Font Size: a A A

Research On Hash Indexing Technique Of High-dimensional Data

Posted on:2016-04-07Degree:MasterType:Thesis
Country:ChinaCandidate:Z Z YangFull Text:PDF
GTID:2298330467979050Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
With the rapid development of Internet technology, communication technology, and multimedia technology, high-dimensional multimedia data has shown exponential growth in recent years. How to index these large-scale high-dimensional data has brought new challenges on the tradition indexing techniques and is becoming one of hot research topics in the field of data mining.As effective indexing technique, hash-based indexing method has received considerable attention in recent years. Based on the iterative quantization (ITQ) hashing algorithm, a cluster driven iterative quantization (CITQ) method is proposed to obtain more compact binary hash codebook. Furthermore, we extend the proposed CITQ model to the indexing of multi-view data.1. For the purpose of developing efficient indexing method for large scale of high dimensional data, a cluster-driven iterative quantization (CITQ) model is proposed. In the proposed CITQ model, the cluster constraint reflecting local data structure relation is imposed to keep the consistency of data structure both in original feature space and the learnt binary compact hashing presentation space. Meanwhile, a well-suited extension of CITQ model to out-of-sample for online querying is also proposed. The experimental results on two open dataset validate the effectiveness of the proposed CITQ model;2. For multi-view data, a shared subspace model is proposed to correlate the multiple views of data. We also show it holds close connection to CCA. In the obtained shared subspace, the CITQ model and the shared subspace model are well combined to realize the joint indexing of multi-view data.
Keywords/Search Tags:Information Retrieval, Hash Indexing, Iterative Quantization, Multi-view Data, Shared Subspace
PDF Full Text Request
Related items