Line-Members – a Novel Feature in On-Line Whiteboard Note Recognition

Confusion-matrices show that character mix-ups between similar looking letters differing in size rather than in shape (like “s” and “S” or “e” and “l”), as well as between tall letters (such as “M” and “t”) and small case letters (such as “s” and “a”) and vice versa can occur in on-line whiteboard note recognition. This paper introduces a novel feature called “line-member” feature that adds discriminance to the feature vector. Thereby, for certain sample points the script line association is estimated using the Viterbi algorithm and taken as a feature. As our experiments indicate, a relative improvement of r = 3.3 % in character level and r = 3.4 % in word level accuracy compared to a baseline system without the novel “line-member” feature can be achieved. In addition, the character confusion as described above can be reduced.