A lightweight deep learning network based on knowledge distillation for applications of efficient crack segmentation on embedded devices
Jun Chen, Ye Liu, Jia-ao Hou
Abstract
Timely crack detection of pavement helps inspectors access road conditions and determine the maintenance strategy, which can reduce repair costs and safety risks. Deep learning has greatly advanced the development of automated crack detection, but there are still challenges that hinder the application of crack segmentation networks in engineering practice such as the bloated models, the class imbalance problem, and the high-performance device dependency. For efficient crack segmentation tasks, this paper proposes a novel high-performance lightweight network termed multi-path convolution feature fusion lightweight network (MCFF-L Net) and utilize the concept of knowledge distillation. The MCFF-L Net with only 1.18 M parameters achieves F1 score of 85.70% and intersection over union (IoU) of 78.22%, which surpasses the popular heavyweight networks and lightweight networks. The proposed network is further implemented on an embedded device of Jetson Xavier NX and the detection speed of pavement cracks can reach 9.71 frames per second (FPS). The combination of embedded device and proposed lightweight networks is suitable for application scenarios where the portable and efficient crack segmentation is needed and the reliable data transmission through network is not accessible.