Professors Koichi Kise and Masakazu Iwamura
Osaka Prefecture University, Osaka, Japan

"Real-time Document Image Retrieval Using Locally Likely Arrangement Hashing"

Thursday, December 4, 2008

4:10 p.m., Packard Lab 466

Refreshments served starting at 3:30 p.m. in the lobby of Packard Lab

The locally likely arrangement hashing (LLAH) algorithm efficiently indexes objects represented as a large number of feature points in the plane.  It is fast enough to enable real-time (e.g. less than 150 ms) retrieval of document images within databases containing more than 10,000 images.  The camera used to capture the query image may be hand-held and low resolution and the image may suffer from perspective distortions and non-linear deformations, and be degraded by uncontrolled lighting. Also, the image may cover only a part of the page and/or may be occluded.  In spite of these challenges, retrieval rates better than 93% have been achieved.  We will give demos, present the theoretical underpinnings of LLAH, and compare it to "geometric hashing," a well-known object recognition algorithm.

(A brief description of LLAH, with demos, software, etc, are available at http://imlab.jp/LLAH/.)