Script recognition

Abstract

This paper describes an approach for word-based on-line and off-line recognition of handwritten cursive script composed of English lower-case letters. The system uses simple and easily extractable features such as the direction of movement and curvature and the relative locations of regions where these suffer discontinuities.

Our approach was evolved based on our concept of ‘shape vectors’ introduced earlier. We visualise script characters as having shapes which are composed of comparatively straight segments alternating with regions of relatively high curvature. We derive the shape vectors from each script character essentially by identifying regions of least curvature and approximating these by straight lines. That these shape vectors carry adequate information about the identity of the character is established by showing that the original character can be faithfully reconstructed from the shape vectors.

We thus use slopes of the shape vectors and relative locations of points of maximum curvature (both highly quantised) as parameters for recognition. The system extracts parameters for individual characters from single specimens written in isolation and uses these to construct feature matrices for words in the vocabulary. These are used for matching with the feature matrices of test words during the recognition phase.

The advantage of the system is that it does not require elaborate training. Recognition scores are in the neighbourhood of 94% for vocabulary sizes of 200 words. The approach has been extended for off-line information as well and performs quite well even in this case.