Mining Repetitive Patterns in Multimedia Data

Mining Repetitive Patterns in Multimedia Data

Junsong Yuan (Northwestern University, USA)
Copyright: © 2009 |Pages: 5
DOI: 10.4018/978-1-60566-010-3.ch200
OnDemand PDF Download:
$30.00
List Price: $37.50

Abstract

One of the focused themes in data mining research is to discover frequent and repetitive patterns from the data. The success of frequent pattern mining (Han, Cheng, Xin, & Yan, 2007) in structured data (e.g., transaction data) and semi-structured data (e.g., text) has recently aroused our curiosity in applying them to multimedia data. Given a collection of unlabeled images, videos or audios, the objective of repetitive pattern discovery is to find (if there is any) similar patterns that appear repetitively in the whole dataset. Discovering such repetitive patterns in multimedia data brings in interesting new problems in data mining research. It also provides opportunities in solving traditional tasks in multimedia research, including visual similarity matching (Boiman & Irani, 2006), visual object retrieval (Sivic & Zisserman, 2004; Philbin, Chum, Isard, Sivic & Zisserman, 2007), categorization (Grauman & Darrell, 2006), recognition (Quack, Ferrari, Leibe & Gool, 2007; Amores, Sebe, & Radeva, 2007), as well as audio object search and indexing (Herley, 2006). • In image mining, frequent or repetitive patterns can be similar image texture regions, a specific visual object, or a category of objects. These repetitive patterns appear in a sub-collection of the images (Hong & Huang, 2004; Tan & Ngo, 2005; Yuan & Wu, 2007, Yuan, Wu & Yang, 2007; Yuan, Li, Fu, Wu & Huang, 2007). • In video mining, repetitive patterns can be repetitive short video clips (e.g. commercials) or temporal visual events that happen frequently in the given videos (Wang, Liu & Yang, 2005; Xie, Kennedy, Chang, Divakaran, Sun, & Lin, 2004; Yang, Xue, & Tian, 2005; Yuan, Wang, Meng, Wu & Li, 2007). • In audio mining, repetitive patterns can be repeated structures appearing in music (Lartillot, 2005) or broadcast audio (Herley, 2006). Repetitive pattern discovery is a challenging problem because we do not have any a prior knowledge of the possible repetitive patterns. For example, it is generally unknown in advance (i) what the repetitive patterns look like (e.g. shape and appearance of the repetitive object/contents of the repetitive clip); (ii) where (location) and how large (scale of the repetitive object or length of the repetitive clip) they are; (iii) how many repetitive patterns in total and how many instances each repetitive pattern has; or even (iv) whether such repetitive patterns exist at all. An exhaustive solution needs to search through all possible pattern sizes and locations, thus is extremely computationally demanding, if not impossible.
Chapter Preview
Top

Introduction

One of the focused themes in data mining research is to discover frequent and repetitive patterns from the data. The success of frequent pattern mining (Han, Cheng, Xin, & Yan, 2007) in structured data (e.g., transaction data) and semi-structured data (e.g., text) has recently aroused our curiosity in applying them to multimedia data. Given a collection of unlabeled images, videos or audios, the objective of repetitive pattern discovery is to find (if there is any) similar patterns that appear repetitively in the whole dataset. Discovering such repetitive patterns in multimedia data brings in interesting new problems in data mining research. It also provides opportunities in solving traditional tasks in multimedia research, including visual similarity matching (Boiman & Irani, 2006), visual object retrieval (Sivic & Zisserman, 2004; Philbin, Chum, Isard, Sivic & Zisserman, 2007), categorization (Grauman & Darrell, 2006), recognition (Quack, Ferrari, Leibe & Gool, 2007; Amores, Sebe, & Radeva, 2007), as well as audio object search and indexing (Herley, 2006).

  • In image mining, frequent or repetitive patterns can be similar image texture regions, a specific visual object, or a category of objects. These repetitive patterns appear in a sub-collection of the images (Hong & Huang, 2004; Tan & Ngo, 2005; Yuan & Wu, 2007, Yuan, Wu & Yang, 2007; Yuan, Li, Fu, Wu & Huang, 2007).

  • In video mining, repetitive patterns can be repetitive short video clips (e.g. commercials) or temporal visual events that happen frequently in the given videos (Wang, Liu & Yang, 2005; Xie, Kennedy, Chang, Divakaran, Sun, & Lin, 2004; Yang, Xue, & Tian, 2005; Yuan, Wang, Meng, Wu & Li, 2007).

  • In audio mining, repetitive patterns can be repeated structures appearing in music (Lartillot, 2005) or broadcast audio (Herley, 2006).

Repetitive pattern discovery is a challenging problem because we do not have any a prior knowledge of the possible repetitive patterns. For example, it is generally unknown in advance (i) what the repetitive patterns look like (e.g. shape and appearance of the repetitive object/contents of the repetitive clip); (ii) where (location) and how large (scale of the repetitive object or length of the repetitive clip) they are; (iii) how many repetitive patterns in total and how many instances each repetitive pattern has; or even (iv) whether such repetitive patterns exist at all. An exhaustive solution needs to search through all possible pattern sizes and locations, thus is extremely computationally demanding, if not impossible.

Top

Background

As a purely unsupervised task, repetitive pattern discovery is different from traditional pattern detection/retrieval problem, where a pattern template or a query example is provided and the task is to find its re-occurrences (Divakara, Peker, Chang, Radhakrishnan, & Xie, 2004). When mining repetitive patterns, we have no a prior knowledge of the pattern. Therefore, it differs from supervised learning problems (e.g. classification and retrieval) that have been widely studied in machine learning research.

Complete Chapter List

Search this Book:
Reset