Rank Based Merge Sorting Network Architecture for 2D Median and Morphological Filters

We present the design principles of Rank Based Merge Sorting Network (RBMSN) architectures for the realization of 2D median and morphological filters used in image preprocessing. The proposed architectures focus on optimization strategies for sorting in terms of the number of comparators and throughput. The minimization in the computational cost is achieved by rank range based merging, column sorting and storing the sorted elements of the overlapping columns of the consecutive windows at each intermediate stage of the sorting network. The proposed architecture uses the concepts of pipelining and grain leel parallelism to accomplish the task of processing one pixel per clock cycle. The architectures for median erosion and dilation filters are synthesized for 3×3 and 5×5 window sizes. The proposed RBMSN Median filter architectures for N=√N × √N window size require N/2(Log
₂
N) comparators and √N(√N - 1), memory registers. The proposed design and implementations are compared with a few of the reported architectures.