Research On Semi-Supervised Hashing Methods

With the rapid development of Internet technologies and widespread use of digital multimedia, a large amount of image data springs up, which leads many challenges to image retrieval, such as storage and computational complexity. Because fast query time and low storage are two basic requirements for image retrieval. However, the traditional image retrieval algorithms cannot fully meet above requirements. Hashing-based method for image retrieval draw much attention in recent years, which embeds each image item into hamming space and expresses the image by binary encoding. Therefore, hashing based method have two advantages, i.e., low storage space requirement and fast image index speed.In this paper, by study of classifier, nonparametric Bayesian model, and label propagation, three hashing methods have been put forward to improve the index speed or retrieval precision. The main contributions of this paper are listed as follows. First, we conduct an in-depth analysis on hashing method and propose a classifier-based supervised hashing method according to the character of hashed binary encoding. The classifier-based hashing method encodes the image data by classical classifiers, which converts hashing problem to classification problem, and improves the precision of image retrieval. Second, a nonparametric Bayes-based supervised hashing method is proposed for the problem that the hashing methods cannot describe data distribution exactly. This method models the distribution of the dataset using dirichlet process, and presents the graph model between database and binary encoding, thus achieves better binary encoding to database. Finally, this paper presents an efficient semi-supervised hashing framework based on label propagation for the problem that unsupervised hashing methods obtain low precision and supervised hashing methods need massive labeled data. Meanwhile, traditional supervised hashing methods and the supervised hashing methods proposed in this paper are extended following the semi-supervised hashing framework.The experimental results show that the classifier-based hashing method and the nonparametric Bayes-based supervised hashing method perform better than traditional hashing methods on the precision of image retrieval. And the semi-supervised hashing framework based on label propagation this paper proposed obtains very high precision of image retrieval even under very few labeled data.