Semi-supervised spectral hashing for fast similarity search

• Published in: Neurocomputing (Amsterdam) Vol. 101; pp. 52 - 58
• Main Authors: Yao, Chengwei, Bu, Jiajun, Wu, Chenxia, Chen, Gencai
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 04.02.2013; Elsevier
• Subjects: Algorithmics. Computability. Computer arithmetics; Applied sciences; Approximate nearest neighbor search; Approximation; Artificial intelligence; Benchmarking; Binary codes; Computer science; control theory; systems; Data processing. List processing. Character string processing; Dimensionality reduction; Embedding learning; Exact sciences and technology; Hashing; Information systems. Data bases; Mathematical models; Memory organisation. Data processing; Pattern recognition. Digital image processing. Computational geometry; Searching; Similarity; Software; Spectra; Tasks; Theoretical computing; Dimensionality reduction; Hashing; Embedding learning; Approximate nearest neighbor search; Nearest neighbour; Computer vision; Euclidean geometry; Hamming distance; Search key; String matching; Binary data; Information retrieval; Orthogonality; Laplacian; Dimension reduction; Supervised learning; Large scale
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2012.06.035

Fast similarity search has been a key step in many large-scale computer vision and information retrieval tasks. Recently, there are a surge of research interests on the hashing-based techniques to allow approximate but highly efficient similarity search. Most existing hashing methods are unsupervised, which demonstrate the promising performance using the information of unlabeled data to generate binary codes. In this paper, we propose a novel semi-supervised hashing method to take into account the pairwise supervised information including must-link and cannot-link, and then maximize the information provided by each bit according to both the labeled data and the unlabeled data. Different from previous works on semi-supervised hashing, we use the square of the Euclidean distance to measure the Hamming distance, which leads to a more general Laplacian matrix based solution after the relaxation by removing the binary constraints. We also relax the orthogonality constraints to reduce the error when converting the real-value solution to the binary one. The experimental evaluations on three benchmark datasets show the superior performance of the proposed method over the state-of-the-art approaches.