Ensembles of Gradient Based Descriptors with Derivative Filters for Visual Object Categorization

Abstract

This paper describes several ensemble methods that combine multiple edge and orientation based histograms with support vector machine classifiers. The aim is to enhance learning speed and accuracy performance by using the chosen classical primitive filters on different edge and orientation descriptors. For efficiently describe images using these descriptors, the combination of a few basis features or edge filters are used. The stronger filter operator responds to edge-like structures, the more sensitive it to orientation. Thus, using more than one edge filter allows to capture more edge information to completely describe the structure of image content. One problem in combining these different descriptors is that the input vector becomes very large dimensionality, which can increase problems of overfitting and hinder generalization performance. The intuitively designed ensemble methods namely product, mean and majority are then used to combine support vector machines classifiers derived from the multiple orientations of edge operators. The results indicate that the ensemble methods outperform the single and naive classifiers.