Natural Image Understanding via sparse coding

Traditional methods for Natural Image Understanding can both be computationally expensive and lack robustness. A recently proposed technique for Natural Image Understanding, based on sparse coding, is computationally less expensive and has demonstrated the capability to correctly identify objects from particular types of noisy images. In this paper we examine the ability of this sparse coding technique to handle broader challenges that are likely to be relevant for Natural Image Understanding systems in practice. We find that it remains robust for varied viewing angles, expressions, and illumination. However, identification accuracy suffers when the size of the training database is significantly less than the size of the testing set. We propose a simple technique that could improve the reliability and accuracy of sparse coding based Natural Image Understanding systems.