Learning effective binary representation with deep hashing technique for large-scale multimedia similarity search

Wu, Gengshen and Han, Jungong (2020) Learning effective binary representation with deep hashing technique for large-scale multimedia similarity search. PhD thesis, Lancaster University.

Text (2020gengshenwuphd)
2020gengshenwuphd.pdf - Published Version
Download (9MB)

Abstract

The explosive growth of multimedia data in modern times inspires the research of performing an efficient large-scale multimedia similarity search in the existing information retrieval systems. In the past decades, the hashing-based nearest neighbor search methods draw extensive attention in this research field. By representing the original data with compact hash code, it enables the efficient similarity retrieval by only conducting bitwise operation when computing the Hamming distance. Moreover, less memory space is required to process and store the massive amounts of features for the search engines owing to the nature of compact binary code. These advantages make hashing a competitive option in large-scale visual-related retrieval tasks. Motivated by the previous dedicated works, this thesis focuses on learning compact binary representation via hashing techniques for the large-scale multimedia similarity search tasks. Particularly, several novel frameworks are proposed for popular hashing-based applications like a local binary descriptor for patch-level matching (Chapter 3), video-to-video retrieval (Chapter 4) and cross-modality retrieval (Chapter 5). This thesis starts by addressing the problem of learning local binary descriptor for better patch/image matching performance. To this end, we propose a novel local descriptor termed Unsupervised Deep Binary Descriptor (UDBD) for the patch-level matching tasks, which learns the transformation invariant binary descriptor via embedding the original visual data and their transformed sets into a common Hamming space. By imposing a l2,1-norm regularizer on the objective function, the learned binary descriptor gains robustness against noises. Moreover, a weak bit scheme is applied to address the ambiguous matching in the local binary descriptor, where the best match is determined for each query by comparing a series of weak bits between the query instance and the candidates, thus improving the matching performance. Furthermore, Unsupervised Deep Video Hashing (UDVH) is proposed to facilitate large-scale video-to-video retrieval. To tackle the imbalanced distribution issue in the video feature, balanced rotation is developed to identify a proper projection matrix such that the information of each dimension can be balanced in the fixed-bit quantization, thus improving the retrieval performance dramatically with better code quality. To provide comprehensive insights on the proposed rotation, two different video feature learning structures: stacked LSTM units (UDVH-LSTM) and Temporal Segment Network (UDVH-TSN) are presented in Chapter 4. Lastly, we extend the research topic from single-modality to cross-modality retrieval, where Self-Supervised Deep Multimodal Hashing (SSDMH) based on matrix factorization is proposed to learn unified binary code for different modalities directly without the need for relaxation. By minimizing graph regularization loss, it is prone to produce discriminative hash code via preserving the original data structure. Moreover, Binary Gradient Descent (BGD) accelerates the discrete optimization against the bit-by-bit fashion. Besides, an unsupervised version termed Unsupervised Deep Cross-Modal Hashing (UDCMH) is proposed to tackle the large-scale cross-modality retrieval when prior knowledge is unavailable.