Damage detection in concrete structures with multi-feature backgrounds using the YOLO network family

Abstract

Image processing and Convolution Neural Networks (CNN) are widely used for structural damage assessment. Datasets with damages on similar backgrounds are commonly used in past studies for training and testing of CNN models. These models will often fail to detect damage in images of real infrastructure. A dataset is created which consists of 3750 real images along with its annotations, having diverse features with varying textures, colours, and architectural elements like windows and doors. This study evaluates the performance of You Only Look Once (YOLO) models (v3-v10) on the created dataset, training them in three distinct scenarios: scenario 1 (instances of damage ≤5), scenario 2 (instances of damage >5), and scenario 3 (the complete dataset). The YOLO models show promising results in detecting and locating damages in images with multi-featured backgrounds, wherein the YOLOv4 showed the best precision of 92.2 %, a recall of 86.8 %, and an F1 score of 88.9 %.