Pixel-level crack segmentation and quantification enabled by multi-modality cross-fusion of RGB and depth images

Pixel-wise crack segmentation and quantification are crucial for structural evaluation and decision-making in asset management to ensure the safety and longevity of infrastructure. Accurately quantifying crack geometry, such as length and width, provides essential insights into structural safety and failure risk, particularly for developing thin cracks. However, current segmentation algorithms struggle with distinguishing crack geometry from noisy backgrounds. To address this challenge, a multi-modality RGB-D crack segmentation network, CSA-Net, is proposed to integrates depth features with RGB using a double cross-fusion module. This module enhances segmentation by leveraging RGB texture, color contrast, and edges, alongside depth variations and surface contours, ensuring robust crack localization across various surfaces. CSA-Net also enables precise crack quantification by using depth map geometry to calculate spatial ratios for pixel-to-millimeter measurements. A Crack RGB-D dataset using the Realsense LiDAR Camera L515 is developed to train and evaluate several segmentation networks. Experimental results show CSA-Net outperforms SP-Net (RGB-D) and U-Net (RGB) with a 4.2 % and 4.54 % increase in mIoU, respectively. Furthermore, the method provides highly accurate crack width measurements with an absolute error of less than 0.05 mm compared to ground truth, demonstrating the advantages of the proposed multi-modality information-fusion method in quantifying the thin crack (width less than 0.3 mm) as specified in concrete structure design codes.