A Pre-Trained CNN based framework for Handwritten Gujarati Digit Classification using Transfer Learning Approach

Abstract

Digit recognition is a software problem to identify numerals of the specific language using computer system. Digit can be printed or handwritten. Handwritten digit recognition is complex task compare to printed because various writing style, thickness and different curve of handwritten digit are difficult to interpret. Numerous work is performed on the native script of India such as Hindi, Bangla, Gurumukhi, and Tamil. However, research efforts on Gujarati Handwritten digit or character recognition are reported very less. This paper aims to demonstrate the efficiency of transfer learning and utilization of a pre-trained model developed for ImageNet dataset to classify handwritten Gujarati digits from zero to nine. The proposed framework developed from the Convolutional and pooling layers of VGG-16, VGG-19, ResNet50, ResNet101, InceptionV3 and EfficentNet pre-trained CNN networks for the feature extraction and newly defined fully-connected layers and output layer for the classification. The proposed framework is investigated on self-created Gujarati Handwritten Digit Dataset. Experimental results show that EfficientNet achieved highest accuracy (training accuracy − 94.9% and testing accuracy − 94.98%) among six pre-trained networks using proposed framework.