Memory optimization of bilateral filter and its hardware implementation

As a method for edge-preserving or noise-reducing, a bilateral filter is widely used. However, because every pixel in a filtering window needs a separate Look-Up Table (LUT) for the parallel processing, its hardware implementation is still bulky. In this paper, we propose Similar Weight Grouping (SWG) which maps multiple indexes with a similar value onto a single index and Zero Value Suppression (ZVS) which removes indexes with a value of almost zero to reduce a size of the LUT. By our scheme, a total size of LUT is reduced by approximately 95% while maintaining its performance. Finally, it is implemented using 7.7 KB on-chip memory and 93.1 Kgates with 65 nm process.