Feature extraction for online handwritten characters using Delaunay triangulation

We introduce a novel feature extraction scheme for online handwritten characters based on utilizing Delaunay triangles for describing each stroke segment. Central to the proposed approach is the idea of associating a unique topological structure with the handwritten shape using the Delaunay triangulation. This allows more ‘meaningful’ groups (i.e., triangles) to be chosen for representing global features, and makes full use of the rich temporal and topological characteristics of handwritten shapes. The Delaunay triangles used for feature extraction, called the Delaunay triangle descriptor, have good discrimination power since they are the only ones satisfying the properties of the Delaunay triangulation. A discrete HMM-based recognition system is used, as the test platform, and shows that the proposed representation can achieve good performance on the chosen data collection, improve recognition accuracy, elevate stability and robustness, and outperform other alternative feature combinations implemented for comparison.