Abstract

Visual re ranking is a method introduced mainly to refine text-based image search results. It utilizes visual information of an image to find the “true” ranking list from the noisy one done by the search based on texts. The process uses both textual and visual information. In this paper, textual and visual information is modeled from the probabilistic perspective visual reranking is in the Bayesian framework, thereby named as Bayesian visual reranking. In this method, the text based information is taken as likelihood, to find the preference strength between re ranked results and text-based search results which is the ranking distance. The visual information of an image is taken as the conditional prior, to indicate the ranking score consistency between the visually similar samples. This process maximizes visual consistency and minimizes the ranking distance. For finding the ranking distance, three ranking distance methods are use . Three different regularizers are studied to find the best results. Extensive experiments are done on text based image search datasets and Bayesian visual reranking proved to be effective.

Authors and Affiliations