Robust and Accurate Image Registration with Pixel Selection

Image registration for real image sequences often fails to estimate accurate motion parameters because of illumination change, occlusion or an approximation error of a specific motion model. In this study, we propose a method to exclude pixels in regions of interest that have inadequately aligned images. We designate the remaining selected pixels as a "mask". We also propose a registration technique that employs the mask to estimate an accurate motion parameter. The proposed technique consists of two stages: the first stage estimates a motion parameter in successive frames for rough-but-robust registration; the second stage estimates a fine motion parameter between the input and the reference frame. Each stage employs its own mask. Experimental results using real image sequences demonstrate the effectiveness of the proposed technique.