Abstract:
To address the high computational cost of tunnel-face fracture identification and the difficulty in directly converting seg-mentation results into geometric parameters, a lightweight framework integrating the segmentation network MDC-Net and the hierarchical topology parsing algorithm C-HTP is proposed. MDC-Net employs bottleneck projection and a dual-attention mechanism to enhance the representation of elongated fractures and suppress interference from complex backgrounds. C-HTP uses skeleton topology and collinearity constraints to distinguish main fractures from branches and to extract pix-el-scale fracture length and maximum width, as well as image-plane orientation angle. Experimental results show that MDC-Net contains 1.831 M parameters, requires 2.520 GFLOPs, and achieves a network-only forward inference speed of 245.57 FPS, with mIoU and F1.5-Score reaching 82.83 % and 82.27 %, respectively. Experiments using manually annotated masks verify the standalone geometric quantification capability of C-HTP. In the end-to-end evaluation using MDC-Net-predicted masks, the mean absolute errors for fracture length, maximum width, and image-plane orientation angle are 17.65 pixels, 1.26 pixels, and 5.22°, respectively. The proposed framework enables an end-to-end workflow from tun-nel-face images to pixel-scale geometric parameters while balancing segmentation accuracy, computational efficiency, and geometric quantification capability.