Categorical Boosting Machine for Tamil Character Recognition Using Shape Based Features

Tamil is one of the world's earliest surviving languages, from which modern Indian scripts draw inspiration. However, there has been limited development of Optical Character Recognition techniques for Tamil characters. This paper describes a model for Tamil vowels-based Optical Character Recognition. The strategy seeks to utilise a CHFR feature extraction technique to extract three distinct feature vectors from the contours of each individual character's centroid. These feature vectors are condensed into a single vector before being sent to the CatBoost classifier for classification. Using the Optuna framework, hyperparameter optimisation and comparative analysis of existing models are performed. The proposed scheme is applied to a dataset of 6,972 handwritten samples evenly divided into 12 classes representing the vowels of the Tamil alphabet. The primary evaluation metric is accuracy, and 84.36% testing accuracy is observed.