Intelligent Mineral Identification and Classification based on Vision Transformer

Abstract

Image recognition technology in deep learning plays a vital role in many research fields. The intelligent recognition of mineral images based on deep learning technology brings new ideas for the development of traditional fields. It improves the efficiency of identification and brings some economic benefits. This paper proposes an intelligent mineral recognition classification model based on Vision Transformer. Firstly, More than 2000 images of twelve minerals, such as biotite, bornite, phoenix, and quartz, were collected. The data set was expanded by the data enhancement method, which was used to train and test the model, and the model's generalization ability was enhanced. Secondly, a self-attentive mechanism is introduced for feature extraction, and a new activation function is used to optimize the convergence speed of the model further. In the end, the accuracy of this model on the test set Top-1 reached 96.08 %, and the F1 score was 95.40 %. Compared with the network models such as ResNet50, VGG16, and DenseNet, the proposed model's recognition accuracy is higher, and the recognition stability is also better. According to the analysis of the experimental results, the pre-processing of the data also has a particular influence on the accuracy of the model, which provides an essential reference for the subsequent intelligent recognition and classification of minerals.