Analysis framework for machine learning experiments based on classifier combination for petrographic images

Abstract

The analysis and description of rocks is very well useful in geological industry and also in rock mining. Igneous rock is most abandoned rock in nature. The classification of Igneous rocks in its two types namely Plutonic and Volcanic is a complex and tedious job because of homogeneity between the classes. In geology this classification process has been carried out on the manual basis regularly, requiring high geological expertise. But being done manually can subject to error and leads to unnecessary delay. Hence for automating the process, pattern classification approach is used. In this paper a locally relevant database of igneous rocks involving both the rock types is considered for experimentation and testing. The grain size is considered as a discriminating textural feature for this classification problem followed by identification statistical features for textural discrimination such as Haralick features, and Laws Masks. A Radial basis function support vector machine classifier is used for the classification providing machine learning approach and giving reasonable results. But In Igneous rocks, the image classes are overlapping in the feature space. So the Classification accuracy can be further increased if the hypothesis of the multiple Support vector machine (SVM) image classifiers are combined to give a single hypothesis for the classification of an image. A method is presented for combining different base classifiers i.e. Adaboost technique.