Pavement crack detection and segmentation using nested U‐Net with residual attention mechanism

Abstract

Abstract As global road maintenance needs grew, automatic technologies for detecting and segmenting pavement cracks were developed. Existing methods faced challenges with background noise interference and the segmentation of fine cracks. This paper proposed the enhanced U‐Net with residual attention (EU‐RA), based on the original U‐Net architecture and inspired by the dense connections in U‐Net++. EU‐RA utilized a pre‐trained ResNet‐152 as the encoder, enhanced feature recognition through a dual attention mechanism, and combined a context module to aggregate multi‐scale information, thereby improving crack detection performance. A mixed loss function optimizes the training process and enhances generalization across different crack types. The decoder integrated multi‐scale feature extraction to capture features of various sizes. In evaluations on the Crack500 and CFD datasets, the F1 scores reached 90% and 96.2%, respectively, outperforming other models. In addition, the EU‐RA model was further evaluated on a self‐created dataset (G242Crack), and the F1 score reached 96.7%. The results indicated that this model performs excellently in pavement crack detection.