Holistic image‐based analysis of damage on concrete surfaces—A multifaceted approach based on supervised machine learning

Abstract Traditional manual methods for inspecting damage on building structures, such as cracks or spalling on concrete surfaces, are laborious, costly, and error‐prone. Despite many attempts to automate this task using digital photographs, most studies primarily focus on detecting damage within images, neglecting the actual dimensions and their implications for structural integrity. To bridge this gap, we present a multifaceted approach for holistic damage analysis that not only detects damage within images but also determines its real‐world dimensions. To achieve this, we first distinguish between linear and areal damage, and apply two separate methods based on deep learning, each tailored to detect these specific types of damage within images. Additionally, we use a cost‐effective 3D‐printed laser projection device to project a grid of laser points onto the surface. This grid, with a known and fixed point‐to‐point distance, serves as a scale reference, facilitating true‐to‐scale measurements of the damage area. Furthermore, for depth estimation of areal damage, we employ models for monocular depth prediction trained in domains distinct from ours. We thoroughly evaluated our methods on realistic and challenging image datasets, which we captured ourselves in public space. The results show that our customized methods for damage detection achieved moderate results for linear damage and more promising results for areal damage. The quantification of damage area resulted in errors less than 10% across all evaluated images, which is suitable for most practical applications. However, estimating the depth of areal damage using models trained on distinct domains proved to be a challenge. Our research expands automated damage detection to include comprehensive, true‐to‐scale analysis of damage and underscores the need for continued refinement.