MS-FPN-Based Pavement Defect Identification Algorithm

False or missed detection may happen in the process of pavement defect detection due to cracks with different shapes and sizes and interference from complex pavement background. To solve this problem, we proposed a Cascade R-CNN detection method based on the MS-Feature Pyramid Network (MS-FPN). First, we introduced a deformable convolution module in the backbone network ResNet101, so that it can adaptively change depending on the pavement defect. Second, we utilized the MS-FPN for the cross-scale bi-directional fusion of feature maps output by the backbone network,in which the multi-branch hybrid dilated convolution (MCE) generates feature maps with multi-scale receptive fields while expanding the receptive field. The dual-channel attention fusion algorithm (ST-A) was used to improve the identification between the background and the object, so that more attention was paid to the location and features of the pavement defect object. The improved Cascade R-CNN can better adapt to the detection of various pavement defects. On the open-source dataset and self-built dataset, the detection precision has improved by 5.02% and 5% respectively compared to that of the original Cascade R-CNN.