A Study of Gaussian Activation Function Based Modular Neural Network for Alternative-Style Handwritten Characters Recognition System

摘要

We propose a design method using Gaussian activation function for alternative-style handwritten character recognition system. While the alternative method can gain high accuracy recognition performance by simplifying the recognition problem to linear discriminant problem, the overfitting problem will occur with small number of learning samples. In this paper, we introduce Gaussian function as activation function of neurons in order to avoid the overfitting problem. Our proposed method can learn the overall distribution of samples and gain higher generalization ability. In the recognition experiment using ETL9B 3036 categories, the proposed method can achieve 97.67% recognition accuracy.