A Curve Fitting Approach to Separation of Non-Linearly Separable Pattern Classes, Applied to Chromosome Classification

This paper proposes a new method by which we can arrive at a non-linear decision boundary that exists between two pattern classes that are non-linearly separable. Chromosomal identification is of prime importance to cytogeneticists for diagnosing various abnormalities. The classification of chromosomes using a classifier is generally difficult and inaccurate due to closeness of feature vectors belonging to various chromosome classes. In this paper a novel method to perform chromosomal classification has been attempted and a good classification accuracy of 94% has been achieved. The technique involves sampling of the feature space within an area bounded by the curves of best fit to the two pattern classes and arriving at the optimal boundary point between the two classes in each sampled region. The boundary points are then smoothened to obtain the non-linear decision boundary.