Text / Non-Text Classification of Strokes using the Composite Descriptor
Martin Bresler · 2013
The task of a single stroke classification into two classes (text and non-text) is the subject of this work. We used an SVM classifier based on a descriptor created as an extension of the existing composite descriptor. It reflects an appearance and a local context of strokes. We achieved over- all accuracy 93.1% on a database of handwritten flowcharts. The state-of-the-arts methods have a quite poor performance on this database with the accuracy 86.3%. Moreover, we showed that our approach allows to learn classifiers favour- ing one class to obtain higher accuracy in classification of strokes in that class while the accuracy in the other class is decreased minimally. This is advantageous for filtering some portion of strokes in the input of specialized recognition en- gines.