Global Similarity with Additive Smoothness for Spectral Segmentation

Abstract

Faithful representation of pairwise pixel affinities is crucial for the outcome of spectral segmentation methods. In conventional affinity models only close-range pixels interact, and a variety of subsequent techniques aims at faster propagation of local grouping cues across long-range connections. In this paper we propose a general framework for constructing a full-range affinity matrix. Our affinity matrix consists of a global similarity matrix and an additive proximity matrix. The similarity in appearance, including intensity and texture, is encoded for each pair of image pixels. Despite being full-range, our similarity matrix has a simple decomposition, which exploits an assignment of image pixels to dictionary elements. The additive proximity enforces smoothness to the segmentation by imposing interactions between near-by pixels. Our approach allows us to assess the advantages of using a full-range affinity for various spectral segmentation problems. Within our general framework we develop a few variants of full affinity for experimental validation. The performance we accomplish on composite textured images is excellent, and the results on natural images are promising.