Equivalence of Julesz Ensembles and FRAME Models
Ying Nian Wu,Song Chun Zhu,Xiuwen Liu
DOI: https://doi.org/10.1023/A:1008199424771
IF: 13.369
2000-01-01
International Journal of Computer Vision
Abstract:In the past thirty years, research on textures has been pursued along two different lines. The first line of research, pioneered by Julesz (1962, IRE Transactions of Information Theory , IT-8:84–92), seeks essential ingredients in terms of features and statistics in human texture perception. This leads us to a mathematical definition of textures in terms of Julesz ensembles (Zhu et al., IEEE Trans. on PAMI , Vol. 22, No. 6, 2000). A Julesz ensemble is a set of images that share the same value of some basic feature statistics. Images in the Julesz ensemble are defined on a large image lattice (a mathematical idealization being Z 2 ) so that exact constraint on feature statistics makes sense. The second line of research studies Markov random field (MRF) models that characterize texture patterns on finite (or small) image lattice in a statistical way. This leads us to a general class of MRF models called FRAME (Filter, Random field, And Maximum Entropy) (Zhu et al., Neural Computation , 9:1627–1660). In this article, we bridge the two lines of research by the fundamental principle of equivalence of ensembles in statistical mechanics (Gibbs, 1902, Elementary Principles of Statistical Mechanics . Yale University Press). We show that 1). As the size of the image lattice goes to infinity, a FRAME model concentrates its probability mass uniformly on a corresponding Julesz ensemble. Therefore, the Julesz ensemble characterizes the global statistical property of the FRAME model; 2). For a large image randomly sampled from a Julesz ensemble, any local patch of the image given its environment follows the conditional distribution specified by a corresponding FRAME model. Therefore, the FRAME model describes the local statistical property of the Julesz ensemble, and is an inevitable texture model on finite (or small) lattice if texture perception is decided by feature statistics. The key to derive these results is the large deviation estimate of the volume of (or the number of images in) the Julesz ensemble, which we call the entropy function. Studying the equivalence of ensembles provides deep insights into questions such as the origin of MRF models, typical images of statistical models, and error rates in various texture related vision tasks (Yuille and Coughlan, IEEE Trans. on PAMI , Vol. 2, No. 2, 2000). The second thrust of this paper is to study texture distance based on the texture models of both small and large lattice systems. We attempt to explain the asymmetry phenomenon observed in texture “pop-out” experiments by the asymmetry of Kullback-Leibler divergence. Our results generalize the traditional signal detection theory (Green and Swets, 1988, Signal Detection Theory and Psychophysics , Peninsula Publishing) for distance measures from iid cases to random fields. Our theories are verified by two groups of computer simulation experiments.