Deep transformer networks for precise pothole segmentation tasks
Iason Katsamenis, Athanasios Sakelliou, Nikolaos Bakalos, Eftychios Protopapadakis, Christos Klaridopoulos, Nikolaos Frangakis, Matthaios Bimpas, Dimitris Kalogeras
Abstract
Potholes on the road surface are a significant safety hazard and can cause severe damage to vehicles. Identifying and repairing potholes is a challenging task that requires efficient and accurate methods. In recent years, deep learning models, such as U-Nets and transformers, have been used for image segmentation tasks with promising results. This paper proposes a transformer-based model and in particular the SegFormer framework, for pothole segmentation using high-resolution images captured from a road inspection vehicle. The proposed network outperformed the traditional U-Net model that demonstrates state-of-the-art performance in various segmentation tasks, achieving an average F1-score close to 80%. The results show that the proposed method can effectively identify and localize potholes, providing a useful auxiliary tool for road maintenance and safety.