624 #17 Bishop, Svensen - Distinguishing text from graphics in on-line handwritten ink

Introduction

This paper presents three methods for analysing text vs shapes in sketches: a multilayer perceptrion neural network (MLP), a hidden markov model (HMM) and a bi-partite HMM. These methods can be layered on top of each other to get a more complete picture of the stroke type. The MLP is the lowest level and attempts to identify strokes by constructing a feature vector of 9 features for each stroke and running them through an MLP to identify their type. The uni-partite HMM combines the individual stroke knowledge of the MLP and combines it with information about the temporal context of the stroke and those that came before it. The intuition with this HMM is that text strokes will follow text strokes and graphical strokes with follow graphical ones. Finally the bi-partite HMM adds information about the spatial context of strokes, how close they are to preceding strokes. In experiments, they found that the addition of the temporal context helped recognition rates for the MLP, but it was not always the case that the spatial context helped.

Discussion

It is interesting that they compared several layers of recognition in this paper, rather than entirely different techniques. Also of interest is the set of features that they chose for their feature vector, though it is not clear why they chose that particular set (perhaps from their own previous work). The combination of static classifiers and

on-line classifiers was particularly interesting as it showed how dependent this method is on training data.