Efficient probabilistic image retrieval based on a mixed feature model

Abstract

In content-based image retrieval (CBIR) methods, color, texture, and shape are extracted from an image as low-level visual features for searching. However, these visual features could not express image sentiment concepts required in applications like garment search. To address this issue, we propose an efficient image retrieval method based on combined visual and conceptual feature spaces. A unified similarity distance between two images is obtained by linearly concatenating similarity measures. To further improve the retrieval efficiency, we propose an efficient probabilistic indexing scheme called M ixed- F eature- b ased P robabilistic Tree(MFP-Tree) to facilitate the retrieval over a large image repository. Different from conventional image retrieval and indexing methods, which only adopt visual similarity as a query metric, our proposed retrieval algorithm allows users to choose among the above three kinds of features as query elements. Moreover, a probabilistic model is introduced to refine the retrieval result with confidence guarantee. Comprehensive experiments have testified the effectiveness and efficiency of our proposed retrieval and indexing method. © 2012 Springer-Verlag Berlin Heidelberg.