Framework for Handwritten Character Recognitionusing Deformable Models

Recently, some deformable models have been proposed for character recognition, due to their ability to capture variations in handwriting. These proposed systems use deformable models to represent characters and to extract features, and subsequently feed the extracted information into a classiier. They often treat the three components { modeling, feature extraction, and clas-siication { as three disjoint and sometimes independent tasks. In this paper, we propose to integrate a deformable model with MacKay's evidence framework 1] as a uniied approach to modeling, feature extraction and classiication, and to apply this framework to handwritten character recognition. Our proposed system begins with tting character models to the raw image and ends at the selection of the most probable output class, all using Bayesian inference. Due to the use of deformable models, the system is invariant to shift, rotation and size changes as well as writing variations. In addition, it does not require user-input parameters (e.g., regularization parameters and character stroke width) and has been empirically shown to achieve a recognition rate of about 95.4% on the NIST database for handwritten digits.