| Summary: | 碩士 === 國立中正大學 === 資訊工程研究所 === 103 === Nowadays many news TV channels broadcast news 24 hours a day. However,
some news stories are repeated again and again. How to browse news efficiently is an
important problem. In this thesis, we propose a novel news story representation using
Fisher vectors to cluster topic-related news stories together.
Typically, bag of visual word (BoVW) model is applied to represent news stories.
In recent experiments, ones verify that Fisher vector achieves better performance in
image classification and action recognition, which inspires us to apply this approach
to improve news story clustering.
We analyze robustness based on SURF- and MBH-based Fisher vectors.
Through the experimental results, we observe that MBH is more robust than SURF in
describing news stories broadcasted across different channels, which can potentially
improve news story clustering performance.
|
|---|
