On the spectral-based reduction of the computational demand for the prediction phase in image classification with artificial neural networks