Non-Neighboring Rectangular Feature selection using Particle Swarm Optimization

Recently, Viola proposed a rectangular features (RFs) based classifier with high accuracy and rapid processing speed for object detection tasks. In this paper, we propose non-neighboring RFs (NNRFs) as an extension of RFs, and a particle swarm optimization (PSO) based feature selection algorithm for NNRFs. NNRFs are the pairs of arbitrary rectangular sub-regions in images, giving us huge number of candidate NNRFs for feature selection (e.g. 1.3 billion NNRFs in 19×19 pixel image). We show that PSO can select the powerful subset of NNRFs efficiently from the various candidates, and the classification accuracy is improved with the same computational cost as compared with that of Viola´s method.