Research And Implementation Of Automatic Classification Technology For Microscopic Rock Images

Classification and identification of microscopic rock images is an inherent and significant task of geology,and is of great important to resource exploration,engineer-ing geology,environmental protection and water conservancy exploration.The classic manual identification is time-consuming and subjective for geologists.Computer aided automatic methods have been developed for rock image classification,and have been proved useful.Sandstones are valuable aquifers and petroleum reservoirs,and classification of microscopic sandstone images is an inherent and significant task for resource explo-ration.However,the variety of geological environments causes the structural differ-ences of sandstones formed in different regions.Current automatic methods seldom consider the problem of automatic cross-region sandstone image classification.In this paper,we make a research on automatic classification technology for mi-croscopic rock images.The main contributions can be highlighted as follows:1.We review the existed methods for computer aided microscopic rock images classification.we firstly introduce the classification process and some problem-s of microscopic rock images classification.Then we make a summary of the existing method for computer aided microscopic rock images classification.Af-ter that,we make a introduction to sandstone classification and the technologies used for sandstone classification in this paper.2.We category all kinds of microscopic rock images,describe features extrac-tion process used in microscopic rock images classification and propose a automatic microscopic rock images identification process based on extract-ed features.We category 2370 microscopic rock images,make a detailed intro-duction to three categories of features(color,texture,grain)and their extraction processes and propose a automatic identification process for microscopic rock images.3.We propose a transfer learning based method Festra for microscopic sand-stone images classification.In this paper,we provide Festra,which uses trans-fer learning to handle the problem of cross-region microscopic sandstone image classification.The method contains two main parts for cross-region sandstone image classification.The first part is feature selection and normalization,which aim to minimize the differences on feature space distributions between the source and target regions.The second part is instance transfer which involves an en-hanced version of Tradaboost.The major enhancements involve both multi-class classification for sandstones,and weighted instance sampling to make full use of the base classifiers.4.We conduct empirical studies to assess the effectiveness of our method.We conduct experiments based on four main research question to evaluating the ef-fectiveness of Festra.Experiments are conducted based on 273 sandstone images taken from four different regions in Tibet.The experimental results have proved effectiveness of Festra,which provides competitive prediction performance on all the four regions.By using the feature selection and E-Tradaboost separately,the results demonstrate the effectiveness of both,and the combination produces the best.