Ground Sampling Distance as a Key Parameter for Automatic Crack Detection in Built Heritage: A Practical Framework With YOLOv5

Abstract. This study presents a practical approach to applying deep learning for the conservation of built heritage, focusing on automatic crack detection in historic masonry using the YOLOv5 object detection model. While most existing research emphasizes model precision under controlled conditions, this work evaluates YOLOv5’s performance in real-world scenarios, accounting for variations in image acquisition conditions. The study contributes a qualitative comparison of deep learning models relevant to automatic surface pathology detection in built heritage and introduces a field-oriented framework to guide experts in selecting and deploying those tools. A key innovation is the investigation of Ground Sampling Distance (GSD), already used in actual inspection methods like photogrammetry, as a critical parameter influencing detection accuracy and model usability. Results show that YOLOv5 can effectively detect both large cracks and microcracks across varied GSD values, and reinforce the value of interdisciplinary practices that combine Deep Learning technologies with established heritage documentation practices.