m-CALP - Yet another way of generating handwritten data through evolution for pattern recognition

摘要

Recently a cellular automata learning and prediction (CALP) model was proposed that used evolution to generate handwritten data for pattern recognition. However, the proposed method by CALP was mechanical and not intuitive, as it did not consider the degree of image deformation that takes place during evolution. The problem with this approach is that it evolved handwritten shapes to a completely different form. We propose a new data generation through evolution method called m-CALP, that evolves handwritten shapes using an objective function that also considers the degree of image deformation as it evolves. We replicated the exact same experimental setup of CALP - with the same 5 handwritten data sets, and classifiers with the exact same settings, and then used it to record the performance of our evolved data. We show that data evolved using our method maintains the interesting properties of a pattern. CALP was shown to have performed better than the state-of-theart synthetic over sampling methods such as SMOTE and BORDERLINE-SMOTE, and m CALP performs better than CALP in most cases, in the same environment. Data generation using evolution is a promising field, as it allows researchers to develop trainers and classifier for even small-sized training data. Therefore, we also show that evolving small-sized training data to generate more data using m-CALP performs better than both CALP and larger sized training data.