This web site is no longer maintained and the content may be outdated.
Please visit www.cmpe.boun.edu.tr for up-to-date information.
 
CmpE RSS
No upcoming events...

Home / Graduate / M.S. Theses Completed
 
 
 
 
  Albert Ali Salah, 2000  [download thesis]    

Thesis Title

A Selective Attention Based Method for Optical Handwritten Digit Recognition


Abstract

Primates have a very complex and efficient system for visual recognition. Primates do not absorb the whole information content of the visual field at once, but selectively and serially attend to locations that contain high information. Since parallel recognition requires great computational resources, doing serial recognition is interesting from an engineering point of view. In this work, we propose a serial recognition model for pattern recognition that is based on selective attention.The selective attention mechanism operates at two levels. The primitive, bottom-up attentive level extracts low-resolution features over the whole visual field. The high-resolution fovea is directed around the visual field to gather more information about spots of interest, and the information is combined in the associative level. Between these two levels, we use local experts to increase classification accuracy. We test our model on a well-studied handwritten numeral recognition problem. We compare different saliency schemes in the attentive level, linear vs. multi-layer perceptrons for the local experts, and hidden vs. observable Markov models in the associative level. We simulate a dynamic fovea to control the number of attended spots for classification. Our results show that this model can be applied to more demanding pattern recognition tasks.
 
 
Boğaziçi University Department of Computer Engineering
Address: 34342 Bebek, Istanbul, TURKEY
Phone: +90 212 359 4523-24 Fax: +90 212 287 2461
general information: infocmpe.boun.edu.tr   webmaster: webmastercmpe.boun.edu.tr