Crack-Att Net: crack detection based on improved U-Net with parallel attention

In recent years, the demand for automatic crack detection has increased rapidly. Due to the particularity of crack images, that is, the proportion of cracks in the entire images is very small, and some cracks in the image are particularly slender and light, it brings challenge for automatic crack detection. In this paper, we propose an end-to-end pixel-level crack segmentation network, named as “Crack-Att Net”. In our approach, firstly, an encoder network is used to extract the crack features; then, crack features generated by the encoder and decoder networks at the same scale are pairwisely fused through a parallel attention mechanism added for accurately locating the cracks; finally, the fused crack feature maps at all scales are further fused into a multi-scale feature-fusion map for crack detection. Experiments results on three existing datasets and an augmented dataset show that our proposed Crack-Att Net outperforms the current state-of-the-art methods.