Crack Detection using Pretrained Deep Learning under Varying UAV Imaging Configurations

Abstract

Crack detection is critical for maintaining the structural integrity of infrastructure, particularly in large-scale applications. Recent advances in deep learning have demonstrated high potential for automating crack detection using unmanned aerial vehicle (UAV) imagery. However, most existing approaches are based on 2D image analysis, limiting spatial context and hindering accurate dimensional measurement. This study investigates the performance of a pretrained deep learning crack detection model embedded in Bentley Systems’ iTwin Capture Modeler. UAV images were acquired at varying camera orientations and distances, and detection outputs from 2D images were projected onto a 3D mesh model generated via photogrammetry, enabling spatially referenced analysis. The UAV configurations included vertical, horizontal, and oblique (60°) flight angles. Among the tested scenarios, the slanting configuration at 40m yielded the highest F1 score of 92.28%. The findings highlight the critical influence of camera orientation and image resolution on detection accuracy and demonstrate the applicability of pretrained models in 3D digital twin workflows for structural inspection tasks.