Handwritten Punjabi Character Recognition using Convolutional Neural Networks

Abstract

Today, computers have influenced the life of human beings to a great extent. To provide the communication between computers and users, natural language processing techniques have proven to be very efficient way to exchange the information with less personnel requirement. In this thesis work, natural handwriting technique is used to recognize the online handwritten Punjabi characters as natural handwritten characters are less error prone as compared to the input taken via mouse or keyboard. This thesis describes the implementation of handwritten Punjabi character recognition using deep learning technique named as Convolutional Neural Networks (CNNs). The main problem occurs in the recognition of handwritten characters is due to the occurrence of variation in the handwriting style of different users because each person has their own style of writing and also the variability in the writing style of his/her own style due to change in mood, speed of writing at different instant of time. Punjabi script is chosen for this research work as it comes on 14th position in the spoken languages and less work is done on Punjabi script as compared to work done on other scripts such as English, Devanagari, Gujarati, Chinese. CNN is chosen for the implementation as it is proven to be very efficient technique to recognize and classify the recognized handwritten characters into their respective classes as it concentrates on the dynamic features of the input handwritten character which is obtained from the random generated character matrices. Here, we used 5-layer CNN having stride value of one for the classification of handwritten images into one of the large number of classes (430 classes) available. Punjabi script has total of 430 classes consisting of 35 consonants, 10 vowel identifiers and their corresponding combination characters. In our dataset, each class contains 100 images thereby providing a total of 43,000 number of character images dataset. We divide our dataset in the ratio of 65:25:10, 55:35:10, 45:45:10 training:testing:validation samples data respectively. Training, testing and validation accuracy at different number of epochs (consist of forward pass and backward pass) for these different sample ratios are calculated and thus compared.