Abstract: An improved Mask R-CNN highway disease detection algorithm (FAC-Mask R-CNN) is proposed to solve the problem of detection difficulty caused by road surface disease and low background pixel contrast. Based on ResNet101, a shallow feature expression with strong position information is added, and the adjacent feature maps are first fused as the final feature output of the backbone network. At the same time, the CBAM module is introduced to reduce the effect of low contrast between target and background pixels. Deep separable convolution and void convolution are used to replace the common convolution applied in backbone network and effective feature layer output process, respectively, which can improve the computational efficiency and mask prediction accuracy of the model. The average accuracy rate of FAC-Mask R-CNN on RDD(road disease datasets) is 89.86%, the recall rate is 88.54%, and the harmonic mean is 90%, which is 3.09% higher than that of Mask R-CNN algorithm. The results show that FAC-Mask R-CNN can effectively complete the task of fine detection and segmentation of road surface diseases.

Key words: disease detection, instance segmentation, feature fusion, attention mechanism, depth-separable convolution, empty convolution