Arun Ravindran

New Features and New Feature Transformations for HMM based script independent OCR System

Date: 07/01/05

Abstract:

The goal of this effort is to explore and characterize the usefulness of some new features and feature transforms for the BBN Byblos OCR system. New computationally cheap “grid” features extracted from the image and “feed back” features derived from the conventional features and training models were added to the feature set and their performance compared with the baseline system. New feature transformation techniques were tested to reduce the dimensionality of the feature space without losing discriminant information. BBN Byblos OCR system uses Linear Discriminant Analysis (LDA) to reduce the dimensionality of the feature space. This work is aimed at investigating the usefulness of a generalized LDA transform, Heteroscedastic Discriminant Analysis (HLDA) transform that removes the equal within-class covariance constraint in LDA. Another transformation that was developed as part of this thesis is the Exponential Feature Transformation. This is a novel nonlinear transformation that could be extended to a wide variety pattern recognition tasks. The new features and new transformations developed in this thesis were tested on the DARPA Arabic corpus and English corpus obtained from broadcast news video sequences. Though the new features and HLDA transformation helped improve the performance on the English corpus, the same was not the case with the Arabic corpus.

Committee:

Dr. John Makhoul (advisor)
Prof. Jennifer Dy
Mr. Prem Natarajan, BBN Technologies, Cambridge MA (external)