Handwritten Numeral Recognition Based on DCT Coefficients and Neural Network

In the paper, we present a numeral recognition method based on multiresolution attributes of DCT coefficients and neural network which is non-linear mapping and error allowance. In this method the images are not subject to traditional preprocessing, but the pixel matrices of images are directly manipulated with DCT transform and multiresolution operation like wavelet decomposition. And the neural network is trained with the extracting features. Computer experiments are based on USPS (US Postal Service) numeral database. The result shows that the feature extraction method mentioned in the paper is more efficient and easier than directly using DCT and wavelet theory, and the structure of neural network could be simpler and it could converge much fast