Precision Recognition of Rock Thin Section Images With Multi‐Head Self‐Attention Convolutional Neural Networks

Abstract Lithological thin‐section image classification is crucial in geology. Traditional manual methods rely on expert experience, being subjective and time‐consuming. Convolutional neural network (CNN)‐based automated classification has potential but is less effective with more rock types and limited training data, restricting its applications. We propose a lightweight framework that integrates the multi‐head self‐attention (MSA) mechanism into classical convolutional neural network (CNN) architectures, and is hereinafter denoted as MSA‐CNN. Specifically, we employ VGG16 and AlexNet as the backbone networks and incorporate the MSA mechanism to enhance the feature extraction from small‐scale lithological thin‐section data sets. The resultant MSA‐VGG16 and MSA‐AlexNet models, after fine‐tuning, can capture crucial geological features more effectively and continuously improve the classification accuracy. We conducted comprehensive experiments on a public data set, which can be partitioned into 3, 34, and 105 categories respectively. The MSA‐VGG16 model exhibits strong generalization ability across all tasks. Notably, in the most challenging scenario with 105 rock categories, the MSA‐VGG16 model outperforms the previously reported best‐performing model on the same data set by approximately 9.61%. These results strongly validate the effectiveness of integrating the MSA mechanism into classical CNNs for lithological classification. They highlight the potential of this method in practical applications and represent a significant advancement in automated geological image classification.