Rock thin section image classification in low data scenarios using few-shot learning

Microscopic rock thin section image recognition is crucial in rock mineral analysis. Typically, deep learning models are used to automate expert knowledge, but the scarcity of samples in certain categories limits the available training data, affecting the performance of traditional deep learning models. This paper proposes a novel few-shot learning model to address the challenge of classifying rock thin section images under limited sample conditions. Based on advanced few-shot learning processes involving pre-training and meta-training, we first introduce a Cross Attention Feature Fusion (CAFF) module. This module generates new features by combining plane polarized light images (PPL) and cross-polarized light images (XPL) of rock thin sections under a microscope, integrating these with the original features through autonomous learning to obtain more comprehensive features. Secondly, we propose a Feature Selection (FS) module based on the prototypical network (ProtoNet). This module enhances the model’s classification capability by extracting key feature dimensions from two perspectives: intra-class representativeness and inter-class distinctiveness. Finally, using the pre-trained ResNet50 and Swim-Transformer on ImageNet-1000k as the backbone network, simulation experiments were conducted on the Nanjing University Rock Thin Section Teaching Dataset. Under the 5-Way 5-Shot few-shot learning task standard, the proposed ProtoNet-CAFF-FS model achieved an average classification accuracy of 96.70% and 99.16%, outperforming traditional modeling methods and demonstrating the effectiveness of the newly added modules.