Interpolation/decimation scheme applied to size normalization of character images

In several pattern recognition systems it is required that input images have the same dimensions, which can be achieved through size normalization. This works presents an interpolation-decimation scheme as a novel form to implement size normalization. The method can deal with any changes of size in images, by integer as well as non-integer scaling factors. This scheme uses cubic spline interpolation in two forms, the polynomial form and polyphase network form, as well as linear decimation by block averaging. A recurrent neural network has been used to test the effects of the proposed method on character recognition. The results obtained indicate that the rate normalization scheme can significantly improve the performance of the NN recognisor