Novel script line identification method for script normalization and feature extraction in on-line handwritten whiteboard note recognition
When writing on a whiteboard, the writer stands rather than sits and the writing arm does not rest. Due to these adverse conditions when writing on a whiteboard, the script lines within the handwritten text suffer from high variations, i.e. they cannot be approximated by polynomials of low order. In this paper, we propose a novel method for identifying script lines in handwritten whiteboard notes by assigning the sample points of the script trajectory to the script lines. The optimal assignment is then found by the Viterbi algorithm. We present two ways to use the script line characterization. First, the script lines are used to normalize the skew and size of the text lines. In a second approach, the feature vector of a standard recognition system is augmented by the explicit script line membership of each sample point, aiming at reducing the confusions between characters differing in size rather than in shape (like “s” and “S” or “e” and “l”). As experiments show, a relative improvement of r =3.3% in character-level and r =3.4% in word-level accuracy compared to a baseline system can be achieved with the proposed script line identification method. In addition, the written character confusion as described above can be reduced. Finally, the proposed utilizations are examined and discussed in further detail.