Coherent bag-of audio words model for efficient large-scale video copy detection

Current content-based video copy detection approaches mostly concentrate on the visual cues and neglect the audio information. In this paper, we attempt to tackle the video copy detection task resorting to audio information, which is equivalently important as well as visual information in multimedia...

Full description

Saved in:
Bibliographic Details
Main Authors: LIU, Yang, ZHAO, Wan-Lei, NGO, Chong-wah, XU, Chang-Sheng, LU, Han-Qing
Format: text
Language:English
Published: Institutional Knowledge at Singapore Management University 2010
Subjects:
Online Access:https://ink.library.smu.edu.sg/sis_research/6522
https://ink.library.smu.edu.sg/context/sis_research/article/7525/viewcontent/1816041.1816057.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Singapore Management University
Language: English
Description
Summary:Current content-based video copy detection approaches mostly concentrate on the visual cues and neglect the audio information. In this paper, we attempt to tackle the video copy detection task resorting to audio information, which is equivalently important as well as visual information in multimedia processing. Firstly, inspired by bag-of visual words model, a bag-of audio words (BoA) representation is proposed to characterize each audio frame. Different from naive singlebased modeling audio retrieval approaches, BoA is a highlevel model due to its perceptual and semantical property. Within the BoA model, a coherency vocabulary indexing structure is adopted to achieve more efficient and effective indexing than single vocabulary of standard BoW model. The coherency vocabulary takes advantage of multiple audio features by computing co-occurrence of them across different feature spaces. By enforcing the tight coherency constraint across feature spaces, coherency vocabulary makes the BoA model more discriminative and robust to various audio transforms. 2D Hough transform is then applied to aggregate scores from matched audio segments. The segements fall into the peak bin is identified as the copy segments in reference video. In addition, we also accomplish video copy detection from both audio and visual cues by performing four late fusion strategies to demonstrate complementarity of audio and visual information in video copy detection. Intensive experiments are conducted on the large-scale dataset of TRECVID 2009 and competitve results are achieved.