LCJ-Seg: Tunnel Lining Construction Joint Segmentation via Global Perception
Lei Tan, Xiaohan Zhang, Xinrui Zeng, Xiaoxi Hu, Fei Chen, Zongyang Liu, Jin Liu, Tao Tang
Abstract
As railway facilities and equipment maintenance becomes a focal point of current railway research, tunnels, as fundamental components of railway infrastructure, rightly garner attention for intelligent railway maintenance. However, detecting tunnel lining joints poses challenges due to their potential confusion with other internal tunnel objects such as pipelines and cables. To address this issue, we propose an effective segmentation method, LCJ-Seg, leveraging Global Perception Module (GPM) and Detail Preservation and Feature Denoising Module (DPFDM). The GPM is incorporated into the backbone for accurately identifying tunnel lining joints, as these joints require global context to properly distinguish them from other internal tunnel objects. The DPFDM preserves fine details and reduces noise of low-level features to enhance feature fusion. It ensures that both high-level and low-level features retain important details crucial for the lining construction joint segmentation task. Comprehensive comparative experiments and ablation studies on a real collected dataset, demonstrating the superiority of our approach over existing segmentation methods.