Crack Detection in Low-Resolution Images Using Attention U-Net

Abstract

With the increasing importance of road and structural safety inspections, this study proposes a novel network for effectively detecting non-uniform cracks. Existing crack detection models often suffer from feature loss and performance degradation when learning the complex structures of irregular cracks. To address this issue, this study aims to develop a network that minimizes feature loss and maintains crack detection performance even in low-resolution images by utilizing receptive fields of various sizes. In this study, we design an Attention U-Net network incorporating a Large Receptive Field Block to enhance crack detection accuracy. The proposed network is validated using the Crack500 dataset, a pavement crack dataset, and demonstrates superior performance compared to existing methods, particularly in low-resolution images. Through this research, we aim to improve the reliability of crack detection and contribute to enhancing the efficiency of road and structural maintenance.