Sorting network based architectures for median filters

Abstract

Sorting network based architectures for computing nonrecursive and recursive median filters are presented. The proposed architectures are highly pipelined and consist of fewer compare-swap units than existing architectures. The reduction in the number of compare-swap units is achieved by minimizing computational overlap between successive outputs and also by using Batcher's odd-even merge sort (instead of bubble-sort). The latency of these networks is reduced by building them with sorting units that sort 2 elements (sort-2) as well as 3 elements (sort-3) in 1 time unit. >