Performance Comparison of Equal Width and Equal Frequency Discretization Methods for Author’s Handwriting Recognition

Abstract

Feature’s reformation or representation is a vital process to give impact towards the study of author’s handwriting recognition. This process is aimed to improve the recognition’s performance accuracy. The generation of features by some feature extraction methods usually resulted in a high dimensionality problem that led to some features redundancy and dependency among features that could lower recognition performance. This study implemented the Higher-Order United Moment Invariant (HUMI) feature extraction method to extract the features of the author’s handwriting image. The Equal Width (EW) and Equal Frequency (EF) Discretization methods are proposed to discretize the HUMI features. The EW discretization method reconstructs the HUMI features by reforming each feature with their representation values. These representation values are their unique general features for each bin. While the EF discretization method repositions the HUMI features into their bins based on the number of features frequency. Each bin will have almost the same number of feature values based on the determined range. As a result, the proposed EW discretization method has achieved the highest performance by classifier DTNB that performs at 99.91% for ten (10) cross-validation experiment setup while EF discretization method managed to reach at best 83% performance by classifier Naive Bayes with the same experimental setup. This study aims to compare the performance of both discretization methods and this shows that the EW discretization method has outperformed the EF discretization method. Nevertheless, both discretization methods have performed at the uppermost level and improve the performance of the author’s handwriting recognition.