Multi-scale fine-grained segmentation and feature extraction for complex granite cracks

Objective To address the challenges in segmenting macro-surface crack images induced by water cooling in high-temperature rock masses—namely, the thin and fine-scale nature of cracks, pronounced variations in crack length, and severe class imbalance—this study proposes TLE-Unet, a semantic segmentation network tailored for rock crack segmentation.

Methods Granite specimens were subjected to thermal treatment followed by uniaxial compression tests, and crack images under different temperature conditions were collected. A self-developed LiteEdgeFusion module was then introduced to finely align shallow high-resolution features with deep upsampled features in a scale-aware manner. By integrating edge detection and a channel-attention mechanism, LiteEdgeFusion enhances boundary perception for crack structures. In addition, an auxiliary decoding head, EWSHead, was designed to achieve multi-scale fusion of shallow and mid-level encoder features via edge guidance and lightweight texture enhancement. During training, Tversky loss was incorporated as auxiliary supervision to improve discrimination of subtle cracks and alleviate the adverse effects of background-dominant class imbalance. Overall, while preserving multi-scale semantic representation, the proposed architecture effectively improves segmentation accuracy for fine cracks and enhances boundary continuity. Results and Conclusions Compared with the baseline U-Net, TLE-Unet increases the crack-class IoU from 38.32% to 46.32%, and improves pixel accuracy from 45.34% to 65.70%. Against mainstream segmentation models, TLE-Unet achieves superior crack IoU performance relative to UNet++, Attention U-Net, and DeepLabv3+, demonstrating stronger capability in recognizing fine-scale cracks. Ablation studies confirm the effectiveness of the LiteEdgeFusion module, the EWSHead auxiliary decoder, and the Tversky loss. Heatmap-based visualizations further indicate that TLE-Unet attends more precisely to crack-edge regions. Moreover, geometric parameters extracted from the segmentation results—such as crack length, maximum width, and mean width-exhibit high consistency with manual measurements, validating the practical utility of the proposed method for quantitative crack analysis.