基于密度峰值的三维模型无监督分类算法

Unsupervised 3D Shape Classification Algorithm Using Density Peaks

针对基于内容的三维模型自动分类问题,提出一种密度峰值驱动的三维模型无监督分类算法.首先利用多种特征描述符分别对每个三维模型提取相应的特征向量;然后将得到的特征向量运用鲁棒主成分分析去除噪声并降维;最后通过计算特征向量分布的密度峰值,并配合决策图,以直观的方式确定三维模型分类类别数,最终实现三维模型的无监督分类.实验结果表明,与传统算法相比,该算法具有易于确定分类类别数、准确率高、鲁棒性强等优点.

In this paper, we propose an unsupervised classification algorithm by using density peaks for automaticcontent-based 3D model classification. Firstly, the algorithm extracts multiple kinds of feature vectors for eachmodel in the given shape collection. Secondly, it uses robust principal component analysis to denoise the featurevectors and reduce their dimensions simultaneously. Finally, the algorithm determines the number of categories ofthe 3D models and realizes an unsupervised classification in an intuitive and visual way by computing the densitypeaks of the feature vectors’ distribution and a corresponding decision graph. Extensive experimental results showthat the number of categories of clustering is much easier to determine and the results are more accurate and robustin our algorithm when compared with the traditional algorithms.