Wavelet transform of Projection profile of character images has been found to be suitable for machine recognition of handwritten characters. In this work, performance analysis of such a feature using twelve different wavelet filters and different training / test data sets is carried out. A total of twelve thousand eight hundred handwritten isolated Malayalam characters belonging to 33 classes were used for the study. An MLP network is used as classifier. It is observed that the performance of different wavelet filters used in the study is more or less same. The average recognition accuracy is 76.8%. The above work is extended by adding one more feature, the aspect ratio and found significant improvement in recognition, the average being 81.3%. Considering the relatively large data set used, the result obtained is promising.
展开▼
