Lightweight detection method for building surface defect identification

Abstract

The invention provides a lightweight detection method for building surface defect identification, and belongs to the technical field of computer vision and intelligent detection. The method comprises the following steps: carrying out standardized preprocessing on three-dimensional point cloud data collected by a terahertz radar, carrying out local scanning by adopting a large-size sliding window, and constructing a local feature vector; a normal material template set is constructed through clustering analysis, a small-size sliding window is adopted for fine-grained scanning, and in the defect recognition process, the defect condition of the previous small-size sliding window is preferentially compared with the state of the previous small-size sliding window, so that rapid recognition of a defect mode is realized. According to the method, the two-stage process of'rough modeling-fine detection 'is adopted, so that the distinguishing capability of the system on local small defects and large-area anomalies is improved; and through dynamic window loading, local cache management and adjacency traversal strategies, memory and computing resource requirements are remarkably reduced, and the method is suitable for edge computing or embedded deployment scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and intelligent detection, and specifically belongs to a method for detecting building surface defects based on 3D point clouds. It is a rapid detection method for building surface defects that combines adaptive extraction of point cloud features and a lightweight recognition strategy, and particularly relates to a lightweight detection method for identifying building surface defects, which is applicable to scenarios such as intelligent building inspection and digital construction management. Background Art

[0002] The detection of building exterior wall defects is a key link in the quality management of the entire life cycle of buildings, covering multiple application scenarios such as quality control during the construction stage, acceptance assessment during the completion stage, periodic structural safety detection during the operation period, and subsequent maintenance and repair. Building an efficient, accurate, and intelligent exterior wall defect detection system is not only an important requirement for ensuring project quality but also an important technical support for urban building safety governance.

[0003] In traditional methods, the detection of building exterior wall defects mostly uses image-based detection methods. Such methods collect high-definition images of the wall surface by arranging a camera array and combine image processing algorithms (such as edge detection, texture analysis, deep learning, etc.) for defect identification. Although this method has a mature implementation path and wide application, it has significant limitations: on the one hand, images are susceptible to external factors such as perspective occlusion, lighting changes, and shadow interference, which easily lead to detection errors and missed detections; on the other hand, image information lacks the ability to express three-dimensional spatial information such as the depth and concave-convex structure of the wall, making it difficult to meet the recognition requirements for defects related to complex geometric features or deep structures.

[0004] To solve the above problems, 3D scanning technology has been introduced into the building defect detection process in recent years. Among them, terahertz radar, as a new type of high-frequency imaging technology, shows great potential in the scanning and recognition of building surface structures due to its strong penetration ability, high resolution, and non-contact acquisition. Compared with traditional lidar or millimeter-wave radar, the terahertz system can obtain high-density and high-precision 3D point cloud information without relying on lighting conditions, providing a data basis for the three-dimensional modeling and quantitative evaluation of complex surface defects.

[0005] Despite the new breakthroughs brought by point cloud technology, existing systems still face several challenges in practical applications. First, the point cloud data itself is characterized by high-dimensionality and density, with a large data scale, resulting in high computational overhead and long response times in subsequent processing, making it difficult to meet real-time requirements. Second, there are a wide variety of wall materials and complex surface textures, and there is still a lack of effective strategies for adaptively extracting local features sensitive to defects from point clouds. In addition, most current point cloud recognition algorithms focus on high-precision modeling and geometric reconstruction, and there are still gaps in the lightweight and highly robust detection and processing of defective targets. Therefore, there is an urgent need to design a point cloud defect detection method with efficient data processing capabilities, adaptive local feature mining capabilities, and good generalization performance to achieve efficient identification and practical deployment of building exterior wall defects. Summary of the Invention

[0006] To solve the problems in the prior art such as high consumption of computing resources, low recognition accuracy, and inability to adapt to walls of different materials in the process of identifying three-dimensional point cloud defects on the building surface. The present invention provides a lightweight detection method for identifying building surface defects. In the process of dynamic sampling, with the core strategy of "dominated by normal regions and extracting abnormal differences", a phased window scanning and adaptive feature learning mechanism are adopted to achieve efficient identification of defective regions under different wall material conditions, with high efficiency, scalability, and transfer robustness. This method can dynamically update the normal wall surface feature representation, identify low-contrast or large-area abnormal regions, and at the same time effectively control memory consumption through modular design, improving the deployability and practicality of the system.

[0007] The specific technical solution of the present invention is as follows:

[0008] A lightweight detection method for identifying building surface defects, the method comprising the following steps:

[0009] Step S11: Perform standardized preprocessing on the three-dimensional point cloud data collected by the terahertz radar to obtain standardized point clouds;

[0010] Step S12: Use a large-size sliding window to perform local scanning on the standardized point clouds, extract three local structural features of roughness, normal vector change rate, and point density, and construct a local feature vector;

[0011] Step S13: Construct a normal material template set through clustering analysis, and perform spatial connectivity analysis during the construction of the normal material template set;

[0012] Step S21: Use a small-size sliding window to perform fine-grained scanning on the standardized point clouds, plan the sliding path of the small-size window according to the spatial adjacency rule, so that any small-size sliding window has spatial overlap or shared boundary points with the previous small-size sliding window;

[0013] Step S22: Perform full normal material template set matching on the local feature vectors of the first small-sized sliding window to determine defects;

[0014] In the process of defect recognition for subsequent small-sized sliding windows, different processes are executed according to the attribution label of the previous small-sized sliding window. If the attribution label of the previous small-sized sliding window is the defect mode, the current small-sized sliding window is preferentially compared with the defect mode of the previous small-sized sliding window to achieve rapid identification of the defect mode; if the attribution label of the previous small-sized sliding window is the normal template, the current small-sized sliding window is preferentially compared with the normal template of the previous small-sized sliding window to quickly determine whether there are no defects.

[0015] Furthermore, step S11 includes the following steps:

[0016] The three-dimensional point cloud data collected by the terahertz radar is expressed as ; where represents the th three-dimensional sampling point on the building surface, represents the set of real numbers, represents the number of three-dimensional sampling points; the calculation process of the standardized preprocessing is as follows:

[0017]

[0018]

[0019]

[0020] where represents the th standardized point after preprocessing, is the geometric center of the point cloud, is the standard deviation of the spatial distribution of the point cloud, represents the square of the Euclidean distance.

[0021] Furthermore, step S12 includes the following steps:

[0022] The th sliding window and its corresponding sampling point set are:

[0023]

[0024] where represents the center of the sliding window, represents the side length of the sliding window, represents the cube region centered on the sliding window center with a side length of . The th sliding window contains points, which is the number of slides required for a complete scan of the building surface;

[0025] The roughness of the th sliding window is calculated as follows:

[0026]

[0027] where represents transpose, represents the normal vector of the local least squares fitting plane of the th sliding window, is the plane offset term of the th sliding window;

[0028] The rate of change of the normal vector of the th sliding window is calculated as follows:

[0029]

[0030] where represents the principal direction normal vector of the neighborhood to which the th standardized point belongs;

[0031] The point density of the th sliding window is calculated as follows:

[0032]

[0033]

[0034] where represents the volume of the convex hull of the point set in the th sliding window, represents the convex hull formed by the point set in the th sliding window, represents the volume function;

[0035] Integrate the three local structural features of roughness, rate of change of normal vector, and point density into a local feature vector. The local feature vector of the th sliding window

[0036] f j =[ r j , ∆ n j , ρ j ] T 。

[0037] Further, the normal material template set in step S13 is constructed as follows:

[0038] Form a spatial structure feature set by all the local feature vectors of the sliding windows , input the spatial structure feature set into the clustering module to perform unsupervised learning, extract the main features, automatically cluster the regions that meet the high consistency and structural continuity, and construct the normal material template set; the normal material template set is represented as follows:

[0039]

[0040]

[0041] Wherein, represents the normal material template set; represents the total number of feature clustering clusters; represents the th normal template; is the th feature clustering cluster, represents the size of the feature value of the th feature clustering cluster.

[0042] Further, the defect judgment of the full normal material template set in step S22 includes the following steps:

[0043] Step S221: Perform fine-grained scanning on the standardized point cloud of the current small-size sliding window, and extract the roughness , the normal vector change rate and the point density of the three local structural features, and construct the local feature vector of the current small-size sliding window;

[0044] Step S222: Calculate the defect response value of the current small-size sliding window. If the defect response value is greater than or equal to the defect threshold, the current small-size sliding window is a normal area, and record the attribution label of the current small-size sliding window; if the defect response value of the current small-size sliding window is less than the defect threshold, mark the current small-size sliding window as a suspected defect window;

[0045] Step S223: Perform local boundary reconstruction and multi-scale verification on the suspected defect window;

[0046] Step S224: Calculate the structural confidence at multiple scales of the radius, and output the defect detection result.

[0047] Further, in step S222, the current defect response value of the small-size sliding window is calculated as follows:

[0048]

[0049]

[0050]

[0051]

[0052] where represents the current defect response value of the small-size sliding window, represents the local feature vector of the current small-size sliding window and the minimum Euclidean distance from the normal material template set; is a tiny positive number to prevent division by zero; represents the Euclidean distance; represents the th local feature vector of the sliding window and the minimum Euclidean distance from the normal material template set, represents the maximum value of the minimum Euclidean distances between all local feature vectors of the sliding windows and the normal material template set; represents taking the maximum value for all .

[0053] Further, in step S23, the comparison between the current small-size sliding window and the defect pattern of the previous small-size sliding window is calculated as follows:

[0054] The previous small-size sliding window has an attribution label of the defect pattern . The local feature vector of the current small-size sliding window is preferentially compared with the defect pattern of the previous small-size sliding window to calculate the similarity, and the defect response value of the defect pattern of the current small-size sliding window is obtained. The calculation method is as follows:

[0055]

[0056]

[0057] where represents the defect response value of the defect pattern of the current small-size sliding window; represents the th local feature vector of the sliding window and the minimum Euclidean distance from the defect characterization library; Represents the maximum value of the minimum Euclidean distance between all sliding window local feature vectors and the defect characterization library.

[0058] Furthermore, the current small-size sliding window is preferentially compared with the normal template of the previous small-size sliding window and calculated in the following manner:

[0059] The previous small-size sliding window The attribution label is the normal template , the local feature vector of the current small-size sliding window is preferentially compared with the normal template of the previous small-size sliding window to perform a fast similarity comparison, and the local defect response value of the current small-size sliding window is obtained , and the calculation method is as follows:

[0060]

[0061] where represents the local defect response value of the current small-size sliding window.

[0062] Compared with the prior art, the present invention has the following beneficial technical effects:

[0063] 1) The present invention dynamically models to enhance the generalization ability: The system automatically learns and updates the normal material template set of the wall through the large-size sliding window scanning process, without large-scale manual annotation, and has good cross-building and cross-material migration capabilities.

[0064] 2) The present invention scans in stages to improve the recognition accuracy: Adopts a two-stage process of "coarse modeling - fine detection" to improve the system's ability to distinguish local small defects and large-area anomalies, and avoid misjudgment and missed judgment phenomena.

[0065] 3) The lightweight processing framework of the present invention supports efficient deployment: Through dynamic window loading, local cache management and adjacent traversal strategies, it significantly reduces the memory and computing resource requirements, and is suitable for edge computing or embedded deployment scenarios.

[0066] 4) The defect response of the present invention has interpretability and reliability: Based on template matching and feature comparison for defect judgment, the results are interpretable, which is convenient for manual review and maintenance operations; the abnormal feedback mechanism ensures the robust operation of the system.

[0067] 5) Through the intelligent and lightweight building surface defect detection technology, the present invention provides efficient and accurate technical support for the quality management of the entire life cycle of buildings. On the one hand, through automated detection, the efficiency and accuracy of building exterior wall defect identification are greatly improved, effectively reducing the safety risks and cost inputs of manual inspections, providing a reliable guarantee for urban building safety governance, and helping to prevent safety accidents caused by structural defects. On the other hand, it promotes the transformation of building detection from the traditional manual mode to the intelligent and digital mode, conforming to the development trends of "new infrastructure" and building industrialization. Its generalization ability across materials and scenarios and lightweight deployment characteristics can be widely applied to scenarios such as intelligent inspection and construction management, improving the overall technical level of the industry, and having important demonstration significance for promoting the safety upgrade and digital construction of urban infrastructure. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments of the present invention. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0069] Figure 1 It is a general overview schematic diagram of the lightweight detection process for building surface defect identification of the present invention.

[0070] Figure 2 It is a schematic diagram of the acquisition of the normal material template set of the present invention.

[0071] Figure 3 It is a schematic diagram of the lightweight detection for building surface defect identification of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the technical ideas proposed by the present invention, any extension, improvement or replacement made by those of ordinary skill in the relevant fields without creative efforts shall fall within the protection scope of the present invention.

[0073] The present invention proposes a lightweight detection method for building surface defect identification. The overall solution is based on the dynamic sliding window strategy and sequentially executes three steps: wall feature modeling, defect fine identification, and data loading optimization in stages, which can greatly reduce the consumption of memory and computing resources while ensuring the detection accuracy, and has good deployability and generalization ability.

[0074] This method is aimed at the original high-density point cloud input and relies on the sliding window mechanism. First, it conducts dynamic scanning and feature clustering on the local area with a larger window, extracts and constructs a representative template set of normal wall materials; then, it conducts a second-stage fine scanning with a smaller window, and compares the current area with the learned normal material template set to identify structural anomalies and surface defects; finally, it cooperates with the block loading and memory scheduling mechanism of the point cloud area throughout the process to ensure that the entire recognition system can operate stably on resource-constrained platforms.

[0075] The present invention proposes a lightweight detection method for identifying building surface defects, as Figures 1-3 shown, the method includes the following steps:

[0076] Step S1: Perform standardized preprocessing on the three-dimensional point cloud data collected by the terahertz radar to obtain a standardized point cloud; use a large-size sliding window to perform local scanning on the standardized point cloud, extract features to construct local feature vectors; construct a normal material template set through clustering analysis, and perform spatial connectivity analysis during the process of constructing the normal material template set.

[0077] Based on the assumption that "most building walls are of normal structure", the present invention first preprocesses the three-dimensional point cloud data collected by the terahertz radar, including standardized operations such as geometric center translation and scale normalization. Subsequently, a large-size sliding window is used to perform dynamic scanning on the local area within the global point cloud. By extracting material-related features such as the roughness, normal vector change rate, and point density of the point cloud within the window, the surface properties and geometric stability of the area are judged. Automatic clustering analysis is performed on the areas that show high consistency and structural continuity during the scanning process, and a set of normal material template sets that can represent different wall material types are gradually established. The normal material template set is used to reflect the typical spatial distribution characteristics of the wall under different material conditions, as a judgment benchmark for the subsequent defect recognition stage. At the same time, to avoid large-area defects being misidentified as normal walls, the present invention also designs a spatial connectivity analysis mechanism: when it is detected that the features of multiple consecutive sliding windows deviate significantly from the current template and form a connected area with an area exceeding the set threshold in space, the system will automatically mark this area as a potential misjudgment area and trigger a manual feedback request to ensure the accuracy and purity of the template set.

[0078] Specifically, step S1 includes the following steps:

[0079] Step S11: Perform standardized preprocessing on the three-dimensional point cloud data collected by the terahertz radar to obtain a standardized point cloud.

[0080] The standardized preprocessing ensures the comparability consistency of the point cloud data among different building detection projects, is a basic operation for regional analysis and structural description, and improves the stability and migration ability of the system.

[0081] The three-dimensional point cloud data collected by the terahertz radar is represented as ; where represents the th three-dimensional sampling point on the building surface, represents the set of real numbers, represents the number of three-dimensional sampling points. To eliminate the influence of building scale differences, scanning area position offsets, and density changes on subsequent feature analysis, the collected point cloud data needs to be preprocessed by standardization. The standardization preprocessing includes operations such as geometric center translation and scale normalization. The preprocessed standardized point cloud is denoted as , and its standardization preprocessing calculation process is as follows:

[0082]

[0083]

[0084]

[0085] where represents the th standardized point after preprocessing, is the geometric center of the point cloud, is the standard deviation of the spatial distribution of the point cloud, represents the square of the Euclidean distance. After the standardization preprocessing, the point cloud is mapped to a unified scale and position framework, so that the subsequent sliding window sampling does not depend on the size of the building object, effectively improving the generality and structural sensitivity of the system.

[0086] Step S12: Use a large-size sliding window to perform local scanning on the standardized point cloud, extract three local structural features of roughness, normal vector change rate, and point density, and construct a local feature vector.

[0087] Use a large-size sliding window to perform local scanning on the standardized point cloud to extract the statistical features of the building wall surface in terms of roughness, normal vector change rate, and point density, and integrate these features into a local feature vector describing the regional properties.

[0088] The th sliding window, and its corresponding sampling point set is:

[0089]

[0090] where represents the center of the sliding window, represents the side length of the sliding window, represents the center of the sliding window as the center and the side length is The cubic region. The sliding window contains points, which is the number of sliding times required for a complete scan of the building surface. Calculate three local structure features of roughness, normal vector change rate, and point density for the window sampling point set.

[0091] Roughness is used to measure the residual variance between the local point cloud and the fitted plane, representing the degree of surface undulation and measuring whether the surface is smooth. The roughness of the th sliding window is calculated as follows:

[0092]

[0093] where represents transpose, represents the normal vector of the local least squares fitting plane of the th sliding window, is the plane offset term of the th sliding window.

[0094] The normal vector change rate is used to measure the consistency of the point cloud normal vector direction in the local area, used to identify corner or subtle distortion areas, and measure the surface consistency. The normal vector change rate of the th sliding window is calculated as follows:

[0095]

[0096] where represents the main direction normal vector of the neighborhood to which the th standardized point belongs.

[0097] The point density reflects the spatial distribution density of points in the area, representing the material particles or loss state. The point density of the th sliding window is calculated as follows:

[0098]

[0099]

[0100] where represents the convex hull volume of the point set in the th sliding window, represents the convex hull formed by the point set in the th sliding window, represents the volume function.

[0101] Integrate three local structural features, namely roughness, normal vector change rate, and point density, into a local feature vector. The local feature vector of the th sliding window is

[0102] f j =[ r j , ∆ n j , ρ j ] T

[0103] The local feature vector comprehensively reflects the wall material characteristics and is the key basis for constructing a normal material template set.

[0104] Step S13: Construct a normal material template set through clustering analysis, and perform spatial connectivity analysis during the construction of the normal material template set.

[0105] Step S131: Construct a normal material template set through clustering analysis.

[0106] Form a spatial structure feature set from all the local feature vectors of the sliding windows , input the spatial structure feature set into the clustering module to perform unsupervised learning, extract the main features, automatically cluster the regions that meet the high consistency and structural continuity, and construct a normal material template set. The clustering method can be -means or density-aware clustering (such as DBSCAN) to cope with the scene complexity. The normal material template set is represented as follows:

[0107]

[0108]

[0109] where represents the normal material template set; the normal material template set is a template set composed of the center vectors of the feature clustering clusters, and serves as a structural comparison reference library in the subsequent defect recognition stage; represents the total number of feature clustering clusters; represents the th normal template, which is the center vector of the feature clustering cluster and represents a typical wall characterization structure; is the th feature clustering cluster, represents the th feature clustering cluster eigenvalue size.

[0110] Step S132: Perform spatial connectivity analysis, spatial anomaly detection, and connectivity elimination during the construction of the normal material template set.

[0111] To avoid template contamination caused by scanning over abnormal regions, spatial connectivity analysis is introduced during the construction of the normal material template set. That is, when it is detected that the features of multiple consecutive sliding windows deviate significantly from the current template and form a connected region with an area exceeding the set threshold in space, the system will automatically mark this region as a potential misjudgment area and trigger a manual feedback request to ensure the accuracy and purity of the template set.

[0112] The spatial connectivity analysis is implemented through the "structural deviation judgment + connectivity detection" mechanism, and the "structural deviation judgment + connectivity detection" mechanism is as follows:

[0113] Calculate the minimum Euclidean distance between the local feature vector of each sliding window and the normal material template set. Taking the th sliding window as an example, introduce the calculation of the minimum Euclidean distance between the local feature vector of the th sliding window and the normal material template set , and the calculation method is as follows:

[0114]

[0115] Among them, represents the Euclidean distance. The determination threshold is denoted as , generally 1e-3; if the minimum Euclidean distance between the local feature vector of the sliding window and the normal material template set is greater than the determination threshold, that is, , then the sliding window is regarded as significantly deviating from the normal template and marked as a "suspected abnormal" region.

[0116] Calculate the maximum value of the minimum Euclidean distances between the local feature vectors of all sliding windows and the normal material template set, and the calculation method is as follows:

[0117]

[0118] Among them, represents the maximum value of the minimum Euclidean distances between the local feature vectors of all sliding windows and the normal material template set; represents taking the maximum value for all .

[0119] Form an abnormal window set with all sliding windows that satisfy , and construct a three-dimensional connected graph based on the spatial adjacency relationship, which is expressed as follows:

[0120]

[0121] Among them, represents the three-dimensional connected graph; It means to find the connected components.

[0122] For each connected subset, calculate its spatial coverage area according to the adjacent relationship of the windows, and judge the relationship between the spatial coverage area and the spatial coverage area threshold. If the spatial coverage area is greater than the spatial coverage area threshold, the connected subset is determined as the "high-risk area of abnormal template pollution", and the area features are removed from the template candidate set, and the manual verification mechanism is triggered to ensure the purity and representativeness of the template.

[0123] Taking the th connected subset as an example, the calculation of the spatial coverage area is introduced. The th connected subset contains sliding windows. The spatial coverage area of the th connected subset is the sum of the areas of sliding windows, which is expressed as . represents finding the spatial coverage area. The spatial coverage area threshold is expressed as . If , then it is determined that the th connected subset is the "high-risk area of abnormal template pollution".

[0124] The "structural deviation judgment + connectivity detection" mechanism effectively prevents large damaged areas from being learned as "normal material templates". By determining the threshold to control the feature deviation degree of a single sliding window, and the spatial coverage area threshold to control the influence range of regional anomalies, the safety and practicality of constructing the normal material template set are improved.

[0125] Step S2: Use small-sized sliding windows to perform fine-grained scanning on the standardized point cloud, plan the sliding path of the small-sized windows according to the spatial adjacency rules, so that any small-sized sliding window has spatial overlap or shared boundary points with the previous small-sized sliding window; perform full normal material template set matching on the local feature vector of the first small-sized sliding window to judge defects; during the defect recognition process of subsequent small-sized sliding windows, different processes are executed according to the attribution label of the previous small-sized sliding window. If the attribution label of the previous small-sized sliding window is the defect mode, the current small-sized sliding window is preferentially compared with the defect mode of the previous small-sized sliding window to achieve fast recognition of the defect mode; if the attribution label of the previous small-sized sliding window is the normal template, the current small-sized sliding window is preferentially compared with the normal template of the previous small-sized sliding window to quickly judge whether there are no defects.

[0126] After completing the normal material template set After the construction, it enters the defect identification stage. In this stage, a smaller-sized sliding window is used to perform fine-grained scanning on the standardized point cloud to extract the structural features of each local area with high resolution, and compare the structure with the normal material template set to perform traversal and discrimination with higher spatial resolution.

[0127] After completing the construction of the normal material template set in step S1, the point cloud is traversed in a refined manner with a smaller-sized sliding window again. For each local area, the system extracts three local structural features: roughness, normal vector change rate, and point density, and compares the features with the normal material template set constructed in step S1. The defect response value is obtained by calculating the similarity. When the defect response value is lower than the defect threshold, the area is determined as a suspected defect area, and strategies such as local boundary reconstruction, multi-scale comparison, and structural confidence estimation are further used to refine the identification and improve the accuracy and robustness of the detection. This stage particularly strengthens the detection ability for defect areas with blurred boundaries or subtle structural changes such as micro-cracks, peeling, and swelling.

[0128] To ensure that the system can cope with the computational resource pressure brought by large-scale building point clouds, the present invention designs an efficient data loading and resource scheduling mechanism. The overall point cloud space is divided according to spatial voxels, and only the point cloud blocks within the area covered by the current sliding window are loaded on demand. A traversal strategy based on spatial adjacency relationships is adopted on the sliding path to ensure the data continuity and processing consistency of adjacent windows. At the same time, the system maintains a fixed number of active windows, and combines cache invalidation and memory recycling strategies to dynamically release the data in invalid areas, effectively avoiding the memory overflow problem caused by full-map loading, and significantly improving the scalability and on-site deployment ability of the detection system.

[0129] To further improve the execution efficiency of the overall system, avoid repeated full-template comparison in wall areas with strong structural continuity, and the risk of resource overflow in large-scale point cloud scenarios, the present invention introduces a set of sliding scheduling and knowledge reference mechanisms based on spatial neighborhood continuity and context reuse to implement a lightweight and efficient dynamic point cloud recognition strategy. The core idea is "recognizing the adjacent by the adjacent", that is, using the material attribution information of adjacent windows in space to guide the current window to make quick decisions, thereby significantly reducing unnecessary computational overhead.

[0130] Specifically, step S2 includes the following steps:

[0131] Step S21: Use a small-sized sliding window to perform fine-grained scanning on the standardized point cloud, and plan the sliding path of the small-sized window according to the spatial adjacency rule, so that any small-sized sliding window has spatial overlap or shared boundary points with the previous small-sized sliding window.

[0132] Specifically, the small-size sliding window path is continuous, enabling the small-size sliding window to slide continuously preferably in one direction (such as the X-axis) to ensure that the current small-size sliding window overlaps spatially or at least shares boundary points with its previous small-size sliding window. There is spatial overlap or at least sharing of boundary points.

[0133] Step S22: Perform full normal material template set matching on the local feature vector of the first small-size sliding window to judge defects.

[0134] Full normal material template set matching to judge defects includes the following steps:

[0135] Step S221: Conduct fine-grained scanning on the standardized point cloud of the current small-size sliding window, extract three local structural features of roughness, normal vector change rate, and point density, and construct the local feature vector of the current small-size sliding window.

[0136] Specifically, the current small-size sliding window , and its coverage area is:

[0137]

[0138] where represents the standardized point under the current small-size sliding window, , represents the side length of the small-size sliding window, is the center of the small-size sliding window; represents a cube region centered on the center of the small-size sliding window with a side length of . The current small-size sliding window contains points.

[0139] Extract three local structural features of roughness, normal vector change rate, and point density of the small-size sliding window.

[0140] The roughness of the current small-size sliding window is calculated in the following way:

[0141]

[0142] where represents the normal vector of the local least squares fitting plane of the current small-size sliding window, represents the plane offset term of the current small-size sliding window.

[0143] The normal vector change rate of the current small-size sliding window is calculated in the following way:

[0144]

[0145] Among them, is the principal direction normal vector of the neighborhood to which the

[0146] current small - size sliding window point density is calculated in the following way:

[0147]

[0148]

[0149] Among them, represents the convex hull volume of the point set in the current small - size sliding window

[0150] Integrate the three local structure features of the extracted small - size sliding window roughness, normal vector change rate, and point density into the local feature vector of the small - size sliding window. The local feature vector of the current small - size sliding window is expressed as follows:

[0151] f i ' =[ r i ' , ∆n i ' , ρ i ' ] T

[0152] In this step, the local feature vector of each small - size sliding window is extracted at a higher spatial resolution , and its definition method is consistent with that in step S1 to ensure structural consistency and comparability. This local feature vector serves as the basis for subsequent comparison with the normal material template set.

[0153] Step S222: Calculate the defect response value of the current small - size sliding window. If the defect response value is greater than or equal to the defect threshold, the current small - size sliding window is a normal area, and record the attribution label of the current small - size sliding window; if the defect response value of the current small - size sliding window is less than the defect threshold, mark the current small - size sliding window as a suspected defect window.

[0154] Compare the local feature vector extracted from the current small - size sliding window with the template features in the normal material template set, measure its structural deviation degree, and output the defect response value.

[0155] Calculate the minimum Euclidean distance between the local feature vector of the current small - size sliding window and the normal material template set as the structural proximity. The local feature vector of the current small - size sliding window The minimum Euclidean distance from the normal material template set is calculated as follows:

[0156]

[0157] Then, perform similarity calculation to obtain the defect response value of the current small - size sliding window , and normalize the value range to [0,1] . The similarity calculation formula is as follows:

[0158]

[0159] where represents the maximum value of the minimum Euclidean distances between all sliding window local feature vectors and the normal material template set, is a small positive number to prevent division by zero, generally taking the value of 1e - 6. The smaller the defect response value of the current small - size sliding window, the greater the deviation of the current area from any normal material template, that is, the more likely it is a defect area.

[0160] Set the defect threshold , and perform material characterization similarity threshold screening, that is, if the defect response value of the current small - size sliding window is greater than or equal to the defect threshold , that is , then the current small - size sliding window is a normal area, and the normal template with the minimum Euclidean distance between the local feature vector of the current small - size sliding window and the normal material template set is recorded as the attribution label of the current small - size sliding window . If the defect response value of the current small - size sliding window is less than the defect threshold , that is , then mark the current small - size sliding window as a suspected defect window and include it in the next - stage refinement recognition.

[0161] This step uses structural similarity measurement, takes the normal material template set as a reference benchmark, generates an interpretable "defect response value" for the small - size sliding window, and quickly screens out potential abnormal areas to avoid generalization bias caused by incorrect learning.

[0162] Step S223: Perform local boundary reconstruction and multi - scale verification on the suspected defect window.

[0163] To further improve the robustness of defect recognition and suppress false detections, for suspected defect windows where the defect response value is less than the defect threshold, that is perform local boundary reconstruction, multi-scale verification, and structural confidence evaluation on the suspected defect windows to ensure that the output results have stability and structural consistency.

[0164] For the suspected defect windows, construct resampling regions with multiple different neighborhood scales , where represents the scale radius, , represents three different radii, reflecting multi-scale; the resampling region is represented as follows:

[0165]

[0166] The resampling region represents the normalized points whose distance from the center of the small-size sliding window is less than the scale radius;

[0167] For each scale radius of the resampling region extract local feature vectors , and the extraction method of the local feature vector is the same as that of the local feature vector within the sliding window. First, calculate three local structural features of the point set within the resampling region , namely roughness, normal vector change rate, and point density, and integrate the three local structural features of roughness, normal vector change rate, and point density into a local feature vector. Calculate the similarity between the local feature vector of the resampling region for each scale radius and the normal material template set to obtain the defect response value for each scale radius . The similarity calculation formula is as follows:

[0168]

[0169] where, represents the scale radius defect response value, represents the scale radius local feature vector of the resampling region and the minimum Euclidean distance of the normal material template set.

[0170] Step S224: Calculate the structural confidence at multiple scale radii and output the defect detection result.

[0171] Statistically, at multiple scale radii, the defect response value is lower than the defect threshold The frequency, defining the structural confidence is defined as:

[0172]

[0173] wherein, is an indicator function. If then ; if then ; is the scale number.

[0174] The structural confidence threshold is denoted as ; if the structural confidence is greater than or equal to the structural confidence threshold, i.e., , then the current small-size sliding window is identified as a structural defect area, and the detection result is output; and the local feature vector of the current small-size sliding window is included in the defect characterization library as a defect candidate set. The defect characterization library is represented as follows:

[0175]

[0176] wherein, represents the total number of defect feature clustering clusters, represents the th defect mode.

[0177] When , the local feature vector of the current small-size sliding window is directly used as the 0th defect mode , and is recorded as the attribution label of the current small-size sliding window .

[0178] When , calculate the minimum Euclidean distance between the local feature vector of the current small-size sliding window and the defect characterization library. The calculation method is as follows:

[0179]

[0180] The defect mode when the Euclidean distance between the local feature vector of the current small-size sliding window and the defect characterization library is the smallest is recorded as the attribution label of the current small-size sliding window .

[0181] Regularly perform clustering analysis on the defect candidate set to extract representative defect modes , which is represented as follows:

[0182]

[0183] Among them, represents the th defect clustering cluster; represents the th defect clustering cluster eigenvalue size. The newly extracted representative defect pattern will be dynamically added to the defect characterization library for priority matching in subsequent recognition processes, and constructing the structural anomaly recognition memory of the system.

[0184] If the structure confidence is less than the structure confidence threshold, that is , it is regarded as a low-confidence noise area and is ignored or re-inspected twice.

[0185] By judging the confidence through the consistency of multi-scale structural responses, it can effectively solve the misguidance of single-scale results caused by factors such as blurred edges, smooth concavities and convexities, and improve the stability and generalization ability of the system under complex building structures.

[0186] Step S23: During the defect recognition process of subsequent small-size sliding windows, different processes are executed according to the attribution label of the previous small-size sliding window. If the attribution label of the previous small-size sliding window is a defect pattern, the current small-size sliding window is preferentially compared with the defect pattern of the previous small-size sliding window; if the attribution label of the previous small-size sliding window is a normal template, the current small-size sliding window is preferentially compared with the normal template of the previous small-size sliding window to quickly judge whether there are no defects.

[0187] The present invention utilizes the structural continuity hypothesis to greatly compress the comparison frequency of model inference, and is particularly suitable for scenarios with extremely high structural continuity such as building exterior walls, which not only ensures the integrity of area coverage but also improves the system operation efficiency.

[0188] Step S231: If the attribution label of the previous small-size sliding window is a defect pattern, the current small-size sliding window is preferentially compared with the defect pattern of the previous small-size sliding window to achieve rapid recognition of the defect pattern.

[0189] Adopt the method of step S221 to perform fine-grained scanning on the standardized point cloud of the current small-size sliding window, extract three local structural features of roughness, normal vector change rate and point density, and construct the local feature vector of the current small-size sliding window .

[0190] If the previous small-size sliding window attribution label is a defect pattern , the local feature vector of the current small-size sliding window is preferentially compared with the defect pattern of the previous small-size sliding window Compare, calculate the similarity, and obtain the defect response value of the current small-size sliding window defect pattern , and the calculation method is as follows:

[0191]

[0192]

[0193] Among them, represents the defect response value of the current small-size sliding window defect pattern; represents the th local feature vector of the sliding window in step S132 and the minimum Euclidean distance of the defect characterization library; represents the maximum value of the minimum Euclidean distances between all local feature vectors of the sliding window in step S132 and the defect characterization library.

[0194] Set the continuous defect threshold . If the defect response value of the current small-size sliding window defect pattern is greater than or equal to the continuous defect threshold, that is, when, it is considered that there are continuous defects in the current small-size sliding window, output the detection result, and record the defect pattern of the previous small-size sliding window as the attribution label of the current small-size sliding window .

[0195] If the defect response value of the current small-size sliding window defect pattern is less than the continuous defect threshold, that is, when, execute steps S222 - S224 to perform defect judgment by matching the full normal material template set.

[0196] Step S232: If the attribution label of the previous small-size sliding window is the normal template, compare the current small-size sliding window with the normal template of the previous small-size sliding window first to quickly judge whether there are no defects.

[0197] Adopt the method of step S221 to perform fine-grained scanning on the standardized point cloud of the current small-size sliding window, extract three local structure features of roughness, normal vector change rate, and point density, and construct the local feature vector of the current small-size sliding window .

[0198] If the previous small-size sliding window attribution label is the normal template , compare the local feature vector of the current small-size sliding window with the normal template of the previous small-size sliding window first for rapid similarity comparison to obtain the local defect response value of the current small-size sliding window , and the calculation method is as follows:

[0199]

[0200] in, Indicates the local defect response value of the current small-size sliding window.

[0201] If the local defect response value of the current small-size sliding window is greater than or equal to the defect threshold, that is , the current small-size sliding window is considered to be a normal area, and the normal template of the previous small-size sliding window is Recorded as the current small size sliding window The attribution label.

[0202] If the local defect response value of the current small-size sliding window is less than the defect threshold, that is, When the defect is determined, execute steps S222 to S224 to match the normal material template set.

[0203] The present invention uses spatial continuity to skip most of the redundant template matching process, improve the speed, and further enhance the ability to respond quickly to "material-dominated continuous segments". Through the "observation-recording-reuse" of defect structures, the system can gradually accumulate experience in identifying defect types during the identification process, and achieve self-growth from cold start to stable recognition performance.

[0204] The present invention is developed around the three stages of "normal material template modeling-small-size sliding window defect discrimination-neighborhood perception scheduling and loading", and constructs a lightweight building surface defect recognition process with structural learnability, judgment responsiveness and computational controllability. Among them, the normal material template set in step S1 provides a reference benchmark with high structural consistency, supporting the rapid similarity calculation of local features of each scanned small-size sliding window in step S2; at the same time, the defect response sequence generated in step S2 continuously builds and optimizes the defect structure response library through a multi-scale confidence mechanism and anomaly clustering process, so that the system has the ability to autonomously identify and memorize unknown defect types. The small-size sliding window scheduling logic is further bound to the adjacent window material context to realize the adaptive sliding strategy of "neighborhood drive-priority comparison-timely recovery", ensuring that low-load operation is always maintained when processing large-scale point cloud data. Through the synergistic integration of the above steps, the present invention realizes high-precision, low-resource, and highly interpretable dynamic recognition of complex building surface defects, and has good adaptability to on-site deployment.

[0205] It can be understood that the present invention describes the technical solutions through multiple embodiments. Those skilled in the art can make reasonable adjustments and equivalent replacements to the process structure, data organization form, or discrimination algorithm without departing from the core idea under the inspiration of the present invention. Therefore, the protection scope of the present invention should not be limited to the specific embodiments in this specification, but should be subject to the scope of the appended claims, covering all variant solutions that substantially achieve the same technical effects.

Claims

1. A lightweight detection method for identifying building surface defects, characterized in that, The method includes the following steps: Step S11: Perform standardized preprocessing on the three-dimensional point cloud data collected by the terahertz radar to obtain a standardized point cloud; Step S12: Use a large-size sliding window to perform local scanning on the standardized point cloud, extract three local structural features of roughness, normal vector change rate, and point density, and construct a local feature vector; Step S13: Construct a normal material template set through clustering analysis, and perform spatial connectivity analysis during the construction of the normal material template set; Step S21: Use a small-size sliding window to perform fine-grained scanning on the standardized point cloud, and plan the sliding path of the small-size window according to the spatial adjacency rule, so that any small-size sliding window has spatial overlap or shared boundary points with the previous small-size sliding window; Step S22: Perform a full normal material template set matching to judge defects on the local feature vector of the first small-size sliding window; Step S23: During the defect recognition process of subsequent small-size sliding windows, different processes are executed according to the attribution label of the previous small-size sliding window. If the attribution label of the previous small-size sliding window is the defect mode, the current small-size sliding window is preferentially compared with the defect mode of the previous small-size sliding window to achieve rapid recognition of the defect mode; if the attribution label of the previous small-size sliding window is the normal template, the current small-size sliding window is preferentially compared with the normal template of the previous small-size sliding window to quickly judge whether there are no defects.

2. The lightweight detection method for identifying building surface defects according to claim 1, wherein Step S11 includes the following steps: The three-dimensional point cloud data collected by the terahertz radar is represented as ; where represents the th three-dimensional sampling point on the building surface, represents the set of real numbers, represents the number of three-dimensional sampling points; the calculation process of the standardization preprocessing is as follows: Among them, represents the th standardized point after preprocessing, is the geometric center of the point cloud, is the standard deviation of the spatial distribution of the point cloud, represents the square of the Euclidean distance.

3. The lightweight detection method for identifying building surface defects according to claim 2, wherein Step S12 includes the following steps: The sliding window, and its corresponding sampling point set is: Among them, represents the center of the sliding window, represents the side length of the sliding window, represents a cubic region centered on the center of the sliding window with a side length of ; The th sliding window contains points, is the number of sliding times required for a complete scan of the building surface; The sliding window roughness is calculated as follows: Among them, represents transpose, represents the normal vector of the least squares fitting plane of the th sliding window local least squares fitting plane, is the plane offset term of the The normal vector change rate of the sliding window is calculated as follows: Among them, represents the th standardized point principal direction normal vector of the neighborhood it belongs to; The point density of the sliding window is calculated as follows: Among them, represents the volume of the convex hull of the point set in the th sliding window, represents the convex hull formed by the point set in the th sliding window, and represents the volume calculation function. Integrate the three local structural features of roughness, normal vector change rate, and point density into a local feature vector. The local feature vector of the ith sliding window is expressed as follows: 。 4. The lightweight detection method for identifying building surface defects according to claim 3, characterized in that In Step S13, the normal material template set is constructed in the following way: Construct a spatial structure feature set from all local feature vectors of sliding windows , input the spatial structure feature set into the clustering module to perform unsupervised learning, extract the main features, automatically cluster the regions that meet the requirements of high consistency and structural continuity, and construct a normal material template set; the normal material template set is represented as follows: Among them, represents the normal material template set; represents the total number of feature clustering clusters; represents the th normal template; is the th feature clustering cluster, represents the magnitude of the feature value of the th feature clustering cluster.

5. The lightweight detection method for identifying building surface defects according to claim 4, characterized in that In Step S22, the full normal material template set matching to judge defects includes the following steps: Step S221: Perform a fine-grained scan on the normalized point cloud of the current small-size sliding window to extract roughness , rate of change of normal vector , and point density These three local structural features are used to construct the local feature vector of the current small-size sliding window ; Step S222: Calculate the defect response value of the current small-size sliding window. If the defect response value is greater than or equal to the defect threshold, the current small-size sliding window is a normal area, and record the attribution label of the current small-size sliding window; if the defect response value of the current small-size sliding window is less than the defect threshold, mark the current small-size sliding window as a suspected defect window; Step S223: Perform local boundary reconstruction and multi-scale verification on the suspected defect window; Step S224: Calculate the structural confidence at multiple scale radii and output the defect detection result.

6. The lightweight detection method for identifying building surface defects according to claim 5, wherein In Step S222, the defect response value of the current small-size sliding window is calculated in the following way: Among them, represents the current defect response value of the small-size sliding window, represents the current local feature vector of the small-size sliding window and the minimum Euclidean distance from the normal material template set; is a tiny positive number to prevent division by zero; represents the Euclidean distance; represents the th local feature vector of the sliding window and the minimum Euclidean distance from the normal material template set, represents the maximum value of the minimum Euclidean distances between all local feature vectors of the sliding windows and the normal material template set; represents taking the maximum value for all of them.

7. The lightweight detection method for identifying building surface defects according to claim 6, characterized in that, In Step S23, the current small-size sliding window is preferentially compared with the defect mode of the previous small-size sliding window in the following way: Previous small - size sliding window The attribution label is the defect mode , the local feature vector of the current small - size sliding window is preferentially compared with the defect mode of the previous small - size sliding window to calculate the similarity, and obtain the defect response value of the defect mode of the current small - size sliding window , and the calculation method is as follows: Among them, represents the defect response value of the current small-size sliding window defect mode; represents the th sliding window local feature vector and the minimum Euclidean distance from the defect characterization library; represents the maximum value of the minimum Euclidean distances between all sliding window local feature vectors and the defect characterization library.

8. The lightweight detection method for identifying building surface defects according to claim 6, characterized in that, In Step S23, the current small-size sliding window is preferentially compared with the normal template of the previous small-size sliding window in the following way: Previous small - size sliding window The attribution label is the normal template , the local feature vector of the current small - size sliding window is preferentially compared with the normal template of the previous small - size sliding window to quickly obtain the similarity, and the local defect response value of the current small - size sliding window is obtained , and the calculation method is as follows: Among them, represents the local defect response value of the current small-size sliding window.

Images