Low-band-shift (LBS) motion estimation with symmetric padding in wavelet domain

Abstract

Low-band-shift (LBS) is a very efficient technique to compute and store the wavelet coefficients of all the shifted versions of an image. It makes it possible to perform motion estimation (called LBSME) directly in wavelet domain overcoming the shift variant property of discrete wavelet transform. However, when motion vectors of boundary wavelet blocks point outside the image boundaries, LBSME does not work well. In this paper, we proposed an algorithm called LBSME-SP to combine LBSME with symmetric padding to minimize the effects of boundary discontinuities. Consequently, the LBS structure is modified. Simulation results suggest that the proposed LBSME-SP can give considerable PSNR gain with only minor additional computation.