Multi-frame image restoration using a neural network

A neural implementation of a multiframe image restoration algorithm is presented. Given the amount of movement between frames, the overlapping area can be restored in the least-square sense using a neural network approach. In the case of different noise levels from one frame to the next, a weighting factor is introduced to optimize the restoration process. Methods for choosing the weighting matrix to obtain a minimum variance linear unbiased estimator of the overlapping portion are discussed. Experimental results on restoring the original image from two noisy frames are presented. Future research directions and perspectives are discussed