Offline Handwritten Basic Telugu Optical Character Recognition (OCR) using Convolution Neural Networks (CNN)

This paper attempts at the development of an OCR of the handwritten Telugu text comprising the words made up of only the basic characters without matras for simplicity. There are 3 steps in this OCR. The first step is concerned with the segmentation of the basic characters from the handwritten Telugu words. The second step involves placing the characters into anyone of 6 groups. The first four groups are those that satisfy one of the four types of openness criteria (top, bottom, left and right). The 5th group satisfies more than one type of openness and 6^th group the closedness criterion. The third step deals with the recognition of the character in each group for which a deep learning network called Convolutional Neural Network (CNN) is employed. The procedure to be applied on each test character involves finding its group first and then identifying the character. The contributions include the development of the segmentation approach, proposition of openness and closedness criteria, and the use of CNN for the recognition of characters. Out of several network architectures tried, the best performance of 98% accuracy is obtained on the created database of the handwritten basic Telugu characters using the modified CNN model that results from changing different parameters such as Adam optimizer of the original CNN. Similarly changing AlexNet parameters, an accuracy of 90.5%is achieved, and by removing some dense layers in it the accuracy is improved to 95.5% whereas LeNet-S provides the accuracy of 89.5% %. As another study for the sake of comparison, we have extracted the Histogram of Gradients (HOG) features from the same dataset and used Support Vector Machine (SVM) as the classifier and the overall accuracy has come out to be 94.13%.