Hierarchical Bayesian image analysis: from low-level modeling to robust\n supervised learning

Abstract

Within a supervised classification framework, labeled data are used to learn\nclassifier parameters. Prior to that, it is generally required to perform\ndimensionality reduction via feature extraction. These preprocessing steps have\nmotivated numerous research works aiming at recovering latent variables in an\nunsupervised context. This paper proposes a unified framework to perform\nclassification and low-level modeling jointly. The main objective is to use the\nestimated latent variables as features for classification and to incorporate\nsimultaneously supervised information to help latent variable extraction. The\nproposed hierarchical Bayesian model is divided into three stages: a first\nlow-level modeling stage to estimate latent variables, a second stage\nclustering these features into statistically homogeneous groups and a last\nclassification stage exploiting the (possibly badly) labeled data. Performance\nof the model is assessed in the specific context of hyperspectral image\ninterpretation, unifying two standard analysis techniques, namely unmixing and\nclassification.\n