DefinePK

Abstract

This paper presents a pilot study on Khmer handwritten symbols recognition using Convolutional Neural Networks (CNNs). The motivation for this study is to develop a recognition system for digitizing large corpora of Khmer handwritten documents. Image data consists of six handwriting sample sets, each of which consists of 33 consonants (root radicals) and 17 vowels, total of 561 syllables. A CNN-based model was trained for offline recognition of root radicals where one CNN was trained for recognition of a particular consonant. All 33 networks have been combined into an assembly. The recognition results are compared against artificial neural network (ANN)-based classifier with full feature set and ANN -based classifier with dimensionality reduction. Feature correlation two-dimensional Fourier transformation (FT2D) and Gabor filters are used for dimensionality reduction. Recognition rate of Khmer handwriting (alphasyllabary system) is increased to 94.85% with Convolutional Neural Networks (CNN).