Guest Lecture: Large-scale Pattern Recognition and Its Application
2011-12-03 10:36:42
By: Information Technology Center
Prof. Koichi Kise Department of Computer Science and Intelligent Systems,
Osaka Prefecture University, Japan
Abstract
Pattern recognition is a task to inferring symbols such as category labels from input signals such as images and sounds.
This research area has a long history and many research efforts have been devoted. In this long history, the main stream has been the research on classifier, which takes as input a representation (mostly a multi-dimensional feature vector) to classify it to one of the known categories. In this paradigm, the strategy can be called a single bullet type, which means a well-tuned single feature vector is employed for the classification by a sophisticated classifier.
However, in recent years, with the explosion of the amount of available data, we are now having another paradigm which I call large-scale pattern recognition. The above traditional way of pattern recognition is not easy to be applied, since it cannot deal with a large number of data. In the large-scale pattern recognition, the classifier is kept simple to deal with the large data. Another point is that the object to be recognized is represented by many feature vectors. This means that the way of pattern recognition is a "shot gun" type. Even if the hit rate of each bullet (accuracy of classification) is low, we can hit the target by using many bullets.
In my talk I first introduce the above two paradigms (traditional and new). Next, I will talk about approximate nearest neighbor search, which enables us super efficient classification at the sacrifice of the accuracy. Then some applications based on this paradigm are introduced with demonstrations. The applications include large-scale document image retrieval, real-time character recognition, and large-scale object recognition. The demo of large-scale document image retrieval shows you the real-time retrieval of document images taking as query a camera-captured part of a document. The size of the database of the demo is 1 million pages, while it has been enlarged up to 50 million pages with a server. The demo of character recognition will show you real-time camera-based character recognition and its applications to retrieve services associated to words. The last demo of object recognition is to recognize pictures by matching the database of 5,000 images, which has been scaled up to 1 million images with a server.
Biography
Koichi Kise received the B.E., M.E., and Ph.D. degrees in communication engineering from Osaka University, Osaka, Japan, in 1986, 1988 and 1991, respectively. From 2000 to 2001, he was a visiting professor at German Research Center for Artificial Intelligence (DFKI), Germany. He is now a professor of the Department of Computer Science and Intelligent Systems, Osaka Prefecture University, Japan. He has received awards including best paper awards of three major international conferences in the field of document analysis, i.e., ICDAR (international conf. on document analysis and recognition), DAS (document analysis systems) and ICFHR (international conf. on frontiers in handwriting recognition). He is now serving as a vice chair of IAPR TC11 (reading systems), and a member of IAPR conferences and meetings committee. He has also worked for international conferences including a co-chair of the document analysis track of ICPR2012, and a program co-chair of ICDAR2013. His research interests are in the areas of document analysis, object recognition and information retrieval.
Presentation: download file (14.8MB)
Information
Date: 03 Dec 2011
Time: 9:30AM - 11:30AM
Location: Royal University of Phnom Penh (RUPP) -first Floor - Room (104)
Contact for Registration
Heng Por
- Email: heng.por(at)rupp.edu.kh
- Tel: +855 17 522 520