Fast 2-D Bilateral Filtering Method without Kernel-size Dependency

Abstract

We propose a fast 2-D bilateral filtering method without kernel-size dependency. Focusing on weighted local histograms and the central limit theorem combined with applying line buffers of column histograms, enables us to reduce the number of necessary memory accesses and calculations. Numerical experiments demonstrated a reduction in both calculations and kernel-size independency. We used a dual core 2-GHz CPU with our method and were able to achieve one million pixels per 0.5 sec operation without the need for downsampling, Single Instruction/Multiple Data (SIMD) or multi-thread operation.