Wavepackets in the recognition of isolated handwritten characters

This work is to apply wavelet packet transformation for the recognition of isolated handwritten Malayalam (one of the south Indian languages) characters. The key idea is that count of zero crossings of wavelet transform coefficients of an image characterize it. A set of 3000 images of 20 selected characters are used for classification. All images are normalized to have same height, binarized and inverted. Two-level Wavelet packet transformation is applied on each character image. Count of zero-crossings in each of the sixteen subbands together with a structural feature forms the feature vector. Feed forward back propagation network is used for classification. We obtained about 90% accuracy in classification and recognition. Further stuy by including more characters and more samples is being carried out.