Reading #17 Distinguishing Text From Graphics in On-line HandWritten Ink

Comments on:

Summary:

This paper is about distinguishing test from graphics. It proposes three methods: (1) the first one treat each stroke separately. It extracts features from the stroke and use feed-forward neural network to train the weight of each input. (2) The second one adds information of temporal context to improve the performance of the first method. The intuition behind this is that users usually draw or write multiple strokes successively. That means the stroke is more like to be followed by a stroke with the same. An HMM can be constructed to represent the sequence of the stroke. (3) The third method uses the gaps features to train the neural network in the first method and combine it with a bipartite HMM in the second method.


Discussion:

The way of distinguishing test from shape is more complex than the one that use entropy. I think entropy or curvature is more intuitive. However, it is a good idea to use the context information. I hope to find or see more different ways to solve this kind of problem since right now I am satisfy with none of them.