Abstract: With the rapid development of society, large infrastructures such as roads, tunnels, and bridges are built, significantly improving people's living standards. But many new or existing infrastructures bring tremendous workload and challenge traditional manual-based security inspection. Therefore, it is of great urgency to automatically monitor infrastructures' health status and map anomalies accordingly. In this paper, we focus on bridges and propose a pavement crack detection method via synergizing low rank representation (LRR) and deep learning techniques to address the problems of low intelligence and insufficient generalization of existing methods. In the first stage, we automatically discriminate most crack frames from the long sequence with a consistent pavement base by LRR. Then we localize the cracks and obtain pixel-wise masks with post-processing based on grayscale and geometry clues in images. In the second stage, we formulate the problem as a semantic segmentation task and propose a network leveraging multi-level features and atrous spatial pyramid pooling (ASPP) for robust performance in various scenes. Finally, we train this network with generated pixel-wise samples from LRR. Extensive experiments on a wide range of pavements demonstrate that our method significantly outperforms many state-of-the-art approaches in terms of both accuracy and automaticity, which can also be adopted in other scenarios.

Key words: crack detection, low rank representation, deep learning, bridge pavement, anomaly mapping, pavement inspection