Pavement crack identification and detection based on multi-task learning

Abstract

Existing detection methods for pavement cracks are difficult to handle many different noises and interferences at the same time due to the limitations of hand-designed feature extraction algorithms. Aiming at the above problems, a multi-task learning detection model based on U-Net for road crack detection images is proposed. First, using the well-trained YOLO crack detection model, the crack image is detected, and the crack image in the recognition box is cropped, so the model has a positioning function; Secondly, for the judgment of crack type and other information, a pavement fracture segmentation method based on U-Net model is proposed, and the fracture type is judged and the length, width and area of the crack are calculated according to the segmentation results. And the use of adaptive weight parameters to optimize the loss function value, improve model performance, with stronger robustness and generalization ability. The experimental results show that the error rates of the proposed pavement crack detection model for transverse cracks, longitudinal cracks and alligator cracks are 0.02, 0.08 and 0.07, respectively, and the method can not only accurately locate, but also read out the crack type and geometric parameter information, which effectively solves the actual needs.