一种基于迭代式稀疏谱聚类的图像集分类算法  

An Image Set Classification Algorithm Based on Iterative Sparse Spectral Clustering

作　　者：谢炜[1] 

机构地区：[1]桂林理工大学理学院,广西桂林541004

出　　处：《计算机工程》2016年第4期242-247,共6页Computer Engineering

基　　金：广西自然科学基金资助项目(2013GXNSFAA278003)

摘　　要：现有图像集分类方法无法直接应用于不同类别的子空间,图像集分类既不独立也不相交。为此,提出一种迭代式稀疏谱聚类算法。每次迭代时利用NCut目标函数将一个母聚类划分为多个更小的子聚类,提升图像集中不同类型噪声的区分性和算法稳健性。为降低谱聚类的计算复杂性,不仅将谱聚类的计算成本降为原先的几分之一,而且提升了聚类质量和最终的分类结果。利用3种标准的脸部图像集合、对象分类以及Cambridge手部姿势数据集进行实验。与7种最新算法的比较结果表明,该算法在各个数据集上的性能均优于其他算法。对于聚类难度最大的Youtube数据集,其性能提升最为明显,比其他算法的最优精度高出11.4%。Aiming at the disadvantages of the existing image set classification methods are not directly applicable of the subspaces and different classes are neither independent nor disjoint, an iterative sparse spectral clustering algorithm is proposed. It performs clustering iteratively and each iteration divides a parent cluster into small number of child clusters using the NCut objective function, which reduces computational complexity and increases discrimination and robustness to different types of noise in the image sets. In order to reduce the computational complexity of spectral clustering, it not only reduces the computational cost of spectral clustering by many folds, but also improves the clustering quality and final classification results. Experiments are performed on three standard face image-sets, an object categorization, and Cambridge hand gesture datasets. Results are compared with 7 states of the art algorithms. The proposed technique achieves the highest accuracy on all datasets. The maximum improvement is observed on the most challenging Youtube dataset where the algorithm achieves 11.4% higher accuracy than the other methods.

关 键 词：图像集分类 子空间 迭代 谱聚类 计算成本 精度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]