动态生长的自组织神经网络点云重建技术

Dynamic Growing Self-organizing Maps for Surface Reconstruction from Point Clouds

为了提高自组织特征映射网络算法中点云重建技术的质量、收敛速度和表面精度,提出一种动态生长的自组织神经网络算法.首先基于自组织神经网络算法,构造了球体三角网格作为神经网络的映射结构,正确选择拓扑邻域的环数,通过对大量无规则节点进行网络训练和学习达到神经元节点的分裂,改变了网络结构的固定性,并删除不稳定的网格节点;然后对网格进行优化,让神经元节点与输入的离散点保持更加的紧密,得到较好的点云重建结果.与自组织特征映射算法训练特性相比,该算法减少了计算量,提高了网络训练的收敛速度和离散点云重建的表面精度,特别是针对海量点云数据或者含有大量噪声点云数据的重建效果更明显.

In order to improve the quality, rate of convergence and surface accuracy of point cloud reconstruction in the self-organizing neural network, the dynamic growing self-organizing neural networks algorithm is proposed in this paper. Firstly by the self-organizing maps algorithm, we construct the spherical triangle mesh as maps of the neural network and select the right loop numbers of the topology neighborhood. Then, we split the nodes and delete the unstable nodes to change the immobility of the network structure by training and learning of neural network for unorganized scattered point clouds. In addition, we optimize the grid to make the neural nodes and discrete points keep closer together. Finally, experiments demonstrate that this method can generate favorable results. Compared with the training characteristics of the self-organizing neural network, the algorithm can reduce the amount of calculation and improve the rate of convergence and surface accuracy of the scattered point clouds reconstruction. Especially, it is more apparent of effect for the reconstruction of a huge amount of data or the point clouds with a lot of noise.