A probabilistic DBNN with applications to sensor fusion and object recognition

Given an input vector x, a classifier is supposed to tell which class is most likely to have produced it. Thus most data classifiers are designed to have K output nodes corresponding to K classes, {w_i : i=1,...,K}. When pattern classes are clearly separated, this kind of data classifier usually performs very well. A specific model is the decision-based neural network (DBNN), which is effective in many signal/image classification applications. This is particularly the case when pattern classes are clearly separable. However, for those applications which have complex pattern distribution with two or more classes overlapping in pattern space, the traditional DBNN may not be effective or appropriate. For such applications, it is preferable to adopt a probabilistic classifier. In this paper, we develop a new probabilistic variant of the DBNN, which is meant for better estimate probability density functions corresponding to different pattern classes. For this purpose, new learning rules for probabilistic DBNN are derived. In experiments on face databases, we have observed noticeable improvement in various performance measures such as recognition accuracies and, in particular, false acceptance/rejection rates. Taking advantage of probabilistic output values of the DBNN, we construct a multiple sensor fusion system for object classification. In a sense, it represents an extension of the traditional hierarchical structure of DBNN