A new stochastic block matching algorithm (SBMA) for video coding based on modified 3-step search

Motion compensated predictive coding is one of the most efficient video coding methods. Full search gives the optimal performance in minimizing the sum of absolute difference (SAD) of pixel values within a block for residual images. We propose a new fast SBMA which exploits spatial and temporal correlations existing in the motion vector field. The new SBMA has a variable search window size and a shiftable search center. Within a given search window, we consider a new search procedure based on a modified 3-step search. Compared to full block matching, we achieve a speed up factor ranging from 70 to 150. As a trade off, we see a 1-12% increase in the error residuals, and the corresponding PSNR value degrades by 0.1-0.5 dB only.