Thinsection Image Classification Method Using Deep Learning Of Integrated Multi-dimensional Information

Thinsection image classification is a very important work in geology.Manual classification was the mostly used method in the past,which was blamed for the influence from human factors and the low efficiency.In recent years,with the advent of some artificial intelligence algorithms,such as machine learning,people began to seek for high precision methods for automatic thinsection images classification.In this paper,we analyze the characteristics of thinsection images,and propose a deep learning based classification algorithm that can integrate multi-dimensional information.This method mainly includes three aspects:multi-dimensional information integrating strategy,block synthesis strategy and maximum likelihood results integrating strategy.It can comprehensively utilize the multi-dimensional information of the thinsection image,and take into account the local and global features of the image,as well as combining the results of different image classification models together,so as to obtain higher classification precision of thinsection images.The research work of this paper mainly includes the following four aspects:(1)Through analyzing the features of thinsection images,we found that thinsection images have the features of extinction,color,shape,texture and global combination.In order to comprehensively utilize the features of extinction,color,shape,texture and global combination of thinsection images,a multi-dimensional information integrating strategy is proposed.Firstly,PCA(principal component analysis)is conducted with the single-polarized and orthogonal light images from all angles integrated together,and the first three principal components were used to generate new images.Then the integrated images as well as the original single-polarized and orthogonal polarized images were trained and modeled separately to get the classification models of single polarized and orthogonal light images and the integrated images(2)In order to fully mine the information in local micro features as well as taking into account the global features of the images,a block synthesis strategy is proposed.First,each thinsection image was classified,then fed into the convolutional neural network model.After gaining the prediction results of all block images,we obtain the final classification results by combining them together.(3)In order to classify thinsection images from different aspects,make different classification results supplement each other and improve the classification precision,images to be classified are divided into three categories,namely single polarized light images,orthogonal light images and integrated images,then fed into the models of their kinds respectively.Finally the maximum likelihood method is adopted to integrate the results of the three classification model.(4)A total of 4752 images of 13 types of rocks,including sandstone,basalt,phonolite,andesite,rhyolite,picrite,granite,gabbro,diorite,conglomerate,limestone,peridotite and schist were obtained under single polarizing light and orthogonal light from different rotation angles by using polarizing microscope.The method proposed in this paper is used for the modeling and classification of these 13 types of thinsection images.Results show that the proposed method has achieved good accuracy in classification.