Malayalam Handwritten Character Recognition Using Transfer Learning

Abstract

A novel Deep Convolutional Neural Network (DCNN) model is proposed for handwritten Malayalam character recognition using Transfer Learning in this work. Popular Transfer Learning models such as Inception-V4, AlexNet, DenseNet, and VGG are used as a feature extractors. The implementation of popular models like AlexNet, DenseNet-121, DenseNet-201, VGG-11, VGG-16, VGG-19 and Inception-v4 was done with 15 epochs. 99% accuracy was achieved by Inception-V4 with an average epoch time of 16.3 minutes. At the same time, 98% accuracy was achieved by AlexNet with an average training time of 2.2 minutes per epoch, which shows that Inception-V4 performs well. Inception framework that has demonstrated excellent performance at a low computational cost. In this paper, we used residual connections within a traditional Inception architecture, which resulted in state-of-the-art learning performance with the highest accuracy of 99.69% and an average epoch time of 15.1 minutes.