Conference Agenda

As bridges age, timely and accurate identification and quantification of cracks and structural defects are essential for ensuring safety and durability. Existing 2D image-based methods are limited by the lack of reliable depth information, hindering precise description of defect shape, severity, and evolution. In contrast, 3D point cloud data provides high-resolution spatial geometric information, supporting more accurate defect analysis. This paper proposes a scale-adaptive cross-modal defect detection and quantification method that integrates 2D image data with 3D point cloud information. First, a 3D semantic segmentation model with a cross-modal registration strategy enables precise segmentation of large structural cracks, while an image-guided fusion mechanism facilitates accurate depth map generation for micro-cracks. Additionally, a 2D segmentation model-guided defect 3D point cloud segmentation algorithm is proposed for detecting small durability cracks. Finally, a 3D defect quantification framework is developed to systematically measure crack length, width, depth, and volume, providing multi-dimensional indicators for structural assessment. Experimental results show that the proposed cross-modal fusion approach significantly outperforms single-modality methods. By integrating PointNet++ and U-Net, the system achieves high-precision multi-scale segmentation with an average mIoU of 0.87–0.89. Furthermore, the framework enables reliable 3D quantification of crack length, width, depth, and volume, providing a more comprehensive and objective basis for structural health assessment than traditional 2D methods. This research provides an effective approach for overcoming the limitations of 2D methods, offering a reliable solution for bridge structural health monitoring and durability prediction.

Ensuring the visual appearance quality of automotive interior buttons requires reliable inspection of painted surfaces. However, powder-coated finishes present complex reflectance behaviors, including directional glare and irregular highlight patterns, which often mask or resemble subtle defect features. To address this limitation, we explore a BRDF-based Photometric Stereo (PS) approach that accounts for complex reflectance characteristics when capturing fine geometric features of painted button surfaces. Leveraging BRDF-based modeling and multi-directional illumination, the PS method derives pixel-wise surface normals that are more robust to specular highlights and better represent true surface geometry. These surface normal maps are subsequently utilized as input to a deep learning model for detecting defective and anomalous surface regions. When integrated with AI-based anomaly detection, PS-derived geometric representations contributed to more reliable identification of fine surface irregularities, effectively enhancing the detectability of micro-defects that are visually indistinct in painted automotive components.

Vision-based structural health monitoring of stay cables has predominantly relied on dynamic imaging to extract vibration characteristics for tension estimation, typically requiring video sequences, high-speed cameras, and stable close-range deployment. This study presents an alternative static-image approach that estimates stay-cable tension from gravitational sag captured in a single static digital image obtained from far-field photography. The method integrates a parabolic sag-tension relationship for inclined cables with a lightweight image-processing workflow, in which global thresholding and Canny edge detection are used to extract the cable profile, followed by quadratic fitting to determine sag and cable force. The procedure enables marker-free measurement and calibration-free deployment, allowing rapid field application with minimal on-site preparation. The proposed approach is intended for inspection scenarios in which cable arrays on one bridge side approximately lie in a common vertical plane and near-orthogonal far-field imaging can be achieved. Under such conditions, basic image segmentation techniques provide sufficient accuracy, although complex backgrounds or low-contrast scenes may reduce robustness and motivate the use of more advanced segmentation strategies. Laboratory strand experiments using a 61 MP customer-grade camera demonstrated small tension estimation errors under near-orthogonal imaging, and sensitivity studies showed that moderate viewing-angle deviations introduce limited bias. Field validation on the Kao-Ping-Hsi Bridge was conducted using far-field images acquired from distances of approximately 550 m. For the longest cables with larger sag amplitudes, estimated tensions agreed with independent reference values within 2%, while errors for the shortest and steepest cables remained below 8%. The results demonstrate that single-image gravitational sag extraction offers a practical, low-intervention alternative to dynamic vision-based techniques for remote stay-cable force assessment when sensor installation or vibration testing is impractical.