An image-guide–PCD modal fusion method for concrete defects spatial feature understanding

Abstract

Cracks inevitably develop in concrete structures during service, facilitating the ingress of water and chlorides, which accelerates steel corrosion and deterioration. Compared with 2D metrics, 3D indicators including crack depth and volume are crucial for determining whether cracks penetrate the protective cover, assessing material loss, and defining repair scope and risk levels. However, commonly used image based inspection methods struggle to provide 3D information of defects. To address this issue, this paper leverages the complementary strengths of 3D point cloud data (PCD) and 2D images and proposes an image-guide–PCD modal fusion method to enable precise defect detection and spatial feature quantitative characterization. First, an image-guided cross modal registration algorithm is developed to align 2D image with 3D PCD of the monitored concrete target. Then, BrC-UNet is proposed to segment defects from image with high-accuracy defect masks, which is further guided defects segmented from PCD data. Furthermore, a 3D defect quantification framework is developed to characterize crack depth, length, width, and volume. Experiments show cross modal registration errors around 2% and image-guided PCD segmentation mIoU of 0.833. The mean quantification errors for defects’ depth, length, and maximum width are 4.87%, 4.29%, and 6.04%, and volume estimation outperforms conventional methods. The proposed extraction and quantification framework transcends 2D limitations, enabling multidimensional analysis for structural evaluation and durability prediction.