Discretization Method to Generalize Features for Authors’ Recognition

Abstract

Features reconstruction or representation to improve the performance accuracy rate is a crucial procedure for handwriting image analysis in the area of authors' recognition. The process of feature extraction usually led to features redundancy with high similarity between features or classes. This will cause the problem of lower performance from classifier models that will have difficulties to differentiate the authors. This study proposes to reconstruct the formulation of discretization method of Equal Width Binning (EWB) to simplify the process. Discretization method is proposed to represent the features and infused the generalization factor into features that are being generated. This is aimed to improve the performance accuracy for authors' recognition. This study deploys the Higher-Order United Moment Invariant (HUMI) and Edge based Directional (ED) feature extraction method to generate Global and Local features respectively from the handwriting images. These generated features were reconstructed by discretization method by representing them with each unique general features. The proposed discretization method has succeeded to improve the average performance of Global Discretized Features that performs at 99.81% and Local Discretized Features achieved up until 99.89%. This shows that the discretization method has managed to improve the features' performance in determining the authors' characteristic.