Open N-Grams and Discriminant Features in Text World: An Empirical Study
Àgata Lapedriza · 2004
Visual word and text recognition is a tractable problem that shares some characteristics with the problem of visual object recognition in cluttered scenes. We use text world as a simple model for testing and developing a new object representation, open n-grams, that can be useful in the later problem. We empirically show that this representation has the same representation power than classical approaches like n-grams, but it is more adapted to problems where clutter and occlusions are common. We also propose a way of selecting the most discriminant features to be used in a classification system.