Fracture Extraction in Thin Sections with Limited Samples Using Attention Mechanisms

To address the challenges of complex micro-fracture morphology, scarce annotated data, and susceptibility to mineral and noise interference in thin-section images, this study proposes a high-precision, few-shot intelligent fracture extraction method based on attention mechanisms. A data augmentation strategy combining multi-scale slicing and random RGB channel permutation was designed to enhance the model’s generalization capability for diverse fracture morphologies. Additionally, an attention mechanism module was embedded into U-Net to strengthen fracture edge feature extraction while suppressing interference from mineral boundaries and micropores. Experimental results demonstrate that with only three annotated samples, the proposed model achieves a 25.5% improvement in F1-score and a 39.2% increase in Intersection over Union (IoU) compared to the original U-Net, significantly reducing false positives in pore- and mineral-interference regions. This study overcomes the limitations of existing methods in small-sample scenarios, offering a novel solution for efficient and intelligent fracture identification in thin-section analysis.