Publication Type : Journal Article
Publisher : International Journal of Engineering and Technology(UAE)
Source : International Journal of Engineering and Technology (UAE), Science Publishing Corporation, Volume 7, Issue 3, p.1599-1604 (2018)
Url : https://www.sciencepubco.com/index.php/ijet/article/view/14053
Keywords : Ancient Documents, Deep neural networks, Degraded Character Recognition, Handwritten Text, Kannada Documents, printed text, South Indian Script.
Campus : Mysuru
School : School of Arts and Sciences
Department : Computer Science
Year : 2018
Abstract : Realization of high accuracies towards south Indian character recognition is one the truly interesting research challenge. In this paper, our investigation is focused on recognition of one of the most widely used south Indian script called Kannada. In particular, the proposed exper-iment is subject towards the recognition of degraded character images which are extracted from the ancient Kannada poetry documents and also on the handwritten character images that are collected from various unconstrained environments. The character images in the degraded documents are slightly blurry as a result of which character image is imposed by a kind of broken and messy appearances, this particular aspect leads to various conflicting behaviors of the recognition algorithm which in turn reduces the accuracy of recognition. The training of degraded patterns of character image samples are carried out by using one of the deep convolution neural networks known as Alex net.The performance evaluation of this experimentation is subject towards the handwritten datasets gathered synthetically from users of age groups between 18-21, 22-25 and 26-30 and also printed datasets which are extracted from ancient document images of Kannada poetry/literature. The datasets are comprised of around 497 classes. 428 classes include consonants, vowels, simple compound characters and complex com-pound characters. Each base character combined with consonant/vowel modifiers in handwritten text with overlapping/touching diacritics are assumed as a separate class in Kannada script for our experimentation. However, for those compound characters that are non-overlapping/touching are still considered as individual classes for which the semantic analysis is carried out during the post processing stage of OCR. It is observed that the performance of the Alex net in classification of printed character samples is reported as 91.3% and with reference to handwritten text, and accuracy of 92% is recorded.
Cite this Research Publication : Shobha Rani N., Chandan, N., Sajan Jain, and Hena Kiran, “Deformed character recognition using convolutional neural networks”, International Journal of Engineering and Technology (UAE), vol. 7, no. 3, pp. 1599-1604, 2018.