A Stroke-Based Structural Deformable Model for Handwriting Recognition

Deformable models (DMs) have been receiving a growing interest on recognizing non-rigid objects for its shape-varying capability. However, there are still no mechanism embedded in existing DMs to account for highly structural patterns and most of them indeed can only model a close or open contour. Also, for this kind of elastic matching approach, many underlying algorithms are indeed adopting pixel-to-pixel matching strategy which induces a high computational cost compared with edge matching one. It leads us to study on a new kind of DM called structural DM to model the patterns together with edge matching strategy to define the image force. Then, the elastic matching process is formulated in a Bayesian framework and solved using steepest descent method. A scheme is introduced to ensure the preservation of global structure in terms of relative stroke ordering during minimization, and a multiresolution scheme is also described. The performance of the new model is demonstrated through a small scale Chinese character recognition experiment.