Image search reranking with Ranking Linear Discriminant Analysis

Abstract

Feature dimensionality reduction is an important step for data processing, which is used to reduce data's dimensionalities in many areas. In this paper, we apply dimensionality reduction to image search reranking. As a supervised dimensionality reduction method, Linear Discriminant Analysis (LDA) performs well in classification applications, but is not the case for ranking tasks. Firstly, it does not take the relevance degrees into consideration, which is important for ranking problem. Secondly, owing to the supervised nature of LDA, a plenty of labeled samples are required, which are often costly and difficult to obtain. Therefore, based on LDA, we propose an improved method named Ranking Linear Discriminant Analysis (RLDA) by using the relevance degrees as labels. Meanwhile, both labeled and unlabeled samples are utilized so that it is a semi-supervised approach. Experiments are carried out to confirm the good performance of the proposed algorithm.