Automatic Contextual Pattern Modeling

An object/scene usually consists of several primitives among which various contextual relations are defined. In this work, an object/scene is defined as a contextual pattern and is represented by an attributed relational graph (ARG). We develop the methodology and theory for automatic contextual pattern modeling to automatically learn a probabilistic pattern ARG model from multiple sample ARGs. The maximum-likelihood parameters of the pattern ARG model are estimated via the Expectation-Maximization algorithm. Particularly, for Gaussian attributed distributions and Gaussian relational distributions, analytical expressions are derived to estimate the parameters of the distribution functions. The learned pattern ARG model characterizes both the appearance and the structure of the pattern, which is observed under various conditions. It can be used for information summarization and retrieval, improving graph matching, and pattern detection/recognition. We demonstrate the theory by applying it to the problem of unsupervised spatial pattern extraction from multiple images.