Abstract: To address the challenges of inadequate feature fusion, loss of fine-grained details, background noise interference, and suboptimal computational efficiency in existing rock fracture identification methods, the present study proposes a fracture recognition approach based on the triple feature efficient-fusion U²-Net(TFE-U²-Net) model. This approach is integrated with an image preprocessing algorithm based on gray-based fracture region preselection(GFRP) to significantly enhance the processing efficiency for large-scale, high-resolution images. The proposed improvements include: (1) integrating a triple feature fusion (TFF) mechanism into the U²-Net framework to strengthen multi-scale feature extraction; (2) embedding an efficient channel attention (ECA) module to suppress background noise and enhance feature representation; and (3) replacing standard convolutions with depthwise separable convolutions to significantly reduce model complexity. Based on a self-constructed rock fracture image dataset, ablation studies and comparative experiments were conducted to validate the model’s performance. Experimental results show that the proposed method reduces the model parameter count by 39.2% compared to the original U²-Net, while achieving IoU and F₁ scores of 81.07% and 89.55%, respectively—significantly outperforming baseline models. The TFF mechanism effectively mitigates detail loss, and the ECA module enhances the model's robustness under noisy conditions.

Key words: fracture recognition, feature fusion, U²-Net model, depthwise separable residual block, grayscale features