An Image Registration Algorithm through Perception based Weighted Cost Optimization of Pixel Labelling

Abstract

Semi-automated image registration techniques depend on human eye validation of final image
alignment. This results in non-uniform registration based on color discrimination based on the
perception of different users. This paper describes a registration method based on perception that is
optimized for a cost function weighted on prominence of features perceived within a group of pixels
after labelling them. The weights are calculated so as to minimize the penalty of over-emphasizing nonprominent features for registration. The key points are then matched using nearest neighbor distance
ratio (NN-DR) followed by motion estimation and a geometric transformation to obtain the final image.
The technique has yielded better results and has been proved excellent especially in the image stitching
application after registering the similar pixels