HMM Topology Skipping Estimation using Two Stage Training for On-line Thai Handwritten Recognition

K. Siriboon (Thailand)


iterative training, Hidden Markov Model, Optimization Handwritten Recognition


Researchers have extensively applied Hidden Markov Model (HMM) to many applications such as speech and handwritten recognition. Most researchers have been using the left-right topology for handwritten and speech recognition. This research studied the methods to add skipping in to the main topology (left-right-left). The limitation in this problem is the local optimum of the HMM iterative training. Hence, we can't use the fully connected topology and let the iterative training adjust the skipping. This paper proposed a two stages training as a method to avoid local optimum training problem by using the first stage training to train the main topology and use the second stage to train the skipping topology. The two stage training method is compared with fixed skipping, fully skipping and no skipping base on isolated on-line Thai handwritten recognition. The recognition results showed that the proposed training method increases the recognition rate in comparison to the other methods.

Important Links:

Go Back