Self-supervised CRF transformers for tunnel face extraction in complex environments
Xiaoting Zhao, Yulin Ding, Rui Hao, Bokai Duan, Han Hu, Qing Zhu · International Journal of Digital Earth · 2026
Tunnel excavation environments present significant challenges for image segmentation due to inconsistent lighting, diverse rock textures, and frequent occlusions caused by personnel, machinery, and debris. These factors often result in segmentation inaccuracies. Moreover, harsh working conditions limit the collection of high-quality labeled data, and much of the available data is either unreliable or insufficient, thereby reducing the effectiveness of supervised learning approaches. To address these challenges, this paper introduces a robust method for tunnel face extraction leveraging self-supervised learning and CRF transformers. The proposed approach integrates the Swin Transformer architecture with self-supervised pretraining to mitigate the constraints of limited labeled data. Additionally, a Conditional Random Field (CRF) module is used to enhance segmentation continuity and detail recognition. Experimental results demonstrate that self-supervised pretraining substantially improves segmentation accuracy, achieving 88.82\% mIoU, 94.15% Precision, and 94.08% F1-Score using only 10% labeled data. Although the CRF module introduces modest metric improvements, it significantly enhances visual segmentation quality by reducing fragmentation and refining detail precision.