A System and Method for Recognition of Handwritten Telugu Characters

Technology: Method for recognition of handwritten Telegu Character;

Industry: Banking Sector, Security;

Applications: Banking Sector & others;

Market: The global optical character recognition market is projected to grow at a CAGR of 14.8% during 2023-2027.

Generally, Handwritten character recognition can be performed either online or offline.
Based on prior arts survey, a Hidden Markov Model based system is used for online character recognition in Telugu & report a top-l accuracy of 91.6%, but said system operates at symbol level, not at character level.
Further there are other issues including accuracy to identify characters both in online and offline mode.
Hence, there is a need to mitigate above challenges, and present invention provides the solution in efficient manner.

The present invention describes a method & system for recognition of handwritten characters based on Convolutional Neural Networks (CNN).
Each network comprises a first, second, third, & a fourth hidden layers of neurons connected to each other.
The method of recognizing handwritten characters is depicted in the smart chart and figures.
The handwritten characters are scanned into an input image & processed by Convolutional neural networks.
Principal component analysis (PCA) system is to identify the output class to which the characters belong.
The support vector mechanism (SVM) is configured to determine support vectors & identify the output class based on the determined support vectors, & also to train weight stage from the last hidden layer of the CNN to the output layer.
The entire hierarchy of CNN layers from the input to the last hidden layer may then be considered as a kernel layer of the SVM.

Technical Perspective:

Data collected splits into training & testing data wherein 90% data used for training & remaining for testing.
Training data divides into two sets & used to train nine different networks.
A classifier is created by combining the networks with least correlation performance on the test dataset. Further, a new classifier pair yielded higher accuracy than individual classifiers.
Facilitates the feature of supporting the recognition of handwritten not only Telegu character but also support other Indian regional language like Tamil, Kannada & Malayalam.

Industrial Perspective:

User friendly with high accuracy & cost-effective system & method.
Very fast process & consumes less time for recognition of Indian Language character (Telegu).

Prof. Srinivasa Chakravarthy V

Department of Biotechnology

IITM IDF Ref. 1023

IN Patent No: 387357

TRL-3/4

Proof of Concept ready, tested in lab

Technology Transfer