Automated Defect Detection on Dry-Hanging Stone Curtain Walls through Colored Point Clouds

Abstract

Stone curtain walls are widely used in contemporary architectures; however, their regular inspection is always labor-intensive, time-consuming, and hazardous due to the complex and enclosed spatial structure of these high-rise building enclosures. To address this issue, this study proposes an automated and novel inspection method, which is composed of the following three steps: First, we utilize 3D laser scanning technology to capture colored point cloud data of the stone curtain wall system; subsequently, by extracting and processing the integration of color and depth information, the stone panels and end sealants are precisely segmented; finally, various defects, such as cracks, unevenness, and irregularities, are automatically identified through artificial intelligence algorithms in a timely manner. To validate the proposed method, an on-site experiment was carried out to demonstrate the effectiveness in detecting multiple defects concurrently on stone curtain walls. The experimental results showed that our proposed method could provide a non-contact and automated inspection alternative for all the stone curtain walls with a high accuracy of anomaly detection, facilitating rational maintenance plans and strategies to ensure the safety and performance of these modern building enclosures.

1. Introduction

Stone curtain wall systems are a popular form of architectural exterior decoration. These systems enhance buildings with sophisticated and elegant design styles. In modern cities worldwide, they account for approximately 30% of curtain walls [1]. Architects and builders primarily use two methods for installing stone curtain walls, namely wet-sticking and dry-hanging [2]. The first is a traditional mortar-set system and uses a steel mesh and layered cement paste to adhere stone panels to walls. However, safety concerns have reduced its recent usage [3]. These concerns include detached and hollowed panels, often resulting from poor construction quality and the irreversible degradation of cement mortar. In contrast, the dry-hanging method secures stone panels to the building structure using metal brackets and specialized anchoring adhesives, avoiding unstable cement-based paste. This system typically conceals curtain mullions, transoms, fittings, connections, and joints behind the stone panels. As a result, it enhances the overall aesthetics and improves the building’s thermal performance [4]. However, installing and maintaining concealed-frame stone curtain walls can be challenging due to limited access for cleaning, inspection, and repairs. Consequently, the accurate and timely identification of various defects in stone curtain wall systems remains a critical issue in their lifecycle management [5]. These defects include cracks, unevenness, panel irregularities, and loosening connections or joints. Such defects not only compromise fire resistance, waterproofing, dust resistance, and lighting protection performance but also pose falling hazards to building occupants and pedestrians [6].

To address these safety concerns, researchers and industry professionals have developed various technologies for curtain wall inspection. The Fiber Bragg Grating (FBG) sensing technology monitors stress and deformation by measuring temperature, pressure, vibration, and displacement [7,8]. On-site vibration testing identifies loose connections between curtain panels and supporting structures [9]. While these methods provide valuable insights, they require micro-destructive deployment, which can impact curtain wall operation. Non-destructive technologies are preferred in practical applications. For metal curtain walls, millimeter-wave imaging detects surface defects [10]. Infrared imaging identifies damage to structural and end sealants in glass curtain walls, exploiting thermal differences between panels and sealants [11]. Research on automated curtain wall construction also offers relevant insights. For example, Wu et al. used semantic segmentation to identify curtain wall frames [12], Johns et al. applied computer vision to measure distances between panels and frames [13], and Yuan et al. utilized non-contact digital imaging to detect bending deformations in glass curtain wall panels [14].

Research in the field of curtain walls predominantly focuses on glass and stone systems, with a notable bias towards glass [15]. This emphasis on glass systems can be attributed to several factors. Firstly, glass curtain walls are more prevalent in contemporary architecture, and their higher maintenance costs have spurred increased research interest. Additionally, stone curtain walls present unique challenges for inspection due to their complex surface features, including varied textures and color patterns. Furthermore, the lack of standardization in stone products has impeded the adaptation of existing inspection technologies to stone curtain wall systems. These factors collectively contribute to the current research landscape, where glass curtain walls have received disproportionate attention compared to their stone counterparts. Glass curtain wall inspection primarily focuses on glass panels and structural adhesive performance. In contrast, stone curtain wall inspection encompasses metal frames, stone panels, and hangers. The closed, narrow space between stone curtain walls and exterior walls complicates access for inspectors and equipment. This limitation significantly increases the difficulty of safety inspections for stone curtain walls during operation [16]. Glass curtain walls, being more open and transparent, offer greater operational space and convenience when applying various detection technologies.

As illustrated in

Table 1. Comparison of the effectiveness of stone curtain wall detection methods.

Detection Method	Effectiveness Description	Limitations
Impact-echo test	Determine panel abnormalities	High workload, relies on experience, risky at high altitudes
Pull-out test	Inspect panel anchoring	Destructive, difficult for high-altitude use, only for near-ground panels
Endoscopy	Peep at local panel connections	Limited scope
Laser vibration frequency method	Analyze defects by obtaining stone’s frequency	Lab research only, accuracy affected, stability lacks verification
Colored point cloud method	Detect defects in stone curtain wall panels efficiently and accurately	Can only detect the surface

, Liu et al.’s comprehensive review of curtain wall safety inspection methods revealed that contact-based techniques predominate in the assessment of stone curtain walls [17]. Initially, impact-echo testing is used to identify panel anomalies, followed by pull-out testing to assess the bond strength between panels and between mullion-transom structures. Additionally, endoscopy can be employed to inspect the local connections of the panels. However, each of these methods has significant limitations. Impact-echo testing requires inspectors to tap each panel individually, making it time-consuming, labor-intensive, and highly dependent on personal expertise. Moreover, high-altitude inspections pose safety risks to workers and pedestrians. The pull-out method, a destructive test, is difficult to implement at heights due to safety and feasibility concerns. Consequently, its application is often limited to panels near ground level. Endoscopy, on the other hand, has very limited applicability and can only detect the edge areas of the curtain wall. Consequently, these methods are suitable only for safety assessments of stone curtain walls, typically conducted post-incident following panel detachment, and are not preventive measures against such accidents [18]. In response to these challenges, Bao et al. proposed a non-contact method using laser vibrometry. This technique measures the first-order natural frequency of the structure system to analyze defects in stone curtain walls. However, this method remains confined to laboratory research. Its accuracy depends significantly on test point location and laser incidence angle, and its stability lacks sufficient validation [19]. The aforementioned research indicates that the safety inspection methods for stone curtain walls require further refinement and improvement, necessitating their evolution towards more efficient, accurate, and non-contact approaches.

Abnormalities in curtain wall panels often indicate unsafe conditions in stone curtain walls. These abnormalities typically result from panel or connection failures. Panel failures primarily manifest as cracking or warping, while connection failures often lead to panel misalignment [20]. Given the potential safety risks associated with these abnormalities, timely and accurate detection is crucial for maintaining building safety and preventing accidents. A rapid inspection of these panels can provide preliminary safety assessment criteria for stone curtain walls [4]. To address this critical need for efficient and reliable detection methods, researchers have explored various rapid detection technologies for architectural surface defects. Among these, point cloud-based methods have gained prominence due to their advantages in data acquisition speed, precision, and non-contact nature [21]. For example, Yu et al. utilized drone-captured images to generate point clouds for rapid post-earthquake facade damage assessment [22]. Wardach et al. employed laser scanners to meticulously examine smooth building facades, enabling high-accuracy defect identification [23]. Additionally, Valero et al. demonstrated the efficacy of ground-based laser scanning in automatically detecting and classifying masonry wall defects. These studies collectively highlight the viability of 3D laser scanning technology in obtaining façade point clouds for defect detection [24].

Despite these technological advancements, the application of point cloud and image technologies to detect panel defects in stone curtain walls remains largely unexplored. Few researchers have investigated this specific application. Point cloud-based analysis holds significant potential for identifying panel abnormalities such as cracks, warping, and misalignments, which could substantially enhance the safety assessments of stone curtain walls. The integration of advanced image processing algorithms with rich point cloud data could enable the detection of subtle variations in surface geometry indicative of panel or connection failures. This approach could incorporate machine learning techniques to classify defects based on geometric and textural features extracted from point clouds. Furthermore, combining multispectral or hyperspectral imaging with point cloud data could provide insights into stone material properties, potentially revealing early signs of degradation or structural weaknesses invisible to the naked eye. Such a comprehensive approach could lead to the development of an automated inspection system for stone curtain walls, ensuring their safety and longevity.

Based on the preceding analysis, this study proposes a novel method for detecting stone curtain wall panel defects using point cloud data. The methodology comprises several steps. First, a 3D laser scan of the stone curtain wall facade is conducted. The resulting point cloud is then transformed into a depth map. Leveraging the sealants between stone curtain wall panels, a computer vision approach is employed to segment individual panels. Subsequently, the presence of cracks and misalignments on panel surfaces is detected by incorporating the architectural façade design information. This process culminates in a comprehensive defect status assessment for each panel. The innovation of this research lies in its integration of point cloud data conversion to depth maps, computer vision techniques, and design information for defect detection on stone curtain wall panels. This method enables the efficient and thorough acquisition of stone curtain wall panel defect data while avoiding risky or destructive sampling tests thus providing a reliable reference for subsequent inspections.

2. Proposed Method

This study introduces a novel approach for detecting defects in stone curtain wall panels, utilizing the Hough Transform as a key analytical tool (Figure 1). The process begins with preprocessing point cloud data from the stone curtain wall to generate depth and color images. Next, the Hough Transform algorithm is applied to the color images to isolate, refine, and categorize inter-panel sealants. The method then employs image fusion to integrate sealant information with the depth image, achieving precise panel segmentation. This results in individual depth images for each panel, encapsulating detailed spatial information. The final step involves inspecting these depth images to identify and annotate defects. This approach offers a rapid, non-invasive technique for extracting defect information from stone curtain wall panels with high accuracy and robustness. It streamlines the defect detection process and enhances result reliability, thereby contributing to the advancement of inspection methodologies in architectural façade maintenance and restoration.

2.1. Point Cloud Data Preprocessing

This study employs a three-dimensional laser scanner capable of generating colored point clouds to scan stone curtain walls and performs registration on the obtained point clouds from multiple measurement stations. In order to achieve high-precision point cloud alignment, a strategy of coarse registration followed by fine registration is typically adopted. First, Principal Component Analysis (PCA) is utilized for coarse registration. PCA computes the covariance matrix of the point cloud as follows:

$$C={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N(x_i-\overline{x})}{(x_i-\overline{x})}^T$$

(1)

where $\overline{x}$ is the mean of the point cloud data. This process extracts the eigenvalues and eigenvectors, with the eigenvectors corresponding to the principal directions of the point cloud. By aligning the main direction of the source point cloud to that of the target point cloud, most of the posed differences between the point clouds can be effectively eliminated, laying the groundwork for subsequent fine registration.

Building on the coarse registration, the Iterative Closest Point (ICP) algorithm is further employed for fine registration. This algorithm iteratively optimizes the alignment error between point clouds to achieve precise registration. In each iteration, ICP first finds the closest point in the target point cloud for each point in the source point cloud and establishes point-to-point correspondences. Then, we proceed to minimize the sum of the squared Euclidean distances between the corresponding points as follows:

$$E(R,t)={{\displaystyle \sum _{j=1}^M‖p_i-(R_{q_j}+t)‖}}^2$$

(2)

where p_i and q_j are the points from the source and target point clouds, respectively, R is the rotation matrix, and t is the translation vector. The optimal rotation and translation vectors are computed to progressively align the source point cloud with the target point cloud. This process continues iterating until the registration error converges to a predefined threshold or reaches the maximum number of iterations. By combining PCA and ICP, high-precision multi-station point cloud registration can be achieved, ensuring that the generated three-dimensional model possesses a high degree of spatial consistency and accuracy.

After completing the registration, we obtained the global point cloud, followed by manually extracting the point cloud corresponding to the target curtain wall.

2.2. Two-Dimensional Image Generation

This study assumes that the ground plane is horizontal and that the design angle of the stone curtain wall with respect to the ground plane is known. Initially, the Random Sample Consensus (RANSAC) algorithm is employed to fit the point cloud to a hypothesized plane, which serves as the reference plane. While RANSAC is robust against outliers, it also has limitations, such as sensitivity to the choice of thresholds and the potential to converge to local optima. To address these limitations, adaptive threshold selection is implemented based on the instrument’s specified accuracy at different distances. Specifically, thresholds of 4 mm are set at distances of 10 m and 8 mm at 20 m from the instrument’s origin, with linear interpolation for intermediate distances. Additionally, multiple RANSAC iterations are performed using different random seeds to avoid local optima. Our sampling strategy involves selecting a minimum of three non-collinear points in each iteration, with the number of iterations determined by the desired probability of identifying a good model.

Subsequently, the BFGS (Broyden–Fletcher–Goldfarb–Shanno) optimization algorithm is used to iteratively update the parameters of the plane equation to ensure that it aligns with the design angle of the ground plane. This is achieved by utilizing gradients and the inverse of the approximate Hessian matrix to determine the search direction, thereby reducing deviations and ensuring the plane equation conforms to the design specifications. The point cloud is then projected onto the reference plane, with projection distances recorded for each point. To optimize processing, the Gram–Schmidt process is applied to compute the orthonormal basis of the plane, which facilitates the mapping of 3D points onto a 2D plane thus reducing data dimensionality. The dimensions of the images are determined based on the boundaries of the point cloud and the pixel size on the 2D plane. The method then involves traversing each point to calculate average projection distances and color values for points within each pixel location, assigning these averages as pixel values for both depth and color images. This process ultimately culminates in the generation of visual representations of the depth and color images (Figure 2).

In this study, the pixel size was determined based on two key factors. First, the study in [25] indicates that the joint width of stone curtain wall panels typically exceeds 8 mm, generally ranging from 10 to 15 mm in the actual designs. To accurately represent these joints, the pixel size should not exceed 10 mm. Secondly, the study in [26] notes that the height of low-rise civil buildings or high-rise podium buildings generally does not exceed 24 m. At this distance, ordinary 3D laser scanners typically have an error of about 10 mm. For instance, the Leica BLK360 has an error of 8 mm at 20 m. Considering these applications, the pixel size should be larger than the 3D laser scanner’s error. Balancing these factors, we set the pixel size to 10 mm, which allows for accurate joint representation while accounting for scanner errors at relevant distances.

2.3. Sealant Line Extraction Based on Hough Transform

In this study, the acquired color image undergoes a series of sophisticated processing steps to enhance its utility for detecting and analyzing linear features. The initial step involves cropping the color image by reducing 10 pixels from each edge to mitigate edge noise, which is a common issue in image processing. This is followed by the application of a 3 × 3 Gaussian filter to smooth the image, effectively reducing the impact of salt-and-pepper noise. To further refine the image quality, distortion correction and perspective transformation techniques are employed to rectify any distortions introduced by the camera lens and to convert the image from an oblique view to a frontal view. The next phase of the process involves the use of the Canny edge detection algorithm to identify edges within the image. To improve the signal-to-noise ratio, a lower threshold is set to enhance the sensitivity of the edge detection. This approach is particularly beneficial for capturing subtle linear features that might otherwise be overlooked. Following edge detection, the Hough Transform is applied to detect sealant lines, which are always constructed by structure silicone sealants. The Hough Transform is a robust technique for identifying lines and other geometric shapes in images. In the context of line detection, this technique maps each point in the image space to a parameter space (ρ, θ), where ρ represents the shortest distance from the origin to the line and θ denotes the angle between the line and the x-axis. For each point (x, y) in the image space, all possible lines in the parameter space can be represented by the following equation:

$$\rho =x\cos \theta +y\sin \theta$$

(3)

In the implementation of the Hough Transform, an integral step involves constructing a two-dimensional accumulator to record the vote count for each (ρ, θ) in the parameter space. For every edge point in the image, the process iterates over θ to compute ρ and the increments at the corresponding positions in the accumulator. Upon completion of the accumulation, peaks in the accumulator signify lines in the image space. By identifying these local maxima, the parameters of the lines are determined, with these peaks corresponding to prominent lines within the image.

The Hough Transform exhibits remarkable efficacy in processing color images, primarily due to the similarity of edge information in color and grayscale images, enabling the effective detection of linear structures. Its robust performance in the presence of noise and interference allows it to mitigate, to a significant extent, the potential confounding effects of chromatic information. This robustness stems from the transform’s ability to abstract geometric features from the image, focusing on structural elements rather than chromatic variations, and its global analysis approach that aggregates evidence across the entire image space. Consequently, the Hough Transform maintains its effectiveness even in complex, multi-chromatic environments characterized by localized noise or partial occlusion.

After detecting the sealant lines, we classify them into two categories based on their angles relative to the y-axis—horizontal (>45°) and vertical (≤45°). Then, these lines are grouped based on their positions along the y- and x-axes, as well as their relative densities, to identify clusters corresponding to sealant connections in the stone curtain wall. In this study, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) method for clustering was employed. This method was chosen for its ability to handle clusters of arbitrary shape and its robustness to noise, making it particularly suitable for grouping potentially imperfect line detections in complex architectural images. Subsequently, the RANSAC (Random Sample Consensus) algorithm is employed to fit lines to the different clusters. RANSAC’s strength lies in its ability to accurately fit models even in the presence of numerous outliers, making it ideal for our application where detected lines may be imperfect due to image noise or architectural complexities. This robustness ensures a more accurate representation of the true sealant lines, adapting to the varying quality of input data.

To further refine the line parameters, the pixel spacing between the horizontal or vertical sealant lines are calculated based on panel dimensions. This step ensures that the line parameters align closely with the actual spacing between the sealant lines. The culmination of this methodological sequence is the extraction of positions corresponding to the sealant lines in the curtain wall (Figure 3). This extraction is crucial for accurately assessing the structural integrity of the stone curtain wall. The refined line parameters, obtained through the meticulous application of the Hough Transform and subsequent optimization techniques, provide a detailed map of the curtain wall’s linear features. This map facilitates the detection of defects and the assessment of maintenance requirements.

2.4. Panel Image Segmentation Based on Image Fusion

In this study, the linear positions corresponding to the stone curtain wall’s sealant lines, extracted from the color image, are mapped onto the depth image and designated as white. This mapping process can be executed directly due to the one-to-one pixel correspondence between the color and depth images, both derived from the same point cloud. Following the mapping, a kernel-based sealant line detection method is employed on the depth image to verify the line positions. This research assumes that the width of the sealant lines in the stone curtain wall is approximately constant. The verification process begins by calculating the coordinates of the intersection points between the lines. At non-intersection points, a series of points are randomly selected as centers to construct 5 × 5 kernels. The depth values of pixels within these kernels, oriented perpendicular to the line direction, are then extracted. The presence of a valley-like trend in these depth values is used to determine whether the current point belongs to a sealant line. The valley-like trend in depth values is indicative of a sealant line, suggesting a local minimum in depth that corresponds to the narrow sealant line between panels. This method is sensitive to the subtle depth variations that characterize sealant line locations. If more than 10% of the points are not identified as belonging to a sealant line, it necessitates a reevaluation of the color image through the steps outlined in Section 2.2 (Figure 4).

In cases where the sealant line width does not exceed two pixels, different kernel sizes (3 × 3, 5 × 5, and 7 × 7) were tested and the 3 × 3 kernel was found to be more sensitive to local variations but also more susceptible to noise. The 7 × 7 kernel’s large range might introduce errors from the panels. Conversely, the 5 × 5 kernel offered a good balance between detail preservation and noise reduction. Furthermore, in the case of unqualified results, several scenarios may arise, i.e., if the unrecognized points are concentrated at one end of the sealant line, it indicates that the extracted sealant line is skewed; if the unrecognized points are clustered within the middle range of a particular panel, it suggests that there may be defects at the edges of that panel; and if the unrecognized points appear at various positions, it implies that the extracted sealant line is incorrect. This multi-faceted discussion helps to better understand the limitations and applicability of the kernel-based detection method.

Upon successful verification of the line positions, the depth image is segmented into multiple grid regions, with each grid region representing a panel of the curtain wall. This segmentation is critical for the subsequent analysis of the structural integrity and alignment of the panels. This methodology ensures that the detected sealant lines are accurately represented in the depth image, providing a reliable basis for assessing the condition of the stone curtain wall.

2.5. Panel Defect Detection

For each segmented grid region, the internal image depth data and the computed minimum and maximum values of the pixels within the region are extracted. These values reflect the range of depth variations within the grid. The difference between the maximum and minimum depths in each region is calculated and converted into actual physical distance. This conversion is achieved by dividing the difference by 255 and multiplying it by the global depth difference (see Equation (4)). The reason for first dividing by 255 is that the depth data in the image are typically stored as grayscale values, and the range of grayscale values is usually from 0 to 255. By dividing the difference by 255, we normalize the depth difference to a scale of between 0 and 1, which can then be multiplied by the global depth difference to obtain the actual physical distance. This process accounts for the scaling of depth values in the image, ensuring that the calculated physical distance aligns with the actual dimensional characteristics of the curtain wall panels.

$$\Delta L_{\mathrm{panel}}=\left|{\displaystyle \frac{D_{\max }-D_{\min }}{255}}\right|\times L_{\mathrm{global}}$$

(4)

where D_max and D_min represent the maximum and minimum pixel depth values within the panel region, respectively, L_global denotes the actual global maximum depth difference, and ΔL_panel signifies the maximum actual depth difference of the panel. The final output is a matrix that encapsulates the depth variation information for each grid region. The numerical positions within this matrix correspond directly to the positions of the panels in the image. The relevant results are overlaid onto the depth map, illustrating the image and depth analysis outcomes for specific grid regions (see Figure 5). This provides a visual tool that facilitates the evaluation and comparison of depth discrepancies across different regions. To further refine our analysis, the Interquartile Range (IQR) method is employed to identify outliers based on the matrix values. This method involves calculating the first (Q1) and third quartiles (Q3), followed by determining the IQR, which is defined as Q3 − Q1. The boundaries for outliers are then established as follows: the lower boundary is set at Q1 − 1.5 × IQR and the upper boundary is set at Q3 + 1.5 × IQR. Any values falling below the lower boundary or above the upper boundary are considered outliers. This statistical approach is effective in identifying anomalies that may indicate defects or irregularities in the stone curtain wall panels. By applying the IQR method, the depth values that deviate significantly from the majority can be systematically filtered out, which may correspond to areas of concern such as cracks, warping, or misalignments. The use of IQR not only helps in detecting outliers but also provides a robust measure of the spread of the data, allowing for a more nuanced understanding of the depth distribution within each panel.

The depth matrix, coupled with the IQR analysis, offers a comprehensive framework for assessing the integrity of stone curtain walls. This methodology enables the detection of subtle depth variations that may escape visual inspection, thereby enhancing the safety and reliability of the assessment process. The visualization of these results on the depth map (Figure 5) provides a clear and intuitive representation of the depth analysis, facilitating the identification of areas requiring further investigation or maintenance.

3. Experimental Validation

This study conducted an experimental investigation on the stone curtain wall façade of a building in Shenzhen, Guangdong Province, China. As illustrated in Figure 6a, one of the stone curtain wall panels exhibited a distinct crack, resulting in surface unevenness with a maximum depression depth of approximately 1 cm. The width of each panel of the curtain wall was 75 cm and the height was 150 cm.

For scanning the stone curtain wall façade, the Leica BLK360 three-dimensional laser scanner was employed, which operates on the principle of laser ranging. This device is capable of providing panoramic color scanning results with a horizontal range of 360 degrees and a vertical range of 270 degrees, producing point cloud data containing both spatial coordinate and color information. This offers a comprehensive and detailed description of the scanned object, with a single measurement accuracy of 4 mm at a distance of 10 m. In our experiment, two measurement stations were established to ensure the acquisition of uniformly dense and sufficient point clouds, and the scanning results are presented in Figure 6b. Considering the plant obstructions in the experimental scene and the excessively low point cloud density in distant areas, after the point cloud registration between the measurement stations, the point cloud of a partial area of the stone curtain wall was manually selected from the obtained colored point cloud for subsequent data processing.

As illustrated in Figure 7, the data processing phase commenced with the acquisition of color and depth images from the stone curtain wall point cloud, utilizing the methodology delineated in Section 2.1. For the color image, a sophisticated combination of edge detection algorithms, such as the Canny operators, was employed to highlight the boundaries of the panels. This was followed by the application of the Hough Transform, specifically tuned to identify linear segments indicative of the sealant lines between panels. The extracted lines were meticulously categorized into horizontal and vertical groups based on their orientations relative to the y-axis and color-coded—green for horizontal and blue for vertical—to visually distinguish their orientations within the image. Subsequently, each group of lines underwent a more granular differentiation process using the DBSCAN algorithm to isolate clusters corresponding to distinct horizontal or vertical sealant lines. Each cluster then underwent a line fitting procedure using the RANSAC method to generate a best-fit line representing the average orientation of the sealant lines within that cluster. The optimized lines, now representing the definitive types of sealant lines, were color-coded with greater specificity for visual distinction and analysis. The detection results of these sealant lines were seamlessly integrated with the depth image through a pixel-matching process that aligned the color-coded lines with their corresponding depth values. By identifying the rectangular regions enclosed by the sealant line pixels, a segmentation algorithm was employed to delineate the different curtain wall panels on the depth map, critical for isolating individual panels and conducting detailed depth analyses to assess their integrity and alignment.

As depicted in Figure 8, the experimental results demonstrated the successful segmentation of a 5 × 7 array of complete stone curtain wall panels from the depth map. For each panel, the maximum and minimum pixel depths were computed, which were then utilized to determine the maximum actual depth discrepancy within each panel. This discrepancy was calculated based on the distance from the point cloud to the fitted plane during the generation of the depth map, as detailed in Table 1. In this experiment, the maximum actual global depth discrepancy was measured to be 25.55 mm. A closer examination of

Table 2. Inspection results of stone curtain wall panels.

Row Column	1	2	3	4	5	6	7
1	7.01	5.81	5.71	4.01	3.41	7.31	5.41
2	5.71	3.31	4.11	4.31	3.61	5.51	5.51
3	5.51	3.21	4.51	4.61	5.11	5.41	4.41
4	5.51	4.91	12.02	4.51	4.11	5.61	4.61
5	5.61	4.21	4.11	3.41	4.31	4.11	3.61

reveals that the stone panel situated at position (4, 3) exhibited an exceptionally high maximum actual depth discrepancy of 12.02 mm, as calculated using the IQR method. This panel’s minimum and maximum distances from the fitted plane were recorded as 8.32 and 20.34 mm, respectively, with the corresponding pixel coordinates at (3, 2) and (19, 23). The detection outcomes unequivocally indicate an anomaly in the stone panel at position (4, 3), which is aligned with the actual conditions observed on-site.

The depth extremes of this panel, as annotated, accurately reflect the real-world scenario, where the maximum depth was found at the juncture of multiple fragmented sections. This congruence between the experimental findings and the physical state of the stone panels underlined the efficacy of the proposed methodology.

4. Discussion

The defect detection method for stone curtain wall panels proposed in this study leverages the integration of three-dimensional laser scanning technology and sophisticated image processing algorithms to achieve an efficient and accurate detection of abnormalities. A key aspect of this methodology is the strategic application of the Hough Transform, which is initially employed in the color image rather than directly in the depth map. This approach is adopted due to the color image’s inherent advantage in providing rich texture and color information, which are critical for distinguishing the sealant lines from the panels. The sealant lines exhibit distinct color and texture characteristics that are markedly different from those of the surrounding panels, allowing the Hough Transform to more sensitively and accurately identify the straight-line-shaped sealant lines.

In contrast, the depth map primarily conveys the three-dimensional geometric information of the object’s surface, where the sealant lines are not as visually apparent as in the color image. The lack of sufficient contrast in the depth map makes accurate sealant line detection challenging. This difficulty is further compounded by the fact that the pixel values in the depth map are derived from the actual global maximum depth difference. When the depth difference between the colloid in the sealant lines and the panel is significantly smaller than the global maximum, the pixel values of the sealant lines and the panel become similar, rendering them indistinguishable. Additionally, the depth map, which is generated from the point cloud, inherits any instrumental errors that occurred during the point cloud collection process. By utilizing the color image, these errors can be circumvented, ensuring a more reliable extraction of the sealant lines.

The lines extracted from the color image are subsequently verified in the depth map using the methodology outlined in Section 2.3 of this study. During the experiment, a point was selected every 20 pixels along each line, and the preliminary verification results are illustrated in Figure 9. With this setting, it is ensured that there are at least three measuring points on the short side of each panel. The results indicate that the verification of horizontal lines is generally more successful, whereas the vertical lines exhibit a gradual degradation from the bottom to the top of the image. This discrepancy is attributed to the operational mode of the three-dimensional laser scanner. When the scanner, positioned on the ground, captures the stone curtain wall panels at higher elevations, the increased distance results in a larger spacing between scanning points under the same scanning angle. Consequently, the point cloud data for the panels at higher positions are less dense, and the features of the sealant lines become less pronounced. This effect is particularly evident in the color image, where the horizontal sealant lines at higher positions appear more blurred and indistinct.

Furthermore, as observed in the enlarged area of Figure 9, there is a noticeable error when data from different measurement stations are registered, leading to a ghosting phenomenon. This artifact adversely impacts the accuracy of the Hough Transform and the subsequent straight-line fitting. Out of the eight vertical sealant lines identified in the color image, two exhibit such issues. To address these challenges, this study incorporates design data to estimate the pixel width corresponding to the width of and the pixel distance between the sealant lines. Armed with this information, the extracted lines are optimized within their respective line clusters, aiming to align the spacing and slope between the lines as closely as possible to the design specifications. This optimization process ensures that the final line detection meets the stringent requirements for defect detection on the stone curtain wall panels.

5. Conclusions

This study introduces a novel preliminary defect detection method for stone curtain wall panels based on colored point cloud data. The methodology involves the acquisition of point cloud data through three-dimensional laser scanning technology, which is then transformed into depth maps and color images. The Hough Transform algorithm is employed to extract the sealant lines between panels, which are subsequently fused with the depth images to segment individual panel depth images. Each panel’s depth image is then scrutinized for anomalies, which are annotated accordingly. This approach enables the rapid, non-contact extraction of defect information from stone curtain wall panels, boasting high accuracy and robustness. Experimental validation of the method has confirmed its effectiveness and precision. The results demonstrate that the method can accurately detect anomalies on stone curtain wall panels and provide an intuitive visual tool for assessing and comparing depth discrepancies across different areas.

The theoretical contribution of this research lies in the introduction of a new defect detection method for stone curtain wall panels, utilizing non-contact three-dimensional laser scanning technology combined with image processing algorithms to identify irregularities as defect indicators. This innovation offers a fresh perspective and approach for the safety inspection of stone curtain walls. In terms of practical contributions, the proposed method allows for a rapid, large-scale detection of panels, enabling the decision for further inspection based on the results. This facilitates a more refined management of stone curtain wall panel inspections, circumventing the need for traditional, high-risk methods such as tapping, which are often subjective and limited in scope. By enhancing detection efficiency, ensuring safety and reliability, and providing a scientific basis for maintenance and management, the method aids in the formulation of rational plans and strategies.

The proposed method can be applied to complex scenarios like stone curtain walls, such as reticulated shell in parks and various pavements on roads. Future research in the field of stone curtain wall panel defect detection can focus on addressing the limitations identified in this study. Firstly, to overcome the constraints of ground-based stations in observing high-altitude curtain walls, the exploration of using drones, cellphones, and other mobile devices for point cloud data collection could be beneficial. This would enable more comprehensive scanning of the curtain wall surface, enhancing the accuracy and completeness of detection. Secondly, for the detection of curtain walls with varying shapes, further research and development of appropriate detection methods are necessary. Integrating multiple sensors and technologies, such as laser scanning, photogrammetry, and structural health monitoring systems, may be essential for the comprehensive detection and assessment of curtain walls with diverse shapes. Lastly, enhancing automation is a critical direction for future research. The development of automated point cloud data acquisition and processing systems, which minimize human intervention, is crucial for improving detection efficiency and accuracy. By advancing these areas, the field of stone curtain wall panel defect detection can continue to evolve, ensuring the safety and integrity of these architectural elements.