Comparison of Reference-set Selection Methods for Reference-based Scheme.

Reference-based representation scheme achieves significant reduced dimensionality and better discriminability compared to using original image features. More specifically, a greater degree of diversity among the reference-set increases the probability that the descriptor of different images will be unique and well-spread in the data space while maintaining low redundancy. To ensure good diversity of reference-set, we study and compare five different methods to automatically construct the reference-set, including max-variation scheme, max-mean scheme, max-correlation scheme, min-correlation scheme, and affinity propagation algorithm to the task of scene categorization on some benchmark datasets. Extensive experimental results demonstrate that the proposed reference-set selection methods improves reference-based scene categorization method and outperforms state-of-the-art methods. [ABSTRACT FROM AUTHOR]