Multiple sensor target classification using an unsupervised hybrid neural network

An unsupervised neural network has been developed to enable a robotic system to detect its target in an unknown environment by fusing multiple sensory information. The neural network consists of a feature extractor for each sensor used in the robotic system and a single classifier which takes input from all the feature extractors. The network is hybrid which combines the following: (a) a modified backpropagation learning rule in a self-organizing fashion, for extracting the features, and (b) the Kohonen linear vector quantization (LVQ) method to classify the objects based on the extracted features. The feature extractor detects the important features of the input image in different subelement groups of its hidden layer(s) by reproducing the input image in its output layer. The features obtained from each of the feature extractors are then fused and fed into a classifier which classifies the target based on these features. The overall hybrid network is unsupervised because it does not need the intervention of a human operator to provide the desired outputs during learning. Weight sharing is incorporated in each of the hidden layer groups of the feature extractors to reduce the number of free parameters. Also, a modified backpropagation learning rule has been used to improve the rate of convergence of the network