Gauss mixture model-based classification for sensor networks

Gauss mixture models (GMMs) provide an approach to the image classification problems, utilizing the robustness and the analytical tractability of the Gaussian distribution. Previous work on the GMM-based image classification algorithms has focused on either the single sensor schemes or schemes with no noise or rate constraints. In our work, we consider a GMM-based image classification problem for a network of sensors with each sensor having a different noisy version of a common image. The goal of each sensor is to classify the image based on its own noisy version and the help it receives from the other sensors under rate constraints. We formulate the image sensor network classification problem as a vector quantization problem and design Lloyd optimal quantizers, minimizing the classification error for the given rate constraints. We then extend our algorithm to include context dependence. Our cross-validated simulations, using a set of aerial images, indicate an improvement in the classification performance (for the given rate constraints) when compared with the network extensions of previously published GMM-based algorithms.