B-spline Surface Reconstruction from Scattered Data Points Based on SOM Neural Network

There often exist some problems,such as long training time and bad training effect etc.,when self-organizing map neural network(SOM) technology is employed in reverse engineering to reconstruct B-spline surface from scattered data points.In this paper,a new initialization method and a divide-and-conquer training scheme is presented.The approach functions as follows: firstly,the scattered data points are split into segments through principal component analysis(PCA);the neurons of output layer with quadrilateral topology are initialized on the least-square fitting planes of every segment.All the mesh surfaces obtained by training every segment respectively are integrated into a whole.Secondly,the boundary and interior neurons in the whole mesh surface are then trained and an approximate bi-linear B-spline surface is reconstructed.Finally,the B-spline surface reconstruction error is improved.Experiments show the proposed method can reduce SOM network training time and improve neural network training effect obviously.