基于神经网络的几何模型区域分割算法

Region segmentation method of geometric model based on neural networks

三维模型简化是近年来计算机图形学中的一个研究热点,现有的简化算法多从全局出发,对几何模型的各个部位统一进行简化,因此模型简化后大量的细节特征丢失。针对三维模型简化中保留细节特征的需要,提出了一种基于自组织特征映射神经网络的三维模型区域分割算法。首先计算三维几何模型中每一顶点的特征向量,然后利用该向量作为自组织特征映射神经网络的输入模式实现对三维模型的聚类分割,最后采取提出的相关性最大准则对过分割区域进行合并,得到最终分割结果。实验表明,该方法能有效地分割出模型的细节区域,满足三维模型简化中保留细节特征的需要。

3D model simplification is a focus of study in computer graphics area in recent years. Most of the existed 3D model simplification algorithms take the model as a whole when making simplification without considering each part of the model, so lots of details are lost. Focusing on detail preserving in 3D model simplification, we put forward a new 3D geometric model region segmaentation method based on self-organizing property mapping neural network. Firstly, the characteristic vector of each vertex of 3D geometric model is calculated out. Then the vector is taken as input mode of the self-organizing property mapping network for segmenting the 3D geometric model preliminarily. Finally, the over-segmentation regions are merged by using max-rehtivity rule and the final segmentation is completed. Experiment shows that this approach is effective, which can meet the requirement of keeping local features in model simplification.