Modified View Based Approaches for Handwritten Tamil Character Recognition

Finding simple and efficient features for offline hand written character recognition is still an active area of research. In this work, we propose modified view based feature extraction approaches for the recognition of handwritten Tamil characters. In the first approach, the five views of a normalized and binarized character image viz, top, bottom, left, right and front are extracted. Each view is then divided into 16 equal zones and the total numbers of background pixel in each zone are counted. The 80 values so obtained form a feature vector. In the second approach, the normalized and binaraized character images are divided into 16 equal zones. Five views are extracted from each zone and the total number of background pixel in each view is counted, resulting in 80 feature values. Further the above two approaches are modified by employing thinned images instead of the whole image. The extracted features are classified using SVM, MLP and ELM classifier. The discriminative powers of the proposed approaches are compared with that of four popular feature extraction approaches in character recognition. The feature extraction time and classification performances are also compared. The proposed modified approaches results in high classification performance (95.26%) with comparatively less feature extraction time.