Image Classification using Gabor Filters and Machine Learning

Abstract

Feature extraction and classification are important areas of research in image processing and computer vision with a myriad of applications in science and industry. The focus of this work is on the robust classification of tree and non-tree areas in aerial imagery of the eastern Andes mountains in Peru. Knowledge of this type of information has strong implications in the study of the effect of climate change on the environment and its conservation. Drawing from recent work on human iris pattern identification, we propose a classification methodology based on Gabor feature space representation of aerial imagery, where the two object classes may be well separated. We evaluate two different distance metrics to discern class separation and use the receiver operating characteristic curve to determine an optimum classification threshold. We then build upon our Gabor representation technique by proposing two additional classification methods based on naive Bayes’ and support vector machine classifiers. Mutual information is used for reducing redundant Gabor features not carrying sufficient object information. Extensive experimentation using real aerial imagery of the Peruvian Andes shows that our approach can provide highly accurate classification, even in the presence of variable illumination, different land features and changing topology. The issue of finding an optimal Gabor feature space where object classes are optimally represented is still a challenging problem to be resolved.