Toward a practical visual object recognition system

Abstract

Recent researches in cognitive science and document recognition have been applied to deal with the problem of categorizing object. Bag-of-Features (BoF) and its extension Spatial Pyramid Matching (SPM) have made a breakthrough in resolving this kind of challenges. Many methods followed this guideline really enhance the recognition accuracy but still have drawbacks in developing a real-world application whose data size is many times bigger. In this paper we propose two kinds of strategy include five criteria to evaluate and select the most appropriate training samples using for building a high performance classifier. We also suggest a method called reinforcement codebook learning to make the codebook training process not only purpose-built to best fits with the most suitable criteria but also much more efficient by reducing significantly its complexity of computation. Experiments on benchmark object dataset demonstrate that our proposed framework outperforms remarkable results and is comparable with the state-of-the-art in spite of using just 20% of 9 · 106 descriptors for training the dictionary. These results give a promise of building a efficient and feasible object categorization system for practical application as so as suggest some ideas to improve the visual feature representation in future.