Nonparametric multisensor image segmentation and classification

Abstract

Nonparametric multisensor systems for image segmentation and classification are presented for which no knowledge of the statistical behavior of the training data and the quantized gray levels from the sensors is required. The joint probability density function of the quantized gray levels is estimated at the fusion center following a density estimation approach which is based on a kernel function and the training data and is implemented via a probabilistic neutral network. The quantizers of the sensors are designed according to a signal-to-noise-type design criterion which is a function of the training data only and couples the data sequences of the various sensors.<>