基于核线性分类分析的三维模型检索算法

Optimized 3D shape retrieval using kernel fisher discriminant analysis

为提高检索精确度,提出了一种利用核线性分类分析来对模型特征进行优化的新方法。其主要思想是通过满足Mercer条件的非线性映射将低维空间下线性不可分的样本映射到高维空间,在高维空间中利用线性分类分析将原有的三维模型特征投影到特定的子空间。该方法能够在保持类间距离基础上得到具有鉴别信息的低维特征用于三维模型检索。实验结果表明,核线性分类分析方法速度较快,可在秒级完成三维特征优化,同时优化特征在本文测试数据集上可平均提高搜索准确度15%。

In this paper,kernel fisher discriminant analysis was adopted to optimize state-of-the-art 3D shape features,such as shape distribution and shape diameter function,to improve the precision of the query results. The main idea was to map the features into high-dimensional space using kernel method and then exploit the ability of linear discriminant analysis to maintain class separation so as to project the high dimensional 3D shape features to a subspace that can better separate classes to improve the discriminative power of features. Experimental results show that the optimization of 3D shape descriptor using kernel fisher discriminant analysis can be completed within a second and can improve the query precision by 15% on average over shape distribution.