A Research On The Theory And Application Of Neural Network In Information Fusion

During the late in 1970s and early in 1980s, as a newly developing subject, information fusion has achieved rapid advances in the fields of robotics, the battlefield exploitation, target recognition and so on. This thesis makes a research on the theory and application of neural network in information fusion on the background of fusion target recognition.Firstly, this thesis summarizes the information fusion technology in target recognition shortly, and analyses the application and research situation of neural network in fusion target recognition at the data level, feature level and decision level.Secondly using for reference of the basic theory and lately research results in neurophysiology and cognitive science, this thesis discusses how human fuse the high level information tentatively with the hope of giving some reference to the research of information fusion. The mechanism that how nerve cell transfer the information and how the low level information is processed is also proposed in this thesis, on the basis of which we emphasize the course of information processing in vision neural system. Furthermore we express some rules obeyed in the research of information fusion and establishing information fusion system.Thirdly the basic theory and applying method of fuzzy neural network in information fusion are addressed here. We discuss the multi-level forward neural network based on fuzzy inference, put forward the training algorithm based on back-propagation, and investigate the general method of applying this model to feature level fusion. Next the general method of applying fuzzy ARTMAP model to feature level fusion is also expounded and we put forward a learning algorithm with adaptive v igilance parameters for each c luster. The simulation of applying the new network to classify the targets indicates this model is effective.In the end we investigate the general method of applying neural network to temporal-spatial fusion at decision level. Aiming at the shortage of common network which cannot make use of the expert knowledge and environmental information sufficiently, a recursive temporal-spatial fusion network at decision level is put forward in this thesis. We also predigest the network structure with the aid of neurophysiology theory and initialize the network. Besides we give an on-line learning arithmetic for the network's connected weights.