Modi Script Handwritten Character Classification using Drop-out Induced Incremental Learning Approach

Abstract

Most of genuine information based on healthcare, food habits and Ayurveda is imbibed in Modi script and very few people are remaining who understand Modi script reading and writing. This motivates us to work with Modi script. After going through the deep literature review, we found InceptionV3 and Residual framework (ResNet152) need to experiment on Modi script dataset. The main objective of the system is to develop an incremental learning model which starts identifying from individual characters with acceptable accuracy, which further trained and tested for words followed by sentences, group of sentences. We are the first one to experiment with whole Modi script character set which covers 360 class labels comprising 35 consonants and 10 vowels which was collected using different people. The individual characters are cut, labelled manually and pre-processed using Otsu, Savaula and To-zero binarization techniques. The model is further explored by adding InceptionResNetV2 framework for classification. Initially the model shows overfitting behaviour as discussed in section 5.2. To regularize the model, first experiment is done using augmented dataset which also not able to show satisfactory results and took un-conventional time for training. Further, drop-out approach is induced which shows good hyper-parameter tuning as shown in table 6. Due to drop-out layers, training is extended till 300 epochs to get better results. After training using drop-out technique, the model is showing proportionate increase in training and validation accuracy. At the same time, it is showing considerable consistent decrease in training and validation loss. This gives us intuition that the newly developed hybrid model helps to reduce overfitting and learns appropriately. The developed incremental model is tested on Modi words where all words are mostly classified correctly except the character “sa” which is in-correctly classified as banacha “na” sinch both the characters look similar.