Symbol Recognintion with Kernel Density Matching

Wan Zhang,Liu Wenyin and Kun Zhang

This paper proposes a new statistical method for symbol recognition. The proposition is similarity measurement of two symbols to satisfy various requests in symbol recognition such as scaling-invariance, etc. A symbol is represented by a 2D joint density estimated from a set of points sampled from the skeleton of the symbol. Matching two symbols is then equivalent to determining whether the two symbols have similar portability distributions. Similarity of two symbols are determined by the Kullback-Leibler (KL) divergence distance. However, this matching is not orientation invariant. So they propose two other methods for determining orientation of the symbol: Angle searching algorithm and independent component analysis (ICA) technique.

For extracting probability distributions Kernel Density Estimator is used (aka Parzen Window Method). And for preprocessing they applied adaptive noise reduction, vectorization of points and uniform length sample points. They also centralized the symbol for zero-mean and applied principal component analysis (PCA) in order to make the symbol representations more robust.

To solve the rotation problem they tried two methods searching and ICA. Searching algorithm was searching the orientation which resulted with least KL divergence which was very time consuming. So they proposed a second ICA method which works faster. ICA's property is that outputs don not change even if some components of the original data are multiplied by a nonzero constant, or data are rotated.

Experiments conducted of proposed method were on GREC 2003 and 2005 datasets. These datasets include rotated and noise added symbols. The results on GREC dataset are impressive and successful. This method might the be the best method for this dataset so far.

Discussion:

Proposed technique is designed to be noise invariant and become successful. However, as my focus is on sketch recognition definition of noise in this paper doesn't relates to the different drawings in sketch recognition as it is not just some random samples and points. For sketch recognition a flexible technique is required rather than noise independent.