Reading #17. Distinguishing Text from Graphics in On-line Handwritten Ink(Bishop)

COMMENTS:

SUMMARY:

This paper presents a method for distinguishing strokes of digital ink as either text or graphics.  In the approach, the authors consider both the features of the strokes as well as the context.  Instead of trying to make hard classifications, the recognizer provides probabilities that the strokes belong to one group or the other.  This improves performance and provides information that can be passed on to more precise recognition processes.

The authors use 3 approaches:

·                     the features of each stroke are extracted and classified

·                     temporal information is used to find correlation between class labels

·                     information is extracted from the gaps between successive strokes

Testing showed that using the temporal context improved performance compared to the classification of individual strokes; the effects of using gap information was not clear.

DISCUSSION:

This is basically a paper on how to quickly distinguish text from graphic shapes.  The recognizer uses the features of strokes which is quite common, but the novelty of this approach is that it also uses the features of the gaps between the strokes.  This method is based on the assumption that strokes that belong to text and strokes that belong to graphics tend to follow other strokes of the same kind.