Design of representaiton scheme towards better image search

Recently, Bag-of-Words(BoW) model has been widely used for feature representation in multimedia search area due to its simplification and effectiveness. While various variants of BoW models have been developed to improve the discriminative cabability, less efforts have been paid on discovering the working mechanism underlying them. In this paper, we systematically investigate the impact of various factors in a quantitive manner, and discover how to improve the image search by jointly optimizing these factors. Subsequently, a rational, which states the rule of factor optimization, is proposed on the basis of the experiments on descriptor matching and used to guide the design of better image search systems. To validate the correctness of the proposed rational, a BoW-based image search system, which consists of image database indexing and searching components, is developed for testing. The comprehensive experiments demonstrate that it is beneficial to employ the proposed rational to develop a better image search system.