Comparación de métodos inteligentes 2D-3D para detección y cuantificación de grietas en morteros de cal

Abstract

CONDITIONING OF PROJECT PUBLICATION. The inspection of civil structures demands objective techniques that ensure safety and durability. This project compares the performance of intelligent methods in two-dimensional (2D) and three-dimensional (3D) domains for the detection and quantification of cracks in lime mortars, using a stereo vision system to obtain volumetric information about the damage. During development, pairs of stereoscopic images were acquired and preprocessed, converting formats through high-performance computing. The fine-tuned YOLO11n model was used for 2D segmentation, and dense point clouds were generated using the RAFT algorithm to estimate 3D metrics such as depth and aperture. The results showed that 3D quantification complements 2D analysis by providing relief information, although volumetric metrics depend on the stability of image segmentation. Parallel processing was observed to significantly optimize data conversion times. In conclusion, the 2D approach showed the best overall performance, while 3D analysis is maintained as a complement that provides volumetric context and interpretive support when segmentation and reconstruction conditions allow it. Keywords: Stereo Vision, Deep Learning, Point Cloud, Structural Inspection