Vectorization of Kernel and Image Subsampling in FIR Image Filtering

Abstract

Image subsampling is a traditional algorithm for accelerating image processing. The subsampling natively causes aliasing; thus, we use randomized sampling to moderate the issue. Also, SIMD vectorization speeds up processing by computing multiple data with a single instruction by hardware acceleration. However, randomized sampling is not suitable for SIMD vectorization. In this paper, we accelerate image filtering by subsampling filtering kernel and images by randomized algorithms. Also, the subsampling is vectorized by vector addressing. We describe how to implement subsampling of only images, the only kernel, and subsampling images and kernels. Also we compare where loop is unrolled is useful. Experimental results show that vectorization of kernel loop unrolling is faster than pixel loop unrolling. Also, kernel and image subsampling is effective for acceleration of bilateral filtering which a smoothing parameter of a range kernel is large.