Applied some new features in off-line recognition of totally unconstrained handwritten numerals using neural network

Some new features in recognition of totally unconstrained handwritten numerals are applied. The new features are based on image projection and do wavelet transformation, then at the deferent scale, they calculate the projection´s fractal dimension, using the fractal dimension as a feature applied to neural network input. The new features have some advantages: it is rotate invariant, and it can represent the image´s characteristic at deferent scale. In order to verify the performance of the new features, experiments with a handwritten numeral database collected from Beijing Postal center were performed. The correct recognition rate in the training set was 99.55% and in the testing set was 96.5%