Fast median‐finding word comparator array

Abstract

Finding the median of a set of data within a window of finite size is computationally challenging on account of the complexity in sorting. Compared with the established nine-stage systolic arrays, an eightstage systolic array to find the median of a fixed 3 × 3 window of n bit integers is proposed. The proposed array requires a fewer number of signal paths, and is faster for fixed resource. The method employed uses selective comparators for finding the median by avoiding comparators required to obtain fully sorted list. This reduces the partial sorting to eight-stage systolic array. However, this is applicable only to median sorting of 3 × 3 fixed window. The method employs combinatorial circuit implementation; consequently there is 12% gain in speed and 7% fewer signal paths. Circuit with memory cells at every stage and eight latency cycles results in 36% speed gain over the state-of-the-art non-sorting based methods. This method provides superior performance in comparison with the methods available for fixed 3 × 3 window-based median filters.

A Low-Cost Design of 2D Median Filter

“…Besides, when the bit width of the test patterns increase, our architecture has the smallest growth rate in area because we adopt the comparison conditions with fixed bit-width for median value finding. As for the throughput, both [8] and [9] can produce a single median result every clock cycle. However, [10], [11] and our design can output four median results simultaneously.…”

“…In other words, our design needs 11 comparators to produce single median value on average but [11] needs 13. Since [8], [9] and [10] need 19, 19 and 13 comparators respectively to produce single median value, our design obviously requires less comparator cost.…”

“…for finding median value of a 3×3 mask. Inspired by [8], J. Subramaniam et al proposed a median-finding method [9] with fewer pipelined stages.…”

A High-Speed Low-Cost Hardware Implementation for Depth Estimation Using Disparity Fusion Method

“…Before the I Š>y 6 is output, the proposed algorithm will let the I Š>y 6 perform vertical median filtering again. Here, we use the serial comparison method, which was mentioned in [21].…”

“…When the disparity value of the ninth row is input, filtering begins. The median finding processing is finished by 19 compare-and-swap circuits [21]. After all disparity values are filtered, the final disparity map is generated.…”