Tunnel face extraction method for drill-and-blast tunnels considering occlusion and blur degradation

Abstract

Reliable extraction of tunnel-face regions is essential for image-based rock-mass interpretation and intelligent construction control in drill-and-blast tunnels. However, field images are often degraded by machinery and worker occlusions, invalid backgrounds, uneven illumination, dust, and blur, which reduce usable rock-surface information and weaken subsequent visual analysis. This study proposes a tunnel-face extraction framework considering occlusion and blur degradation. A multi-class segmentation dataset was constructed to distinguish tunnel-face, occlusion, and invalid regions, and a theoretical tunnel-face mask was introduced to quantify internal occlusion. Blur degradation and internal occlusion were quantified using a relative blur score and an occlusion ratio, respectively. Under a unified training and evaluation protocol, U-Net, DeepLabV3+, FPN, and SegFormer were compared, with a composite loss function designed to preserve tunnel-face regions and suppress occlusion misclassification. The results show that FPN with a ConvNeXtV2-Tiny backbone achieved the best performance, with Dice coefficient, intersection over union, precision, recall, mean intersection over union, and pixel accuracy values of 0.974, 0.949, 0.970, 0.978, 0.877, and 0.951, respectively. Degradation analysis indicates a stronger performance response to occlusion than to blur, with additional deterioration observed when high blur and occlusion co-occur. Field application in the Yangmeishan Tunnel verifies the practical applicability of the proposed framework for extracting usable tunnel-face regions under complex construction conditions.