A Deep Learning Method Of Petrographic Thin Section Images For Mineral Identification

Petrographic thin section is commonly used in geological research,in which mineral identification is essentially required.Traditional methods rely on manual identification,which are often costly,time-consuming and subject to expert experience.With the development of computer science and artificial intelligence,the use of computer-related methods to establish automatic mineral identification models on petrographic thin section images has gradually become a research direction.Among them,deep learning is a stateof-the-art approach that has strong capabilities in feature extraction and feature learning,which can better deal with some complex problems on images.This paper proposes a deep learning-based method of petrographic thin section images for mineral identification,which can not only strengthen the boundary information of the images and compensate for the blurred boundary problem of the recognized object caused by the convolutional down-sampling operation,but also take into account the results of different polarized image recognition models,so as to achieve high prediction accuracy.The specific work of this paper is as follows:(1)The paper makes petrographic thin sections of granite samples,collects plane polarized images,cross polarized images and back-scattered electron images,and designs a special data processing method according to the characteristics of these images,so that they can be quickly transformed into the deep learning training dataset with accurate labels and balanced classes.(2)The paper analyzes the common rock-forming minerals in petrographic thin section images,and introduces their color characteristics and geometric characteristics from the perspective of plane polarized images and cross polarized images,after which a deep learning-based fully convolutional neural network that can learn these features comprehensively for mineral identification is proposed.(3)The paper enhances the boundary information in the fully convolutional neural network to improve the blurred boundary problem of the recognized object caused by the convolutional operation during feature extraction.The mineral boundary information is fused into the input layer and result layer of the semantic segmentation network to obtain a more accurate and clearer identification result of the edge segmentation.(4)The paper proposes a fusion strategy of different model results to improve the final accuracy of mineral identification.Through combining the weighted prediction probabilities between pixels,the comprehensive utilization of the two recognition models is realized,and the result with higher accuracy than a single model is obtained.