Online Handwritten Amharic Word Recognition Using Fisher Discriminant Analysis and Hidden Markov Model
No Thumbnail Available
Date
2014-10
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Technology advancement has enabled human being to use electronic devices for recognizing and processing human languages. Amharic, which is the working language of Ethiopian government and which has its own script, is also encoded into computers with available computer keyboards. The purpose of this research is to develop an online handwritten Amharic word recognition system which allows using handheld devices to engrave Amharic scripts.
In this thesis, a writer independent, online Amharic word recognition is presented along with different tests for character recognition. The underlying principle for word recognition is that a word is comprised of characters. Hence by segmenting a given word into character blocks and by using character recognition engine, a given input sequence for Amharic word can be predicted. Finally, hypothesis filtering will limit the number of words hypothesized. As part of character recognition, three approaches were adopted and tested. The first one using Fishers Linear Discriminant Analysis (FDA) to discriminate vectors. The second approach is to extract features from a given input sequence using a predefined set of primitives using HMM model. And the third approach is by scanning the input sequence horizontally, vertically and hybrid of the two scanning. By taking those points into vector and by using FDA for vector classification, discriminate the characters. For training and testing of characters, data from 108 users, 264 character from each user, were used. Likewise data from 34 users, where each user wrote 200 words, is used for word recognition purpose. The result for the character recognizer diminishes as the number of character increases for the first case. For the case of HMM the character recognizer engine predicted an average of 3.94 %. Using the scanning approach, first a vector of 300 length is used and resulted in an average 40.51%, 44.41% for vertical scanning and 63.11% for the hybrid. However, when the vector size is reduced to 70 to increase operation performance, the result is impacted accordingly to 25.66% for the horizontal scanning, 18.77% for the vertical scanning and 39.85% for the hybrid approach. Word recognition using the hybrid approach resulted in 37.9% recognition performance.
Keywords: Online Handwriting Recognition, Amharic Handwriting Recognition, Fisher Discriminant Analysis, Hidden Markov Model
Description
Keywords
Online Handwriting Recognition; Amharic Handwriting Recognition; Fisher Discriminant Analysis; Hidden Markov Model