Structure-preserving image filtering with soft power iteration clustering

The structure-preserving filter is an indispensable operator in the field of computational photography and imaging. The objective is to decompose natural images into meaningful structures and fine-scale textures. The existence of possible high-contrast oscillations makes the problem rather challenging. The kernels of the gradient-based filters are mostly based on the Euclidean distance, which can not work out a satisfactory result. In this paper, we propose a novel structure-preserving image filter based on the soft power iteration clustering. Our filter kernel is built upon the diffusion/spectral distance, which exhibits promising capability in distinguishing structures from textures. The proposed filter consists of two iterative processes. The former one maps the image into the diffusion space by the power iteration filtering. The latter one smooths the image with the affinities derived from a soft clustering process in the diffusion space. The proposed filter is efficient, where each iteration takes linear time. Our filter outperforms the state-of-the-art structure-preserving filters in smoothing quality. Experimental results validate that our filter benefits various applications, including texture manipulation, edge detection, compression artifact removal, inverse halftoning, and image composition.