Development of an Incremental Supervised Learning Model for Recognition of Handwritten Modi Script

Abstract

Handwritten Modi text recognition has a prominent role, among others, in business, healthcare or cultural heritage preservation. Deep learning is a multilayer neural network learning algorithm which emerged in recent years. It has brought a new wave to machine learning and making artificial intelligence and human-computer interaction spread with big strides. India has a cultural heritage where traditions, religions and languages are quite varied. Modi script is one of the oldest written forms of media. Most of the early written knowledge on subjects like medicine, Buddhist ideology, food habits and horoscope has been written using Modi script. MODI is one of the languages that present special challenge to OCR. The main challenge in MODI script is that it is mostly cursive, and few characters look similar. The deep learning methods like InceptionV3 and RestNet architecture seems not experimented yet as per literature review on Modi script. This motivates us to apply the deep learning methods to offline handwritten character recognition using Residual and InceptionV3 framework. newlineThe handwritten Modi barakhadi dataset is prepared by collecting samples from around 25 different people. We are the first one to experiment with whole Modi barakhadi as dataset. The dataset of size 7721 is experimented. To start with, experiment is done on 10 Modi characters using InceptionV3 and ResNet50 frameworks and then explored for whole dataset. The individual characters are cut, labelled and pre-processed using Otsu, Savaula and To-zero binarization techniques. The performance evaluation of pre-processed data is done using three deep learning algorithms ResNet152, InceptionV3 and InceptionResNetV2 on the real-world handwritten character database written by different people. Pre-processed images has given more than 80% precision for all algorithms at 30 epochs for 1786 validation set. We have introduced drop-out technique to reduce the overfitting. The DL model with drop-out is experimented till 300 epochs to get satisfactory result