Research On Surface (Line) Reconstuct Based On RBF Neural Networks

RBF Neural Networks' simple structure and excellent approximationcapability arouse scholars' broad attention. Because of RBF Neural Networks'special connective structure and training method, make it use on functionapproximation and non-linear forecast system.Presently domestic and overseasresearch on how to use RBF Neural Network to reconstruct surface is veryactivity. For example, people use three independently RBF Neural Networks togain the relationship of surface's three coordinate and parameters, and then gainthe relationship of three coordinate indirectly, but this method affect network'straining speed precision of surface reconstruct. This dissertation research on howto use one RBF Neural Network gain the relationship of surface's threecoordinate directly, this relationship is amended by networks' power value andvalve value.The method adopt gradient descent d arithmetic ,use this arithmeticto training mapping relationship, from training and study, and get reconstructivesurface step by step. This arithmetic haye very high precision and it's veryrobust, This dissertation also research on how to choose center vector, because if choose too many center vectors will lead over imitate, and will let down it'sgeneralize ability; If choose too few center vectors, then will not get enoughlearning information and let down it's generalize ability too; Besides thisdissertation via experiment study how shape parameter affect network'sperformance and how to use RBF Neural Network to reconstruct surface onunorganized data.