Hypergraph learning with collaborative representation for image search reranking

Image search reranking has received considerable attention in recent years. It aims at refining the text-based image search results by boosting the rank of relevant images. Hypergraph has been widely used for relevance estimation, where textual results are taken as vertices and the hypergraph ranking is performed to learn their relevance scores. Rather than using the K-nearest neighbor method, recent works have adopted the sparse representation to effectively construct an informative hypergraph. The sparse representation is insensitive to noise and can capture the real neighborhood structure. However, it suffers from a heavy computational cost. Motivated by this observation, in this paper, we leveraged the ridge regression for hypergraph construction. By imposing an ℓ2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ell _2$$\end{document}-regularizer on the size of their regression coefficients, the ridge regression enforces the training samples to collaborate to represent one query. The so-called collaborative representation exhibits more discriminative power and robustness while being computationally efficient. Thereafter, based on the obtained collaborative representation vectors, we measured the pairwise similarities among samples and generated hyperedges. Extensive experiments on the public MediaEval benchmarks demonstrated the effectiveness and superiority of our method over the state-of-the-art reranking methods.