Abstract: Because mineral identification depends on labor to a large extent, it is of high costs and greatly affected by the subjective influence of mineral-identifying people. In this paper, mineral RGB images and hyperspectral image samples were collected by simulating the way how experts visually identify minerals. With the above samples, the convolutional neural network was trained and the mineral type identification model obtained. The experimental results show that due to the relatively single information contained in the mineral RGB image, it is not enough to distinguish mineral types, the recognition effect is poor, with the recognition accuracy of only approximately 39.52%. However, the hyperspectral images of minerals contain more abundant information and can effectively express the characteristics of mineral types. Therefore, their recognition performance is excellent, with the model recognition accuracy of more than 94.7%, which can meet the actual production needs.