Hybrid Classification for Handwritten Character Recognition of a Subset of the Tamil Alphabet

In this paper, we present a hybrid classification approach using a weighted contribution of feature point-based and convolutional neural network (CNN) methods to improve classification and recognition of handwritten Tamil vowels. We explore different feature point methods, including BRISK, ORB, and KAZE, such that we featurize input images with a visual bag-of-words representation before applying standard machine learning classifiers including a multi-layer perceptron. We also consider different CNN configurations for this classification task. We find that since feature point methods and CNNs learn different representations of underlying input images, our hybrid classifier outperforms the individual feature point and CNN classifiers with respect to accuracy, precision, recall, and F1 score.