基于权重能量自适应分布的三维形状分割算法

针对三维形状分割问题,提出一种引入权重能量自适应分布参与深度神经网络训练的全监督分割算法.首先对三维形状表面进行过分割得到若干小块,提取每一个小块的特征描述符向量作为神经网络的输入,计算权重能量自适应分布,将经过加权后的分割标签作为神经网络的输出,训练深度神经网络.对于新的未分割的三维模型,提取模型表面三角面片的特征向量后输入到神经网络中进行预测分割后,对预测分割的边缘进行修整得到分割结果,实现三维模型的自动分割.在普林斯顿三维模型分割数据集上的实验结果表明,算法通过在训练过程中引入权重能量自适应分布,可以大幅降低神经网络训练时的均方误差,提高神经网络预测结果的准确率;与传统算法相比,该算法具有高准确率、强鲁棒性、强学习扩展能力等优点.

3D Shape Segmentation Algorithm Using Weighted Energy Adaptive Distribution

In this paper, a fully supervised segmentation algorithm is proposed by using weighted energy adaptive distribution(WEAD). Firstly, the 3 D shapes are divided into several small patches using an over-segmentation method. Secondly, feature vectors of the patches are extracted as the training input and WEADs are used as the training output to re-weight the labels of segmentation. Finally, a corresponding deep neural network is trained. For an unlabeled 3 D shape, our algorithm automatically segments it by using the trained deep neural network followed by the graph cuts method. Extensive experimental results show that the mean square error(MSE) in the training process can be greatly reduced by using WEAD in our algorithm. And our method performs better than other fully supervised and unsupervised algorithms on the Princeton Segmentation Benchmark(PSB) dataset.