Pavement crack detection method based on multi-scale dilated spatial attention

Abstract

Various types of cracks on the road are the main reasons that affect the service life of the road structure. However, automatic detection of pavement crack is a challenging industrial problem due to the background noise interference and complex crack structures. In this paper, a novel crack detection method based on multi-scale dilated spatial attention (MSDA) mechanism is proposed for pavement crack detection, named as PCMSDA. First, an MSDA module is designed for feature extraction, which aggregates semantic information to enhance crack features and obtain richer texture details, thereby strengthening the representation of crack features. The MSDA module can effectively preserve the crack features of shallow networks, avoid discontinuities in cracks, and reduce interference of background noise. Meanwhile, in the upsampling process, a content-aware reassembly of features (CARF) module is designed. In the CARF, the input feature maps are utilized to predict the upsampling kernels, where each position employs a different upsampling kernel. Subsequently, feature reassembly is carried out based on the predicted upsampling kernels. While introducing relatively less computational overhead, this approach can aggregate contextual information and effectively preserve the detailed features of cracks, enabling the model to effectively detect pavement cracks. To validate the effectiveness of the proposed PCMSDA, extensive experiments are validated on the real pavement crack dataset. Experimental results demonstrate that the proposed PCMSDA achieves better performance than state-of-the-art pavement crack detection models.