Neural network high precision processing for astronomical images

In this paper we deal with the diagnosis and removal of chromaticity, a relevant source of error in high precision astrometric measurements, using a feed forward neural network and focuse on the usefullness of a carefully optimised image processing. The first problem we study is the image construction via Fourier transform so we suggest a method to effectively evaluate it no longer involving FFT algorithm but via direct matrix multiplication. The second problem is related to the necessity of a good choise of the parameters used to encode images, that we solved with a careful preprocessing and filtering; these parameters are then used as inputs to a feed forward neural network trained by backpropagation to remove chromaticity.