Unconstrainted Handwritten Character Recognition Based on Extended Stroke Primitives

초록

This paper addresses a recognition technique for unconstrained handwritten characters based on a novel feature extraction and decision tree classifier. The features are strokes, bays, junctions, tip points, etc. In general, feature extraction of character recognition consists of segmentation and characterization of strokes or stroke primitives. Intrinsic dilemma in feature extraction is that optimal characterization of stroke primitives can only follow optimal segmentation of extended stroke primitives, vice versa. Introducing a new concept of ESP(Extended Stroke Primitive), the proposed technique can avoid the dilemma. ESP is defined as a part of stroke, a stroke itself, or consecutive multiple strokes. Input character image is thinned using a parallel thinning algorithm, and feature nodes are extracted and characterized from the thinned image. ESPs are extracted from the feature nodes and edge pixels, and they are represented as feature vectors. Feature vectors are classified using a decision tree classifier. We have achieved very encouraging experimental results