Saliency Network with Pyramidal Attention for Crack Detection

Abstract

Road crack is a common road disease and can endanger the safety of vehicular traffic. To solve the problem of the low accuracy of traditional deep neural networks in detecting crack images with more complex background and interference, this paper proposes a crack detection network (GLPANet) based on human visual cognitive mechanism. We construct three key modules for extracting crack image features extraction, namely Global Correspondence Modelling (GCM), Local Correspondence Modelling (LCM), and Pyramidal Attention Network (PANet). Specifically, GCM directly fuses all internal features outside through three-dimensional (3D) convolution, LCM also uses 3D convolution to decouple multi-image relationships into multiple local pairs (LP) of image correspondences. PANet learns the spatial geometry of cracks through a network of pyramidal attention mechanisms to create associations between crack features, in the PASCAL VOC dataset, PANet achieved a mean IoU of 77.92%, an 11.5% improvement over FCN-R101. In the crack dataset, with the mean absolute error (MAE) of 0.0172, GLPANet outperforms the state-of-the-art competitors. GLPANet network can improve the accuracy of fracture detection in complex and disturbed backgrounds.