Web Media Integrated Retrieval Based On Multiple Feature Fusion

With the exponential growth of Internet development in21st century, all things are digitalized as data. The dramatic development brings not only the sharp expansion of data scale, but also the complexity of data structure, which tends from the original single text modality to the fusion of multiple modalities. Traditional information retrieval methods cannot meet rapidly changing demand. How to deal with large scale data effectively and how to get useful information from multi-modal data accurately and efficiently have become an urgent problem.This paper starts with the introduction of traditional information retrieval technology and then explores the issues with the development and present situation of multi-modal multimedia retrieval technology in detail. With these widely used technologies, including analysis methods, indexing methods, fusion algorithms, result-merge algorithms on multimedia, this paper proposes an integrated retrieval framework based on multiple feature late fusion. The framework was realized and tested in the D-Ocean project (No.2010ZX01042-002-003) and achieves better effect in experiment. In the framework, this paper presents ηTA result-merge algorithm to improve retrieval efficiency for large scale data. In addition, for multi-modal social media data, this paper presents a two-stage clustering multiple feature early fusion method which takes media characteristics into consideration. The two-stage clustering method gets higher quality compared with direct CCA fusion algorithm in experiment. At the same time, this method speeds up retrieval rate using a unified Lucene index.