A Computationally Efficient HMM-Based Handwriting Verification System

In this paper, we present a novel framework for HMM- based handwriting verification in which the training is performed using a one-shot algorithm for segmentation and HMM parameter estimation using a constrained k-means clustering procedure, instead of the recursive expectation maximization algorithm. This new framework allows training based on a single observation set which results in a straight forward reference model construction and elimination of computationally expensive re-training. Results of a human study using this verification system for handwritten signature and password verification demonstrate that this new efficient approach is still able to maintain high accuracy of 99 % while only three training sets were used.