Gauss mixture image classification for distributed sensor networks.

摘要

Classification algorithms based on Gauss mixture models (GMMs) provide robust and analytically tractable solutions to image classification problems. Training a GMM might use the expectation-maximization (EM) algorithm or Gauss mixture vector quantization. Application of the EM algorithm is based on the assumption that the underlying data follow a Gauss mixture distribution. The goal is to fit a GMM to the data. Gauss mixture vector quantization is a Lloyd clustering algorithm, and it requires no assumptions about the statistics of the underlying data.; I first extend Gauss mixture vector quantization to design a tree-structured GMM-based classifier. Tree-structured classifiers provide a method to focus on the difficult regions of the training vector space by growing classification trees into those regions. I design a tree-structured Gauss mixture vector quantizer (TS-GMVQ) by first growing the tree into "difficult" regions, and then pruning it optimally, using the Breiman-Friedman-Olshen-Stone (BFOS) algorithm to avoid overfitting.; I then focus on the GMM-based classification problem for sensor networks. Previous work on the EM algorithm and Gauss mixture vector quantization has emphasized single sensor classification problems. I generalize the GMM-based classification problem to include multiple, distributed sensors. In particular, I consider a set of sensors, communicating with each other under rate constraints for the purpose of classification. Each sensor has a different noisy version of a common image and aims to classify the image based on its own noisy version and the help it receives from the other sensors. I view the sensor network classification problem as one in vector quantization and provide a Lloyd-optimal solution to minimize classification error for the given rate constraints. In particular, I use a TS-GMVQ to partition the vector space during the Lloyd design. I also include context dependence into my algorithm, making use of the concepts developed in conjunction with hidden Markov models.