Paper:
Citation:
Bhat, Akshay, and Tracy Hammond. "Using Entropy to Distinguish Shape Versus Text in Hand-Drawn Diagrams." IJCAI. Vol. 9. 2009.
Summary:
Distinguishing between text and shape is a critical task. The approach followed for it in this paper is finding the entropy of each thing and generally the entropy of text is much higher. This is because entropy measure is generally the measure of randomness of an information source. Also, handwritten text is more difficult to describe than common shapes. A zero order entropy is used.
In the entropy model used, each symbol corresponds to a range of angles. The first step in the classification process is to group the strokes. This step uses using the intuition that strokes that are part of the same entity should be close to each other and drawn within a small amount of time. After strokes are grouped, strokes are re-sampled to smooth and sub-stroke angles so that all points are equidistant from each other. Probability estimate is calculated by dividing the total number of occurrence of a symbol the total number of points. Then, all the probabilities within a stroke group are added.
Based on a threshold calculated from training dataset, the input stroke group is classified as either shape, text, or unclassified. Confidence is measured and modeled by using arctan function.
Discussion:
The paper is written in a very easy to read and understand way. The accuracy of the system with maximum accuracy mode is 95% and with more classification mode is 92%, which are pretty impressive results. I think it should be tested with more misclassification symbols also other than a resistor or a dot for better results. Also, some more inputs and results for higher order entropy models would have explained the situation a lot better.
No comments:
Post a Comment