Abstract: Pavement cracks constitute a major potential threat to driving safety.Previous manual detection methods are highly subjective and inefficient.Current computer vision methods have limited applications in crack detection.Existing models have poor generalization capabilities and limited detection effects.To address this problem,a dense network structure of pavement crack detection,called Crack U-Net,is proposed to improve the model generalization capabilities and detection accuracy.Firstly,the dense connection structure of Crack U-Net adopts the network design from the encoder-decoder backbone network U-Net.Similar to the encoder-decoder backbone network,this structure of Crack U-Net is able to improve the utilization of feature information and to enhance the robustness of the model,as well.Secondly,the Crack U-block composed of residual blocks and mini-U is proposed as the basic convolution module of the network,which can extract more abundant crack features compared with the traditional dou-ble-layer convolution layer.Finally,dilated convolution is used in the middle layer of up sampling and down sampling in the network to fully capture the crack features,which is at the edge if the image.Crack U-Net runs on public fracture dataset and produces a series of experimental results.The experimental results show that the AIU value of this method on the dataset is 2.2% higher than the previous method,and it is better than the existing fracture segmentation accuracy and generalization.The experimental results also show that Crack U-Net model can be pruned,and the pruned model is suitable for loading to mobile devices for road crack detection.

Key words: Crack detection, Deep learning, Image segmentation, Road pavement