Gabor filter based entropy and energy features for basic scene recognition

Our research objective is to develop a supervised learning based hierarchical classification framework built upon Gabor features. Specifically, we experimented on the Oliva Tor alba data-set from the Corel stock photo library. This data set consists of 2688 natural and artificial scene color images, of size (256X256X3) each, from 8 sub-categories. In this paper, we restrict our goal to categorization of images into natural and artificial groups. The methodology consists of feature extraction and binary classification stages. In this, we propose to use the complex Gabor (CG) filter based global features (depending on the overall layout and scene structure, but invariant to object details) from each image. Initially, in the feature extraction process, a 20-CG filter is applied to images for producing Gabor output images in terms of spatial frequency and orientation. Guided by Gabor uncertainty principle, we choose the resolution of spatial frequencies and orientations. At each of these coordinates, energy and entropy features are computed from image's real and imaginary components. We investigate the viability.ofusing the global features with a SVM classifier for basic scene categorization. Next, after applying the PCA based dimensionality reduction, a 2 class SVM discriminant function based on quadratic kernel is applied to the Gabor features for image classification. By using a 10-fold cross validation, we obtained a classification accuracy of 95.79 % and kappa accuracy of 0.9148.