Image search re-ranking with an improved VisualRank algorithm and multi-layer DCNN features

Ai Wei, Xinmei Tian

Abstract: Image search re-ranking is a post process for image retrieval to further improve the accuracy by investigating image's visual information. VisualRank, as a classical method, re-ranks images based on their popularities and improves the retrieval performance effectively. Many re-ranking algorithms have been proposed based on VisualRank and treat it as a denoising process. However the power of Visualrank has not been fully developed yet. In this paper we thoroughly discuss the mechanism of each part of VisualRank and design an effective approach based on it to fully develop its strength and improve retrieval performance significantly. Besides, local and holistic features from deep convolution neural networks are adopted in our algorithm. We evaluate their respective merits and make them cooperate in the designed framework. Experiments are conducted on the INRIA web353 dataset and the results demonstrate that our approach achieves a competitive performance with state-of-the-arts.