Handwritten Character Recognition Based on Convolution Neural Network Models

Inhalt:
Handwritten character recognition (HCR) aims to recognize the characters written in the paper and translate them to a readable form for the machine. This paper applies four kinds of models in Convolutional Neural Network (CNN) to classify different handwritten English characters from a test dataset. The main focus of this paper is to find out how different CNN models perform on the same dataset. In fact, this is important since the best architecture of models can be identified in the specified dataset. LeNet-5 CNN model as a famous neural network was used in this research with particular changes on the size of kernels in convolutional layer and the number of neurons to fit a dataset called A-Z Handwritten Alphabets. The other three models are designed with the same architecture but have different numbers of convolution and pooling layers to observe how they perform on the same dataset. The testing accuracy of LeNet-5 is 98.18%, while the other three models own 98.57%, 98.70% and 98.76%, respectively.