Towards efficient similarity search with semantic hashing techniques

Abstract

Content-based similarity search is a fundamental research problem in multimedia analytics. The explosive growth of multimedia data puts considerable pressure on the efficiency and effectiveness of similarity search of large-scale data, due to data complexity, computational cost and memory consumption. Thus, designing efficient and accurate similarity search engines to provide real-time analysis with economical computation and a minimal memory footprint has become a matter of urgency. As an advanced similarity search technique, hashing, which aims to learn compact hashcodes for high-dimensional data, has drawn substantial attention in the past decades, especially in data management, data storage and computational operations, due to its promising performance in both efficiency and accuracy with acceptable memory consumption. However, several key challenges have not yet been fully addressed when performing hash learning, such as the complexity of binary learning, the semantic consistency during projection, the complex data types, and its applicability for a real-world recommendation. This thesis concentrates on building discriminative hashing models to tackle those challenging issues to achieve efficient, accurate and scalable similarity search in the context of multimedia applications. The first phase of this thesis addresses the unsolved issues of unsupervised graph hashing, i.e., the complexity of graph modelling and unavoidable noises in raw data. This involves a new robust graph learning strategy that separates graph construction and hash learning, studying the complexity of graph modelling, and looking into the unavoidable noise in raw data. I present the Robust Graph Hashing(RGH) model which automatically learns robust graph construction based on self-representation of the samples, to alleviate noise in the data. The second phase investigates the feasibility of supervised hashing for fashion recommendation, while addressing the challenges of inference efficiency, label quality and fashion understanding. A supervised hashing framework called Discrete SupervisedFashion Coordinates Hashing (DSFCH) is introduced, which DSFCH learns meaningful yet high-level features of clothing items represented as hash codes, as a basis for efficient and accurate fashion recommendation. Phase three focuses on improving the multiple semantic preservations during hashcode learning. How to effectively leverage the intrinsic semantic correlations across three domains(i.e. the visual, semantic and hashing spaces) for concurrent knowledge transferring and structure preservation has become a crucial bottleneck in the current hashing models. The proposed InductiveStructure Consistent Hashing (ISCH) framework interactively coordinates the semantic correlations between the visual feature space, the binary class space and the discrete hashing space. The final phase is focused on the multi-view image hashing learning problem. The challenge of this task is that directly projecting multi-view visual features onto the binary label space may lead to semantic inconsistency, due to the large quantisation loss during space transformation. A Flexible DiscreteMulti-view Hashing (FDMH) model is proposed, in conjunction with collective latent feature learning by combining multiple views of data and consistent hash code learning by fusing visual features and flexible semantics. Overall, four discriminative hashing models are developed and applied to a series of challenges similarity search tasks. Extensive experiments on various publicly available benchmark databasesdemonstrate the superior performance of these proposed models over state-of-the-art alternatives.