Automated Pavement Crack Detection Based on Multiscale Fully Convolutional Network

Automatic pavement crack detection is essential for fast and efficient pavement maintenance and health measurement. And crack image data is the basis of crack detection. The existing data collection methods have disadvantages such as high cost, easy loss of frames, blurring, and loss of crack information. Therefore, a new method of data collection using target detection and perspective transformation is introduced. The CRACK2000 dataset with multiple complex backgrounds is constructed by this method. Also, a multiscale fully convolutional network by improving U-Net, named U-multiscale dilated network (U-MDN), is proposed. The network uses U-Net as the base network and introduces U-multiscale dilated convolutional module (U-MDM) after U-Net downsampling. In addition, the U-MDM is compared with U-MCM and MDM, and the result shows that U-MDM has a better effect. Finally, U-MDN is compared with U-Net, CrackSeg, DeeplabV3+, Basnet, and PDDF-Net on CRACK2000 and other data sets, respectively. The experimental results demonstrate that the U-MDN is better than other algorithms in terms of precision, recall, F1-score, and AUC. This paper presents a new data acquisition method based on object detection and perspective transformation. This method is used to construct a CRACK2000 dataset with multiple complex backgrounds. On this basis, a U-Net-based multi-scale fully convolutional network UMDN is proposed. Based on U-Net, the network introduces U-multiscale dilated convolutional module (U-MDM) after subsampling. Experimental results show that U-MDN is superior to other algorithms in precision, recall, F1-score, and AUC.image