Reading #15 An Image-Based Trainable Symbol Recognizer

Comments on:

JJ

Summary

In this paper, the authors aim to develop a trainable symbol recognizer. Since this recognizer only needs one template to do the template matching, it allows users to easily train the system by adding, removing or overwriting.

In this paper, four methods are used to do the template matching: Hausdorff Distance, Modified Hausodff Distance, Tanimoto Similarity Coefficient and Yule Coefficient. Then the recognizer combines these four methods with the help of parallelization and normalization.

To deal with the problem of rotation, the authors transform the symbol in Cartesian coordinates into polar coordinates. During the process, the recognizer throws away the definition that is obviously dissimilar with to the candidates.

This paper conducts four tests, which differ in the number of definition and whether independent from users. The accuracy is above 90%.

Discussion:

This paper combines four different methods to evaluate the similarity between templates and candidates. By doing this, the recognizer result in a satisfactory accuracy rate. The method of dealing with rotation is different from what is done in $1. Since $1 rotate candidate while here, the templates are rotated to match the candidate.