The branch of Information Theory employs a concept called Entropy to compute the randomness/variation in the Signals. This paper directly uses that idea of Entropy to distinguish between Shapes and Texts. This paper is pretty intuitive, at the same time, very innovative in the sense that it makes use of the information that there exists a lot of randomness in texts as compared to Shapes.
Approach:
This paper exploits the fact that generally there is lot more randomness in the shapes than in the text. In order to capture that randomness in any given figure, the authors find out the angle between two consecutive points from a given point. Based on the angle value, they then classify it into one of the pre-defined categories which are given below:
A [0, pi/6)
B [ pi/6, 2*pi/6)
C [2*pi/6, 3*pi/6)
D [3*pi/6, 4*pi/6)
E [4*pi/6, 5*pi/6)
F [5*pi/6, pi)
X End points
For all the points, a sequence of such categories/labels is created and by using a Classifier, it is being determined whether the stroke is a shape or a text or unclassified. For unclassified strokes, the Entropy value is always between the value for a Shape or a Text. Also, this representation of the stroke can be used for Compression and the author also compared that with the GZIP compression.
However, for very Complex Shapes or random strokes, this approach can classify them as texts. For those kinds, the author states that this can be used an one of the features and with other additional features, they can be properly labelled as Shape or Text.
Thoughts:
This is one of the very simple approaches and very intelligently exploits the idea of randomness presence in the texts to distinguish them from Shapes. Moreover, it runs pretty fast.
Future Ideas and Research:
I think the idea of Entropy can be applied to distinguish between various different domains of stroke having different Entropy values. It can even be used to distinguish between characters from two different languages where one has very less Entropy and other has a lot.
No comments:
Post a Comment