Fast edge-aware image filtering via domain splitting and transform

Abstract

Edge-aware filtering is a fundamental and important tool in CG, CAD, and image processing applications. In this research, we propose a novel and fast edge-aware image filter. Our filter is based on the domain splitting and domain transform techniques. First, we transform pixel coordinates according to their geodesic distance on the image manifold. Next, we split a pixel domain into representative regions where we can efficiently perform discrete L1 Gaussian convolutions. Our filter consists of linear computational complexity without ring-artifact (which could be generated by box-like kernels).