MDAU-Net: A Multi-Scale U-Net with Dual Attention Module for Pavement Crack Segmentation

Abstract

Crack segmentation is a critical and demanding task in pavement engineering. Recent studies using Convolutional Neural Networks (CNNs) or vision Transformer (ViT) for crack segmentation have shown promising results, but there are still challenges due to the inhomogeneous crack intensity, complex pavement environments, limited labeled training datasets, and noise in most crack images. These challenges make it difficult to distinguish cracks from noise. To address these challenges, we propose MDAU-Net, a multi-scale U-Net with dual attention module based on an encoder-decoder architecture for pavement crack segmentation. MDAU-Net exploits multi-scale input to preserve spatial awareness in all encoder layers, and the dual attention module (DAM) combines global average pooling (GAP) attention module and position attention (PC) module to extract more comprehensive and discriminative features for crack boundary recognition and improve crack segmentation effectiveness. For optimizing the segmentation network and capturing detailed crack information, a hybrid weighted segmentation loss function is developed. Comprehensive experimental evaluations on two benchmark datasets (DeepCrack and Crack500) demonstrate that MDAU-Net achieves state-of-the-art segmentation performance. These results show that MDAU-Net is a promising approach for pavement crack segmentation and can be used to improve the safety and durability of pavement infrastructure.