Pattern Classification of Handwritten Kannada Digits Using Customized CNN

Abstract

In the area of pattern classification, handwritten digit classification is a challenging problem. Handwritten digits seem different due to writing styles and sizes. There is a wide scope of research on regional languages like Kannada. As Kannada digits symmetric as well as curvy, they are difficult to recognize accurately. Customized convolutional neural network architecture is proposed in this paper for Kannada digits’ classification which are handwritten. This provides automatic learning feature facility of handwritten digit and predicts its class. Layered model of customized CNN architecture is also given which shows flow of working of the proposed method. Large-sized dataset is unavailable for handwritten Kannada digits, so prepared dataset of 20,200 samples. Out of 20,200 total samples, 16,000 samples used for training and the remaining 4200 samples used for testing. The proposed method gives average 93.13% testing accuracy and 91.58% validation accuracy.