Parallel algorithms for 3D multi-level median filtering with motion compensation

This paper aims to develop effective parallel processing techniques for 3-dimensional (3D) multi-level median filtering (3DMMF) with motion compensation on video sequences. Due to the nature of unbalanced load of the filtering algorithm and the objective for the parallel approach, two methods of dynamic load balancing have been proposed and compared. They are sender-initiated-load-balancing (SILB) and receiver-initiated-load-balancing (RILB) algorithms. We propose a SILB algorithm which utilises the spatial-temporal characteristics of the processed sequences for load prediction to achieve dynamic load balancing. Both theoretical analysis and experimental results on the IBM SP2 computing surface have been presented in this paper.