Offline handwriting recognition using Artificial Neural Network and Hidden Markov Model
Cursive handwriting is the most natural way for humans to communicate and record information. The developments of automatic systems that are capable of recognizing human handwritings offer a new way of improving human-computer interface and of enabling computers to perform repetitive tasks of readin...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2002
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/4393/1/TayYongHaurPFKE2002.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Cursive handwriting is the most natural way for humans to communicate and record information. The developments of automatic systems that are capable of recognizing human handwritings offer a new way of improving human-computer interface and of enabling computers to perform repetitive tasks of reading and processing handwritten documents more efficiently. The aim of this thesis is to design an offline handwritten word recognition system based on the hybrid of Artificial Neural Network (ANN) and Hidden Markov Model (HMM). The Input space segmentation (INSEG) approach proposes various ways to segment word into characters. This approach creates the problem of junks - character hypotheses that are not true characters. Two training approaches have been introduced, namely character level discriminant training and word-level discriminant training. The latter shows integration of the ANN and HMM by using the gradient descent algorithm. Different topologies of the ANN have been investigated for modeling of junks. Three isolated word databases, namely, IRONOFF, AWS and SRTP, have been used as the evaluation of the proposed system. Experimental results have shown that the ANN-HMM hybrid with word-level discriminant training consistently yield better recognition accuracy compared to character level discriminant training and discrete HMM-based recognition system. It achieves recognition accuracy of 97.3%, 88.4%, 90.5% and 95.8%, on IRONOFF-1 96, IRONOFF-1 991, SRTP-Cheque, and AWS, respectively. |
|---|