Point Feature Based Recognition of Handwritten Meetei Mayek Script

In this article we are proposing a hybrid point feature based recognition of Meetei Mayek optical handwritten characters. Four different interesting point features namely Harris corner detector, Laplacian-of-Gaussian (Log) detector, Harris-Laplacian detector, and Gilles feature are used to detect different fixed number of interesting feature points. The [X, Y] coordinates of these interesting points are used as feature vector. Four different feature vectors are generated out of these and we have developed a fused feature out of it. Feature reduction is done by PCA, and accuracy calculations for all these features are performed using support vector machine. A comparative analysis is done to get a feature that shows consistency in accuracy before and after applying feature reduction. Experimental result shows that fused feature FV5 shows better accuracy in all different situations. Before reduction it shows 97.16% accuracy, on reduction by probabilistic PCA it shows 94.15% accuracy, and PCA-based reduction 97.16% accuracy, which is best in all situations as compared to other features.