Machine Learning Final Project: Handwritten Sanskrit Recognition using a Multi-class SVM with K-NN Guidance

We develop an optical character recognition (OCR) engine for handwritten Sanskrit using a two-stage classifier. Inside the standard OCR pipeline, we focus on the classification problem assuming characters have been preprocessed decently. One challenge we face is that the language of Sanskrit has about a hundred core characters where model driven methods, like Support Vector Machine (SVM), have to search in the exponentially growth of the combinatoric model space during training, while data driven methods, like k nearest neighbor (kNN), becomes costly in computation during testing. To address this challenge, we propose a two-stage multiclassifier, using non-parametric to reduce the model space to search, and parametric models to relieve computation burden with better generalization. In the first stage, we apply kNN to coarsely assign the test data into the possible group of k classes, and a multiclassifier of k classes to label the sample in the second stage. Our method is fully automatic, highly accurate, and computational efficiently.