A Deep Learning Framework for Segmentation of Road Defects Using ResUNet-a

Iason Katsamenis, George Kopsiaftis, Athanasios Voulodimos, Ioannis Rallis, Ioannis Georgoulas, Charalampos Zafeiropoulos, Anastasios Doulamis

Abstract: We present a deep learning framework leveraging the ResUNet-a framework for pixel-wise semantic segmentation of cracks and potholes. By integrating key components including a U-Net encoder/decoder backbone, residual connections, atrous convolutions, pyramid scene parsing pooling, and multi-tasking inference, the proposed method exhibits robustness in capturing intricate spatial details and inter-pixel contextual relationships essential for accurate road defect segmentation. Experimental results validate the efficacy of the proposed approach, with ResUNet-a consistently surpassing the conventional U-Net model, demonstrating its superior performance in crack and pothole segmentation tasks and thus providing a useful auxiliary tool for road maintenance and safety.