Mobile image retrieval using multi-photos as query

A novel image retrieval scheme, where multi relevant images are input as queries to improve the retrieval performance and exploit sufficient information provided by multi query images to reduce distractor features, quantization loss and learn visual synonyms.

Abstract

In this paper, we propose a novel image retrieval scheme, where multi relevant images are input as queries to improve the retrieval performance. We exploit sufficient information provided by multi query images to reduce distractor features, quantization loss and learn visual synonyms. During learning synonyms, consisting of visual synonyms detection and visual synonyms expansion, some identical and unique details semantically important to the query are captured. We represent images using a set of visual synonyms, each of which comprises several visual word paths, quantizing a descriptor from the root to a leaf of a hierarchical vocabulary tree. Spatial layout is also introduced for geometry constraint as an information source independent from descriptor space. Hierarchical visual word path and synonyms learning provide multiple choices for feature matching. Finally we evaluate our approach on two image datasets, where images from 5K Oxford building dataset are used as query; a 227K image dataset act as distractor.