A Deep-Learning-Based Multiple Defect Detection Method for Tunnel Lining Damages

Abstract

Diagnosing tunnel lining structural damages is vital to ensure safe tunnel operations. However, the detection of multiple defect is challenging task due to the size imbalance between cracks, spalling, and backgrounds. Currently, deep-learning-based methods for multiple defect are dependent on multiple-stage networks, which have limited their scalability and complex frame working processes. To accurately recognize the multiple defect at the pixel-level using only one-stage networks, a new method was proposed, which integrated the basic SegNet with a focal loss function, and was referred to as an FL-SegNet method. The focal loss function was adopted to address the problem of the size imbalance by down-weighing the losses assigned to the well-classified samples, and then the training was focused on the hard samples. Furthermore, comparative experiments were performed to evaluate the performances of the different methods. The experimental results demonstrated that FL-SegNet method was capable of accurately predicting the profiles of small-sized cracks and overlapping damages even under various noise conditions, and successfully outperformed the two-stream method and the basic SegNet method in this regard. The performance metrics (MPA and MIoU) of the FL-SegNet method were significantly higher than those of other multiple defect detection approaches in different scenarios (images with small-sized damages attained to 81.53% and 69.86%, increased by 11.99 % and 4.88% compared with two-stream method, and increased by 17.78% and 7.69% compared with basic SegNet). Therefore, this paper provides an effective solution for the future detection of multiple defect in tunnel linings.