Shift Invariant Support Vector Machine for Image Classification in Automatic Target Recognition Systems

Abstract

In this paper, we present a new method that combines the concepts of distance classifier correlation filters (DCCF) and support vector machines (SVM) to enable a new shift-invariant classification algorithm. A DCCF-based kernel function is developed to use with the SVM classifier for image classification. We demonstrate that the proposed kernel satisfies Mercer’s condition, and thus a viable SVM kernel. Our proposed algorithm is shift invariant and exhibits high discrimination when tested on moving and stationary target acquisition and recognition (MSTAR) datasets, a standard benchmarking resource for ATR algorithms. Our proposed solution outperformed two state-of-the-art shift-invariant algorithms: Unconstrained maximum average correlation energy(UMACE) and Optimal tradeoff synthetic discriminant function(OTSDF). Furthermore, our results indicate that the proposed algorithm outperforms the SVM-Gaussian when relatively small datasets are available.