Visual search reranking with RElevant Local Discriminant Analysis

Abstract

Visual search reranking is a promising technique to refine the text-based image search results with visual information. Dimensionality reduction is one of the key preprocessing steps in it to overcome the “curse of dimensionality” brought by the high-dimensional visual features. However, there are few dimensionality reduction algorithms employing the relevance degree information for visual search reranking. This paper proposes a novel dimensionality reduction algorithm called RElevant Local Discriminant Analysis (RELDA) for visual search reranking. As a semi-supervised combination of improved Linear Discriminant Analysis (LDA) and Locality Preserving Projections (LPP), the proposed RELDA algorithm preserves the local manifold structure of the whole data as well as controls the relevance between labeled examples. Moreover, RELDA algorithm has an analytic form of the globally optimal solution and can be computed based on eigen-decomposition. Extensive experiments on two popular real-world visual search reranking datasets demonstrate the superiority of the proposed RELDA algorithm.