Kernelizing Semantic Similarity Measurement Using Bi-directional Learning Ranking for Cross-Modal Retrieval

Abstract

Aiming at measuring the inner semantic similarities between different modal data, cross-modal retrieval tries to map heterogenous features to a hidden common subspace in which they can be reasonably compared. In this paper, our supposed method considered learning bi-directional ranking examples with five similarity measurements. Particularly, by optimizing our model with kernelized semantic similarity, more complex non-linear correlations between different modalities can be measured in order to learn a discriminative subspace which is more suitable for cross-modal retrieval tasks. By analyzing the experimental results basing on the public dataset, we demonstrate better performance of the proposed method.