We apply a replica inference based Potts model method to unsupervised imagesegmentation on multiple scales. This approach was inspired by the statisticalmechanics problem of "community detection" and its phase diagram. Specifically,the problem is cast as identifying tightly bound clusters ("communities" or"solutes") against a background or "solvent". Within our multiresolutionapproach, we compute information theory based correlations among multiplesolutions ("replicas") of the same graph over a range of resolutions.Significant multiresolution structures are identified by replica correlationsas manifest in information theory overlaps. With the aid of these correlationsas well as thermodynamic measures, the phase diagram of the corresponding Pottsmodel is analyzed both at zero and finite temperatures. Optimal parameterscorresponding to a sensible unsupervised segmentation correspond to the "easyphase" of the Potts model. Our algorithm is fast and shown to be at least asaccurate as the best algorithms to date and to be especially suited to thedetection of camouflaged images.
展开▼
