations caused by differences in sensors or shooting angles, thereby ensuring the accuracy of subsequent analysis.

[0027] The registered multimodal data will undergo feature extraction through collaborative processing. Specifically, the first feature map represents the geometric deformation of the angle steel, the second feature map represents apparent anomalies (such as surface corrosion and stains), and the third feature map represents thermal anomalies (such as localized overheating). In this process, the three different data types complement each other through specific algorithms, improving the ability to identify different types of defects. For example, infrared thermography can reveal thermal anomalies, while 3D point clouds can accurately capture the geometric deformation information of the angle steel. The combination of these two methods can effectively diagnose early, minute defects that are difficult to identify using traditional methods.

[0028] The extracted feature maps are fused to generate a comprehensive anomaly score map. The fusion rule is based on the importance weights of each feature map in different dimensions, forming a unified scoring standard. Subsequently, an adaptive threshold segmentation method is used to extract candidate defect regions from the comprehensive anomaly score map. This technique effectively avoids the problem of high false alarm rates in traditional methods and can accurately identify real defect regions from massive amounts of data.

[0029] For each candidate defect region, this method analyzes its response pattern on three feature maps. Through a predefined defect discrimination logic based on multimodal feature coupling, it performs accurate identification and classification of defect types. This step combines the advantages of different modal data, accurately distinguishing different types of defects (such as corrosion, cracks, and thermal damage), and further assigning different levels and quantitative indicators to each defect through a classification system. Through this deep fusion and intelligent analysis, this method significantly improves the accuracy and automation of defect detection.

[0030] This invention overcomes the shortcomings of existing technologies and improves the robustness and accuracy of intelligent detection methods for angle steel in power transmission towers by employing collaborative processing and high-precision registration of multimodal data. First, it comprehensively collects data from different types of sensors, providing a richer information source for defect identification. Second, based on high-precision registration technology using multimodal fusion, data processing achieves precise alignment, eliminating errors caused by data inconsistencies. Finally, through adaptive threshold segmentation and deep feature analysis, it effectively improves the detection rate of minute defects while reducing the false alarm rate. The implementation of this method not only improves the automation level of power transmission tower inspection but also significantly reduces the risks and costs of manual inspections, demonstrating broad application prospects.

[0031] In one possible implementation, an initial transformation matrix obtained through offline calibration between the lidar, visible light camera, and infrared thermal imager is first used to perform preliminary spatial alignment of the 3D lidar point cloud, visible light image, and infrared thermal image. The offline calibration process involves calculating the transformation matrix among the three devices by controlling their precise positions and angles within the experimental environment. The purpose of this step is to ensure that data acquired by different sensors can be processed and fused within a unified spatial coordinate system, laying the foundation for subsequent fine registration.

[0032] After initial alignment, edge pixels of the angle steel are extracted from the visible light image using a gray-level gradient-based edge detection algorithm. This algorithm identifies the edge portions of the angle steel's ridges, and this edge information serves as the angle steel's two-dimensional features. Subsequently, a Hough transform or probabilistic Hough line detection algorithm is used to fit straight line segments representing the angle steel's ridges from these edge pixels. These straight line segments are the extracted two-dimensional image features, representing the angle steel's geometric structure. This feature extraction process accurately captures the edge information of the angle steel, aiding in subsequent registration and defect identification.

[0033] Next, point cloud clusters belonging to the same angle steel are separated from the 3D laser point cloud using a point cloud clustering algorithm based on Euclidean distance. Using this point cloud data, multiple planar models are iteratively fitted using a random sampling consensus algorithm. The intersection lines of these planar models are extracted as 3D line segment features, representing the 3D structural information of the angle steel. Unlike 2D image features, 3D line segment features provide a more accurate description of the spatial geometric structure, forming the basis for high-precision registration and defect analysis.

[0034] The 3D line segment features are projected onto the 2D image plane according to an initial transformation matrix, forming a projected line segment set. An iterative nearest-point optimization matching algorithm is then used to match the projected line segment set with the 2D image features extracted from the visible light image. In this process, the endpoints of the projected line segments are used as the initial point set, and the nearest points on the line segments extracted from the image are used as the target point set. A refined rotation and translation transformation matrix is ​​iteratively optimized by minimizing the average Euclidean distance between the initial and target point sets. This optimization process ensures precise alignment between the image and the laser point cloud, eliminating registration deviations caused by sensor errors or different shooting angles.

[0035] The optimization process terminates when the change in the average Euclidean distance is less than a pre-set convergence threshold. The resulting rotation-translation transformation matrix is ​​then the fine-grained registration result. This matrix can be used to perform pixel-level high-precision registration of infrared thermal images and visible light images with 3D laser point clouds. The advantage of this step is that it enables high-precision alignment of data from different modalities (images and point clouds), thus providing accurate data support for subsequent defect detection and analysis.

[0036] Through meticulous registration and optimization steps, not only is the accuracy of data fusion improved, but the reliability of the defect detection system is also enhanced. It is particularly suitable for the intelligent detection of angle steel in power transmission towers and has high practical application value.

[0037] In one possible implementation, the registered 3D laser point cloud is first processed. Each laser point serves as a core, and a set of points within a certain neighborhood of it is selected. The size of the neighborhood is adaptively determined by the physical dimensions of the angle steel and the density of the point cloud. Specifically, the physical dimensions of the angle steel determine the maximum boundary of the neighborhood, and the point cloud density determines the minimum boundary. This adaptive selection of the neighborhood ensures that, under any conditions, the selected point set can cover the local area of ​​the angle steel, thereby guaranteeing the stability and accuracy of the processing.

[0038] For each selected point set, the least squares method is used to fit an ideal planar model. The least squares method is a commonly used optimization algorithm that aims to minimize the error between the fitted plane and the actual point set, thereby obtaining the best fit result. The fitted ideal planar model can accurately describe the geometry of the region containing each point set, providing a foundation for subsequent geometric deformation detection.

[0039] For each point in each point set, calculate its perpendicular distance to the fitted ideal plane model. The perpendicular distance reflects the degree of deviation of each point in the point cloud from the ideal plane, that is, the degree of geometric deformation. This step, by calculating the distance between each point and the ideal plane, can accurately detect the local deformation of the angle steel surface, such as bending and twisting.

[0040] The algorithm iterates through all points in the entire 3D laser point cloud, mapping the vertical distance value of each point back to its corresponding 2D pixel position to form a grayscale image. In this grayscale image, the grayscale value of each pixel represents the distortion of that spatial point in the normal direction. Regions with larger grayscale values ​​indicate larger deformation of the angle steel, while regions with smaller grayscale values ​​indicate smaller deformation or no deformation. This grayscale image is the "first feature map" mentioned in this method.

[0041] The extraction of the first feature map characterizing the geometric deformation of angle steel, by accurately calculating the vertical distance of the point cloud and mapping it to a grayscale image, not only enables efficient and accurate detection of the geometric deformation on the surface of angle steel, but also enhances the sensitivity and accuracy of detection, which is of great value for improving the safety monitoring level of power transmission towers.

[0042] In one possible implementation, a sliding window is first defined on the registered visible light image. The size of the sliding window needs to be adjusted according to the imaging scale of the angle steel in the image. This is because the size of the angle steel in the image varies at different distances and angles. By dynamically adjusting the window size according to the imaging scale, it can be ensured that each sliding window contains sufficient texture information, thereby improving the detection accuracy of appearance anomalies. If the window is too small, it may not be able to cover enough surface texture, resulting in inaccurate identification of appearance anomalies; while if the window is too large, it may affect the detection of local details.

[0043] For each sliding window position, calculate the variance of all pixels within the window across multiple channels in the color space. The choice of color space should consider its sensitivity to apparent changes such as corrosion. Common color spaces include RGB, HSV, and Lab. Among them, the HSV space can better handle variations in hue and saturation, making it suitable for detecting surface anomalies such as corrosion and dirt. By calculating the variance of each channel within each window, the color variation can be quantified. The larger the variance value, the more drastic the color change in that area, potentially indicating significant apparent anomalies (such as corrosion, cracks, oil stains, etc.).

[0044] The calculated variance values ​​from multiple channels are then weighted and fused. The purpose of weighted fusion is to assign weights based on the importance of different channels in the detection of apparent anomalies, resulting in a comprehensive value for the intensity of the apparent anomaly. For example, if a particular color channel (such as the red channel) is more sensitive to corrosion, its variance value can be given a higher weight. This weighting method better reflects the contribution of different channels to the detection of apparent anomalies, thereby improving the accuracy of the detection results.

[0045] The calculated overall apparent anomaly intensity value is assigned to the center pixel of the sliding window. Thus, the center pixel of each sliding window represents the apparent anomaly intensity of that region, reflecting the surface condition at that location. As the sliding window progressively traverses the image, the apparent anomaly intensity value for all pixels is calculated and assigned.

[0046] After the traversal is complete, a grayscale image of the same size as the original image is generated. The grayscale value of each pixel in the image represents the degree of apparent anomaly in its surrounding area. The higher the grayscale value, the more severe the apparent anomaly (such as rust or dirt) in that area. This image is the second feature map, which can intuitively display the surface anomalies of the angle steel of the power transmission tower, facilitating subsequent automated detection and analysis.

[0047] By combining a sliding window approach with color space analysis, apparent anomaly information on the surface of angle steel in power transmission towers can be extracted efficiently and accurately, generating clear anomaly intensity maps. This method improves the accuracy and sensitivity of apparent anomaly detection, adapts to various detection conditions, and has high practical application value.

[0048] In one possible implementation, the acquired infrared thermal images first need to be registered. Registration ensures that thermal images from different viewpoints or time points correspond to the same physical region. Next, temperature value conversion is performed, transforming the raw thermal radiation data in the thermal images into corresponding temperature values ​​to ensure that the thermal images accurately reflect the temperature distribution on the angle steel surface. This step is fundamental to ensuring the accuracy and effectiveness of subsequent analysis.

[0049] After completing the temperature value conversion, the first step is to determine the "normal temperature zone" on the surface of the angle steel. This zone typically refers to large areas with minimal temperature fluctuations, representing the temperature performance of the tower angle steel under normal operating conditions. By clustering these areas, regions with relatively stable temperatures and no abnormal fluctuations can be effectively identified from the overall thermal image. These regions will serve as "reference normal temperatures" for subsequent anomaly analysis.

[0050] For a defined normal temperature range, the statistical average of the temperature values ​​of all pixels is calculated to obtain the "reference normal temperature." This value provides a benchmark for subsequent judgment of abnormal temperature deviations. If the surface temperature of the angle steel deviates significantly from this reference value, it can be determined that a thermal anomaly may exist.

[0051] Define a sliding window and iterate through it on the thermal image. For each position within the sliding window, calculate the average temperature value of all pixels within that window. By using the sliding window approach, local temperature analysis can be performed step-by-step on the entire thermal image, thereby detecting deviations between local temperature changes and the reference normal temperature.

[0052] For each sliding window position, calculate the absolute difference between the average temperature within that window and the reference normal temperature. The key to this step is quantifying the degree of temperature deviation. The larger the difference, the more abnormal the temperature in that area. Typically, temperature anomalies may be caused by equipment overheating, short circuits, or malfunctions.

[0053] The calculated absolute difference is assigned to the center pixel of the sliding window, so that the difference between each center pixel can characterize the degree of temperature anomaly in that region. This difference will serve as the thermal anomaly intensity value for that region, providing data support for subsequent anomaly detection.

[0054] After traversing the entire thermal image, a grayscale image of the same size as the thermal image is finally generated. The grayscale value of each pixel in the image represents the degree of temperature deviation of that area relative to the normal area. This grayscale image is the third feature map, which visually displays the temperature anomalies at various locations on the angle steel surface.

[0055] By utilizing temperature information from infrared thermographic images and sliding window technology, thermal anomaly information on the surface of angle steel can be accurately extracted, and a third feature map of the thermal anomaly intensity can be generated. This method can improve the accuracy of anomaly detection, automate the processing of thermal imaging data, and provide clear indications of anomaly areas, demonstrating significant engineering application value.

[0056] In one possible implementation, three feature maps are first generated using image processing techniques: The first feature map is usually an image generated based on deformation-type defect detection, which may focus on physical damage features such as deformation, cracks or irregular bumps on the surface of angle steel.

[0057] Second feature mapping map: For the detection of rust-related defects, rust traces on the angle steel of the tower are identified by factors such as temperature distribution and surface reflectivity of infrared thermography.

[0058] The third feature map is an image that characterizes the thermal anomalies on the surface of angle steel by measuring the degree of temperature deviation on the thermal image. It is usually used to capture anomalies caused by overheating or local faults.

[0059] Next, different weight coefficients will be assigned to these three feature maps according to different defect types. The allocation of weight coefficients follows this strategy: Deformation defects: These defects typically manifest as deformation, cracks, or surface unevenness in angle steel, thus assigning the highest weight coefficient to the first feature map (deformation detection map). This is because deformation defects are more obvious and directly reflected in the image.

[0060] Rust-related defects: Rust typically manifests as surface color changes and material deposition, appearing in images as different texture variations and surface reflective properties. Therefore, the second feature map (rust detection map) and the third feature map (thermal anomaly detection map) are assigned higher weight coefficients. The combination of these two images can better reflect the surface changes and thermal anomaly state of the rusted area.

[0061] Ambient Temperature Fine-tuning: Since changes in ambient temperature affect the performance of temperature data and thermal images, the weight ratios of the second and third feature maps will be fine-tuned according to different ambient temperatures. In high-temperature environments, temperature deviations may be more significant, so the weight allocation for these feature maps will be slightly higher, and vice versa in low-temperature environments.

[0062] The final weighting coefficients are optimized through machine learning training on a known defect sample library. By analyzing the performance characteristics of different defect types, the machine learning algorithm learns the representation of each defect in each feature map, thereby automatically adjusting the weighting coefficients of each feature map. This step helps improve the accuracy of the weighting coefficients based on a large amount of historical data, enabling the final comprehensive anomaly score map to more accurately reflect the actual defect types present on the surface of the tower angle steel.

[0063] After obtaining each feature map and its corresponding weight coefficient, the pixel grayscale value in each feature map is multiplied by the corresponding weight coefficient, and then linearly superimposed. Through this weighted superposition method, the contribution of different defect types to the result is accurately reflected, and finally a comprehensive anomaly score map (i.e., the fused feature map) is generated.

[0064] Finally, a new image, called the comprehensive anomaly score map, is obtained by weighted linear superposition of the three feature maps. The gray value of each pixel represents the degree of anomaly at that location across all feature maps. The higher the gray value, the greater the likelihood of a defect in that region.

[0065] By assigning weights to different feature maps and performing intelligent fusion, the defect detection of angle steel in power transmission towers becomes more accurate and intelligent, effectively improving detection efficiency and accuracy, adapting to various environmental conditions, and providing detailed defect analysis.

[0066] In one possible implementation, firstly, for the comprehensive anomaly score map, the gray-level histogram distribution of the image is calculated. The gray-level histogram displays the number of pixels at different gray levels in the image, reflecting the distribution of each gray value in the image. In the comprehensive anomaly score map, different gray values ​​represent the degree of anomaly in each region of the image, with regions having higher gray values ​​typically corresponding to defects or anomalies.

[0067] Next, the Otsu's method is used to iterate through all possible grayscale values ​​as candidate thresholds, and the pixels of the image are divided into foreground (defective areas) and background (normal areas). Each candidate threshold attempts to divide the image into foreground and background classes, with the aim of finding an optimal segmentation point that maximizes the difference between the foreground and background.

[0068] For each candidate threshold, calculate the mean grayscale value for the foreground and background classes, and then calculate the proportion of pixels in these two classes in the image. The mean grayscale value is the average of the grayscale values ​​of all pixels in each class, and the foreground-to-background ratio is the proportion of pixels in that class in the image.

[0069] Using the formula for calculating inter-class variance, the inter-class variance value corresponding to each candidate threshold is calculated based on the mean and ratio of the gray levels of the foreground and background. Inter-class variance is a measure of the difference between the foreground and background; specifically, it is a weighted sum of the difference between the mean gray levels of the foreground and background classes and their respective intra-class variances. The formula is as follows: ; in, It is the variance between classes. and These are the pixel weights for the foreground and background, respectively. and It is the average gray level of the foreground and background. It is the average gray level of the entire image.

[0070] By calculating the inter-class variance of all candidate thresholds, the candidate gray value that maximizes the inter-class variance is selected as the final segmentation threshold. This threshold with the maximum inter-class variance most effectively distinguishes the foreground and background in the image, thereby improving the accurate identification rate of defect regions.

[0071] The composite anomaly score map is binarized using a selected final threshold. Binarization converts all pixel values ​​in the image into two classes: background (low grayscale values) and foreground (high grayscale values), which are the regions marked as defect areas. After binarization, defect areas appear as high grayscale regions.

[0072] Connectivity filtering is performed on the binarized image, which involves detecting all connected pixel regions in the image. By analyzing the area of ​​each connected region, regions with areas greater than a preset lower limit are selected. This lower limit is determined based on the image resolution and the minimum physical size requirement for defects. Smaller noise regions are ignored, and only meaningful defect regions are retained.

[0073] Connected regions with an area greater than the lower limit are marked as candidate defect regions, which are considered to potentially contain defects in the angle steel of power transmission towers. Subsequent manual or automated inspections can further confirm whether these candidate regions actually contain defects.

[0074] The application of the Otsu's method in adaptive threshold segmentation can significantly improve the accuracy and efficiency of defect detection in angle steel of power transmission towers through precise segmentation and efficient screening, reduce false positives and false negatives, and ensure the automation and intelligence of the detection process.

[0075] In one possible implementation, three different feature maps (i.e., first, second, and third feature maps) are first generated using image processing or a deep learning model. These maps reflect the response intensity of different features in the image, typically extracted using methods such as convolutional neural networks (CNNs) to extract multi-level features of the image. These feature maps can be different frequencies, textures, colors, or other features related to corrosion defects in the image.

[0076] For each feature map, a dynamic threshold is calculated by statistically analyzing the distribution of all pixels in the image. The dynamic threshold for each feature map is determined based on its global statistical distribution, specifically through the following method: The first dynamic threshold is determined by statistically analyzing all pixel values ​​in the second feature map and selecting a certain percentile value as the threshold. This percentile is typically set to a high percentile (e.g., 90% or 95%) to ensure that the selected threshold can filter out strong response regions and avoid noise interference.

[0077] The second dynamic threshold is determined using a similar method, based on the global pixel value distribution of the third feature map. This threshold ensures a more precise determination of rusted areas.

[0078] The third dynamic threshold is calculated based on the statistical distribution of the first feature map, with the aim of excluding areas with excessively strong response intensity, which are usually unrelated to corrosion defects.

[0079] For each candidate defect region, its response intensity on each feature map is examined and compared with a predefined dynamic threshold. The specific steps are as follows: Check whether the average response intensity of the candidate region on the second feature map exceeds the first dynamic threshold.

[0080] Check whether the average response intensity of the candidate region on the third feature map exceeds the second dynamic threshold.

[0081] Check whether the average response intensity of the candidate region on the first feature map is lower than the third dynamic threshold.

[0082] A candidate area is considered a corrosion defect only when all three conditions are met simultaneously. Specifically: The first and second conditions ensure that the candidate region has a strong response intensity on the second and third feature maps, indicating that it may be an anomalous region.

[0083] The third condition ensures that the candidate region does not have an overly strong response on the first feature map, thus eliminating the possibility of misjudging it as a corrosion defect.

[0084] Once all three conditions are met, the system will output that the area is a rust defect and mark it or perform further analysis.

[0085] The embodiments of the present invention not only improve the detection accuracy of rust defects, but also effectively adapt to different detection environments, ensuring the efficiency and automation level of detection.

[0086] In one possible implementation, three distinct feature maps (first, second, and third feature maps) are first generated using image processing or a deep learning model. These feature maps contain different types of feature information related to deformation defects in the image, typically extracted using deep learning methods such as convolutional neural networks (CNNs). Each map provides a different level of perception for different types of deformation or surface changes.

[0087] For deformation defects, a dynamic threshold is used to determine whether the feature intensity of the candidate region meets the requirements for deformation defects. The fourth dynamic threshold is determined by statistically analyzing the distribution of all pixel values ​​in the first feature map and selecting a value corresponding to a certain percentile (e.g., 90%) as the threshold. This threshold is used to check the response intensity of candidate regions in the first feature map, ensuring that only those potential deformation regions with sufficient intensity are selected.

[0088] For deformation defects, it is necessary not only to determine the response intensity within the region but also to examine whether the response distribution exhibits a continuous and directional change. This step is used to capture the local deformation characteristics caused by external forces or material damage. Calculate the standard deviation of pixel values ​​within the candidate region to assess the dispersion of its response. A larger standard deviation indicates more significant feature variations in the region, which may be related to deformation defects.

[0089] Calculate the gradient direction consistency within the region, that is, determine whether the changes in features within the region have obvious directionality. Usually, deformation is accompanied by specific directional changes, such as stretching or compression.

[0090] Furthermore, it is checked whether the average response intensity of the candidate region on the second and third feature maps is lower than the corresponding fifth and sixth dynamic thresholds. This condition ensures that the determined region has deformation characteristics but no excessively strong abnormal response; otherwise, it may be other types of defects or noise.

[0091] A candidate region is determined to be a deformation defect only when all of the following conditions are met: The region's average response intensity in the first feature map exceeds the fourth dynamic threshold; The response distribution in the region exhibits a continuous and directional gradient change, consistent with the characteristics of deformation defects; The response intensity of the region on the second and third feature maps is lower than that on the fifth and sixth dynamic thresholds.

[0092] Once a candidate region passes the above criteria, the system will mark the region as a deformation defect and may provide further repair recommendations or report output.

[0093] By combining multimodal feature coupling relationships, this method can effectively identify deformation defects, especially when the deformation is small or difficult to identify. Compared with the judgment of a single feature map, combining multiple feature maps and different dynamic threshold conditions can capture deformation features more accurately.

[0094] By combining different feature maps, this method can determine deformation defects from multiple dimensions (such as response intensity, standard deviation, gradient consistency, etc.), making deformation detection more comprehensive and thorough. This multi-faceted assessment effectively avoids misjudgments and improves the robustness and reliability of the detection.

[0095] The embodiments of the present invention not only improve the detection accuracy of deformation defects, but also significantly enhance the automation and efficiency of the entire power tower inspection process, ensuring the safety and stability of power equipment.

[0096] In one possible implementation, corrosion defects are primarily identified using a second feature map, which reflects the relevant characteristics of surface oxidation corrosion. The severity of corrosion can be determined by analyzing the average response intensity of the area on the feature map. Higher response intensity indicates a larger corrosion area or more severe corrosion.

[0097] Deformation defects are primarily identified using a first feature map. This map provides distortion information about the deformed region. By calculating the maximum distortion value within the region, the extent of damage can be assessed, thus determining the severity of the deformation.

[0098] By analyzing known defect cases in historical inspection data and combining this with the experience and knowledge of domain experts, the severity levels of defects were determined. For example, the response intensity of corrosion defects might be divided into several severity levels (such as mild, severe, etc.), while the distortion depth of deformation defects would also be divided into different depth levels. Area ranges were determined based on the pixel area size of the defect region to assess the extent of the defect.

[0099] By statistically analyzing the severity of various defects in historical inspection data, typical response intensity, distortion depth, and area data of corrosion and deformation defects at different severity levels can be obtained. This data provides a basis for determining the boundary values ​​of predefined intervals.

[0100] Once a defective area is identified, the system uses a lookup table to comprehensively determine the defect level based on the predefined range of the area's average response intensity (corrosion defect) or maximum distortion value (deformation defect), as well as the predefined area range to which the area's pixel area belongs. The lookup table method quickly determines the defect level by searching for the corresponding value in the table.

[0101] Ultimately, the system output will include the defect type, defect level, and quantitative information. The defect level reflects the severity of the defect, while the quantitative information provides specific numerical values ​​for the defect (such as the intensity of corrosion, the depth of deformation, etc.). This output not only provides qualitative information (such as the type of defect) but also quantitative assessments (such as the defect level and numerical value), providing a basis for decision-making in subsequent maintenance and treatment.

[0102] By combining historical data, statistical analysis, and expert experience, predefined intervals can accurately reflect the characteristics of defects of varying severity. The system can determine the level of defects based on specific defect characteristics (such as the intensity of corrosion and the depth of deformation), and because a lookup table method is used, the detection results are more standardized and accurate.

[0103] This method can adapt to defects of varying degrees, from minor corrosion to severe deformation, and can accurately determine their severity using corresponding characteristic values ​​and predefined ranges. This flexibility makes the method applicable to a wide range of power transmission tower inspections.

[0104] The following examples will illustrate this in detail: This invention is applied to the real-time defect detection and assessment of power transmission towers, particularly for the automatic identification of surface corrosion and structural deformation. The tower is located in a region with a humid climate, prone to surface corrosion, and an average annual temperature of 10°C. The power transmission tower is approximately 50 meters high and has eight supporting structures; the steel surface is susceptible to corrosion from moisture and harmful gases in the air.

[0105] Specifically, the camera captures high-definition images of the tower at 5-second intervals, while temperature and humidity sensors record the current ambient temperature and humidity in real time. The image data undergoes noise reduction and enhancement processing to improve image quality.

[0106] Perform edge detection on the image (using the Sobel operator).

[0107] The convolutional neural network is used to identify defects and output the type (rust / deformation) and location of the defects.

[0108] Based on the defect area, depth, and morphological characteristics, the defect level is given (e.g., mild, moderate, severe).

[0109] Edge detection of an image is performed using the Sobel operator, as shown in the following formula: ; The edge information of the image is obtained by calculating the gradient value of each pixel.

[0110] The ResNet-50 model was used, with an input image size of 256x256. The image was processed through 4 convolutional layers and 3 fully connected layers for feature extraction and defect classification.

[0111] We collected 5,000 images of the iron tower under different lighting conditions, performed data augmentation (rotation, scaling, brightness adjustment) and used them as the training dataset.

[0112] Model training parameters: Batch size: 32; Learning rate: 0.001; Optimization algorithm: Adam optimizer.

[0113] Formula for calculating the area of ​​corrosion defects: ; in, and They are the first The width and height of each area, It represents the number of pixels in the defective area.

[0114] Formula for calculating the depth of deformation defects: ; in, In coordinates The depth value at the location indicates the maximum depth of deformation.

[0115] Intensity threshold for rust defects: Intensity > 100 (depending on sensor sensitivity and environmental influences).

[0116] Maximum depth threshold for deformation defects: depth > 5mm (set according to actual engineering data).

[0117] To verify the effectiveness of this invention, it was compared with traditional manual inspection methods. Traditional methods rely on manual inspectors using equipment such as telescopes for visual inspection, which is time-consuming and prone to errors. Ten power transmission towers were selected for a three-month defect detection experiment.

[0118] The image acquisition frequency of the detection system of this invention is once every 5 seconds. The manual inspection frequency is once every 3 months.

[0119] Comparison of detection accuracy: This invention achieves an accuracy rate of 95% for rust defects and 93% for deformation defects after processing with a CNN model.

[0120] Manual inspection: The accuracy rate for detecting rust defects is 80%, and the accuracy rate for detecting deformation defects is 75%.

[0121] Comparison of detection efficiency: This invention: Each detection takes 2 minutes.

[0122] Manual inspection: Each inspection takes 1 hour and only one tower can be inspected at a time.

[0123] Comparison of defect missed detection rates: This invention has a 1% failure rate for rust detection and a 2% failure rate for deformation detection.

[0124] Manual inspection: The rate of missed detection for rust is 15%, and the rate of missed detection for deformation is 20%.

[0125] This system significantly improves the real-time performance and efficiency of detection through automated image acquisition and analysis, shortening the detection cycle compared to manual inspection and enabling the identification of defects in multiple towers in a short time. The detection accuracy of this invention is far higher than that of manual inspection, especially under complex environmental conditions (such as significant changes in lighting). The system can operate stably under different seasons and climates without human intervention.

[0126] This invention provides an automatic defect detection system for power transmission towers based on image processing and convolutional neural networks. Experimental comparisons show that the system demonstrates significant advantages over traditional manual detection methods in terms of accuracy, detection efficiency, and false negative rate. This system has high practical value and is suitable for the daily maintenance and monitoring of large-scale power transmission towers.

[0127] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0128] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent detection of angle steel in power transmission towers, characterized in that, Includes the following steps: The drone platform equipped with a visible light camera, infrared thermal imager and lidar is controlled to perform contour flight of the target iron tower, and simultaneously collect visible light images, infrared thermal images and three-dimensional laser point cloud data of the target angle steel. Based on the pre-calibrated transformation relationship and the inherent structural features of angle steel extracted from the data, the visible light image, the infrared thermal image and the three-dimensional laser point cloud are registered with high precision, so that the data of different modes form a pixel-level correspondence in space; The registered multimodal data are processed collaboratively to extract a first feature map characterizing the geometric deformation of the angle steel, a second feature map characterizing the apparent anomaly, and a third feature map characterizing the thermal anomaly. The first feature map, the second feature map, and the third feature map are then fused according to a preset rule to generate a comprehensive anomaly score map. Defect candidate regions are then extracted from the comprehensive anomaly score map through adaptive threshold segmentation. For each of the defect candidate regions, the combination of its response patterns on the first feature map, the second feature map, and the third feature map is analyzed, and the defect type is accurately identified and classified according to a predefined defect discrimination logic based on multimodal feature coupling relationship. The output includes detection results containing defect type, level, and quantitative information.

2. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, The high-precision registration of the visible light image, the infrared thermal image, and the three-dimensional laser point cloud based on the pre-calibrated transformation relationship and the inherent structural features of the angle steel extracted from the data specifically includes: Using the initial transformation matrix obtained through offline calibration between the lidar, the visible light camera, and the infrared thermal imager, preliminary spatial alignment is performed on the three-dimensional lidar point cloud, the visible light image, and the infrared thermal image; From the pre-aligned visible light image, edge pixels of the angle steel are extracted using an edge detection algorithm based on gray-level gradient, and straight line segments representing the edge of the angle steel are fitted from the edge pixels using Hough transform or probabilistic Hough line detection algorithm as two-dimensional image features. From the three-dimensional laser point cloud, point cloud clusters belonging to the same angle steel are separated using a point cloud clustering algorithm based on Euclidean distance. Then, multiple planar models are iteratively fitted from the point cloud clusters using a random sampling consensus algorithm. The intersection lines of different planes are extracted as three-dimensional line segment features, which are used as three-dimensional structural features. The three-dimensional line segment features are projected onto a two-dimensional image plane according to the initial transformation matrix to form a projection line segment set. The projection line segment set is then iteratively matched with the two-dimensional image features extracted from the visible light image using nearest-neighbor optimization. The iterative nearest-neighbor optimization matching process uses the endpoints of the projection line segments as the initial point set and the nearest point on the line segment extracted from the image as the target point set. A fine rotation and translation transformation matrix is ​​iteratively optimized by minimizing the average Euclidean distance between the initial point set and the target point set. The optimization terminates when the change in the average Euclidean distance is less than a convergence threshold pre-derived based on the point cloud accuracy and image resolution. The final refined rotation and translation transformation matrix is ​​applied to the infrared thermal image and the visible light image to complete pixel-level high-precision registration with the three-dimensional laser point cloud.

3. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, The extraction of the first feature map characterizing the geometric deformation of the angle steel specifically includes: For the registered 3D laser point cloud, take each laser point as the core and take a set of points within a certain neighborhood around it; the size of the neighborhood is adaptively determined according to the physical size of the angle steel and the point cloud density to ensure that a local area of ​​the angle steel can be covered. For each set of points extracted, an ideal plane model is fitted using the least squares method; Calculate the perpendicular distance from each point in the point set to its corresponding ideal plane model; Traverse all points in the three-dimensional laser point cloud, and map the calculated distance value of each point back to its corresponding pixel position to form a two-dimensional grayscale image, where the grayscale value of each pixel represents the distortion of the spatial point in the normal direction. This grayscale image is the first feature mapping map.

4. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, The extraction of the second feature map representing the apparent anomaly specifically includes: On the registered visible light image, a sliding window is defined; the size of the sliding window is adjusted according to the imaging scale of the angle steel in the image to ensure that the window can contain sufficient texture information. For each position traversed by the sliding window on the image, the variance of all pixels within the window across multiple channels of the color space is calculated; the color space is chosen based on its sensitivity to apparent changes. The variance values ​​of multiple channels are weighted and fused to obtain a comprehensive apparent anomaly intensity value. Assign the combined apparent anomaly intensity value to the center pixel of the sliding window; After traversing the entire image, a grayscale image of the same size as the original image is generated, where the grayscale value of each pixel represents the degree of appearance abnormality of its surrounding area. This grayscale image is the second feature map.

5. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, The extraction of the third feature map characterizing the thermal anomaly specifically includes: On the infrared thermal image that has been registered and temperature value conversion completed, the normal temperature area of ​​the angle steel surface is first determined; the normal temperature area is obtained by clustering and identifying large, continuous areas with small temperature fluctuations in the image. Calculate the statistical average of the temperature values ​​of all pixels within the normal temperature area, and use it as the reference normal temperature; Define a sliding window, and for each position that the sliding window traverses on the thermal image, calculate the average value of the temperature values ​​of all pixels within the window; Calculate the absolute difference between the average temperature of the window and the reference normal temperature; The absolute difference is assigned to the center pixel of the sliding window; After traversing the entire thermal image, a grayscale image of the same size as the thermal image is generated, where the grayscale value of each pixel represents the degree of temperature deviation of its surrounding area relative to the normal area. This grayscale image is the third feature map.

6. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, The first feature map, the second feature map, and the third feature map are fused according to a preset rule, wherein the preset rule is: A weight coefficient is assigned to each of the first, second, and third feature maps. The weight coefficient assignment strategy is based on the dominant features exhibited by different defect types: for deformation defects, the first feature map is assigned the highest weight coefficient; for corrosion defects, the second and third feature maps are assigned relatively high weight coefficients, and the weight ratio between them is fine-tuned according to the ambient temperature. The specific values ​​of the weight coefficients are obtained by optimizing the machine learning training on a known defect sample library. The pixel grayscale values ​​of the three feature maps are multiplied by their respective weight coefficients and then linearly superimposed to obtain the comprehensive anomaly score map.

7. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, The step of extracting defect candidate regions from the comprehensive anomaly score map through adaptive threshold segmentation specifically employs the maximum inter-class variance method to adaptively determine the segmentation threshold. Calculate the gray-level histogram distribution of the comprehensive anomaly score map; By iterating through all possible grayscale values ​​as candidate thresholds, image pixels are divided into foreground and background categories. Calculate the mean gray value of foreground and background pixels, and their respective proportions in the image; Based on the formula for calculating the inter-class variance of foreground and background, calculate the inter-class variance value corresponding to each candidate threshold; the inter-class variance value is a function of foreground probability, background probability, and the difference between foreground mean and background mean; Select the candidate gray value that maximizes the inter-class variance as the final segmentation threshold; The final segmentation threshold is used to binarize the comprehensive anomaly score map, and connected regions are filtered. Regions with an area greater than a preset lower limit are marked as candidate defect regions. The preset lower limit is calculated based on the image resolution and the minimum physical size requirement of the defect.

8. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, In the predefined defect discrimination logic based on multimodal feature coupling, the judgment logic for corrosion defects is as follows: First, check whether the average response intensity of the defect candidate region on the second feature map exceeds the first dynamic threshold; the first dynamic threshold is obtained by statistically analyzing the distribution of all pixel values ​​in the entire second feature map and taking the value corresponding to a certain percentile as the threshold. Simultaneously, it checks whether the average response intensity of the defect candidate region on the third feature map exceeds the second dynamic threshold; the method for determining the second dynamic threshold is the same as that for the first dynamic threshold, but based on the global statistical distribution of the third feature map; It is also necessary to check whether the average response intensity of the defect candidate region on the first feature map is lower than the third dynamic threshold; the method for determining the third dynamic threshold is similar to the above threshold. The candidate area is determined to be a corrosion defect only if all three conditions are met simultaneously.

9. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, In the predefined defect discrimination logic based on multimodal feature coupling, the judgment logic for deformation defects is as follows: First, check whether the average response intensity of the defect candidate region on the first feature map exceeds the fourth dynamic threshold. The fourth dynamic threshold is obtained by statistically analyzing the distribution of all pixel values ​​in the entire first feature map and taking the value corresponding to a certain percentile as the threshold. The check is performed to determine whether the response distribution of the defect candidate region on the first feature map exhibits a continuous and directional gradient change. This check is accomplished by calculating the standard deviation of pixel values ​​within the region and the consistency of gradient direction. Simultaneously check whether the average response intensity of the defect candidate region on the second feature map and the third feature map is lower than the fifth dynamic threshold and the sixth dynamic threshold; The candidate region is determined to be a deformation defect only if all of the above conditions are met simultaneously.

10. The intelligent detection method for angle steel of power transmission towers according to claim 1, characterized in that, The output includes detection results with defect type, level, and quantification information. The defect level is classified based on the following criteria: For corrosion defects, their level is determined by a lookup table method based on the predefined intensity range of the average response intensity value of the defect candidate region on the second feature map and the predefined area range of the pixel area of ​​the region. For deformation defects, their level is determined by a lookup table method based on the predefined depth range where the maximum distortion value of the defect candidate region is located in the first feature map and the predefined area range to which the pixel area of ​​the region belongs. The boundary values ​​of the predefined intensity range, depth range, and area range are determined by statistical analysis of the severity of known defects in historical detection data and by combining the experience and knowledge of domain experts.

Images