Modular multi layer perceptron (MMLP) to identify objects composed of characteristic sub-parts

A scheme that identifies 3D complex objects on short range images is briefly reviewed. This scheme was developed to identify armoured vehicles on IR images taken at short range and addresses the problems of occlusion and 2D aspect variability of 3D objects. One key-point of the approach is a MMLP. This neural network is composed of small sub networks (MLP's), each one learning properties of one characteristic sub-part. The sub network output combination method is grounded on a bayesian approach and can be interpreted as the fusion of a number of experts (the sub-networks), each of which identifying the object by looking at one particular detail. The underlying hypotheses are presented and discussed. One hypothesis is shown to be questionable and is relaxed by means of a renormalisation process.