Distance-based Loss Function for Machine Learning in Image Segmentation of Concrete Cracks

Crack detection is essential for maintaining the integrity and safety of infrastructure such as bridges, roads, and buildings. Traditional manual inspection methods are labor-intensive and error-prone, which underscores the need for automated solutions. Advances in machine learning have shown promise in automating this task, but further improvements are necessary to address specific challenges, such as the thin and elongated nature of cracks, varying lighting conditions, and background noise. This paper presents a refined loss function designed to enhance the performance of machine learning models in detecting fine-structured concrete cracks. This loss function considers the spatial distances between predicted and actual crack locations, penalizing false positives that are far from true crack positions. In this way, our approach emphasizes the thin nature and continuity of cracks, ensuring that even the smallest features are accurately detected. Our initial experiments demonstrate that this strategy offers a more accurate and reliable solution to crack detection.