Categorization of digital ink elements using spectral features

Abstract

In sketch-based interfaces, the separation of text and graphic elements can be essential when a system has to react to different kinds of input. Even if the interaction with the interface consists in drawing graphic elements, text input may be considered for some purposes, such as annotation, labelling, or input of recognizable text. This work deals with the detection of textual patterns in a set of digital ink elements. The main idea is that text needs a special hand behaviour to be produced, different from the behaviour employed to draw symbols or other graphic elements. Inspired by the models that describe handwriting as a system of coupled oscillations, we believe that the frequencies of these oscillations contain some information about the symbol nature. Therefore, we employ a descriptor that works in the Fourier space. Results show that this representation leads to distinguished patterns for text and graphic elements. The performance of our system is close to the performance one would obtain by using a handwriting recognition engine tuned for this task, while being much faster. Some benefits are also present when both approaches - the proposed and the engine - are combined. © 2008 Springer-Verlag Berlin Heidelberg.