Combining representations for improved sketch recognition

Abstract

Sketching is a common means of conveying, representing, and preserving information, and it has become a subject of research as a method for human-computer interaction, specifically in the area of computer-aided design. Digitally collected sketches contain both spatial and temporal information; additionally, they may contain a conceptual structure of shapes and sub shapes. These multiple aspects suggest several ways of representing sketches, each with advantages and disadvantages for recognition. Most existing sketch recognitions systems are based on a single representation and do not use all available information. We propose combining several representations and systems as a way to improve recognition accuracy. This thesis presents two methods for combining recognition systems. The first improves recognition by improving segmentation, while the second seeks to predict how well systems will recognize a given domain or symbol and combine their outputs accordingly. We show that combining several recognition systems based on different representations can improve the accuracy of existing recognition methods.