Semantic segmentation of pavement cracks based on an improved U-Net

Abstract

Cracks are one of the types of pavement defects that can affect the safety and quality of roads, so identifying such defects is an important part of road maintenance. In this paper, based on the U-Net coding-decoding structure, an ECA channel attention module is added to the coding stage, thus improving feature extraction without introducing too many extra parameters and computational effort. The insertion of the FCNhead decoding dock in the decoding stage can improve the performance of the model on semantic segmentation tasks while maintaining its efficiency and interpretability, thus better meeting the needs of practical applications. Experimental results on the CFD dataset and the crack500 dataset show that the algorithm improves the accuracy of crack detection and has better robustness.