Handwritten Character Recognition Using KNN and SVM Based Classifier over Feature Vector from Autoencoder

Abstract

Optical character recognition system is a necessity for the field of man-machine interaction. Handwritten character recognition is a subset of OCR technique by which computer classifies the handwritten alphabets as well as digits. In this work, we present four methods using a vanilla Autoencoder and a Convolutional Autoencoder. For classification purpose we have used KNN, SVM based classifiers such as hybrid KNN-SVM and ν-SVM. We evaluated our proposed models on different handwritten scripts such as EMNIST, Devanagari Handwritten Character, and Kannada-MNIST. Autoencoders are generally used to reduce the dimension of the dataset in a non-linear manner and hence extract features for efficient data representation. Baseline of our approach has been different combination of deep learning based feature extraction methods with classifiers. Our developed CNN models consists of less number of layers yet achieved results comparable to other state-of-the-art. A detailed justification for differences in accuracy of our proposed models are discussed in the main article.