Objective and Method  Water cooling-induced macroscopic surface cracks in high-temperature rock masses are characterized by fine scales, pronounced length variations, and severe class imbalance in images, posing significant challenges for image segmentation. To address these challenges, this study proposed TLE-UNet, a semantic segmentation network for rock crack images. First, through thermal treatment and uniaxial compression tests of granite specimens, crack images under varying temperatures were acquired. Then, using a self-developed Lite Edge Fusion module, high-resolution shallow features were finely aligned with deep semantic features at varying scales. Furthermore, edge detection and channel-attention mechanisms were combined to enhance crack boundary perception. Additionally, this study designed an auxiliary decoding head module, EWS Head, to achieve multi-scale fusion of features from shallow- and mid-level encoders through edge hint and lightweight texture enhancement. During training, the Tversky loss function was incorporated as auxiliary supervision to improve discrimination of fine cracks and alleviate the adverse effects of background-dominant class imbalance. Generally, the proposed architecture effectively improved the image segmentation accuracy and boundary continuity of fine cracks while preserving multi-scale semantic representation.

Results and Conclusions  Compared to the baseline U-Net, TLE-UNet increased the crack-related IoU from 38.32% to 46.32% and pixel accuracy from 45.34% to 65.70%. In contrast to mainstream segmentation models, TLE-UNet outperformed UNet++, Attention U-Net, and DeepLabv3+ models in crack-related IoU, demonstrating a higher capability to identify fine cracks. Ablation experiments verified the effectiveness of various modules in the TLE-UNet algorithm. Heatmap-based visualization analysis further indicates that TLE-UNet can more precisely highlight crack edges. Finally, based on the image segmentation results and relevant calculation methods, this study derived the geometric parameters of cracks, including their lengths, maximum widths, and mean widths. The analysis of these parameters across multiple images reveals a high consistency between TLE-UNet-derived results and actual measurements, further verifying the effectiveness of the TLE-UNet algorithm in crack formation extraction.