Multichannel segmentation using contour relaxation: Fast super-pixels and temporal propagation

Abstract

The contribution describes a statistical framework for image segmentation that is characterized by the following features: It allows to model scalar as well as multi-channel images (color, texture feature sets, depth, ...) in a region-based manner, including a Gibbs-Markov random field model that describes the spatial (and temporal) cohesion tendencies of 'real' label fields. It employs a principled target function resulting from a statistical image model and maximum-a-posteriori estimation, and combines it with a computationally very efficient way ('contour relaxation') for determining a (local) optimum of the target function. We show in many examples that even these local optima provide very reasonable and useful partitions of the image area into regions. A very attractive feature of the proposed method is that a reasonable partition is reached within some few iterations even when starting from a 'blind' initial partition (e.g. for 'superpixels'), or when - in sequence segmentation - the segmentation result of the previous image is used as starting point for segmenting the current image. © 2011 Springer-Verlag.