基于特征表示与学习的图像集合分类算法  

On Image Set Classification Algorithm Based on Feature Representation and Learning

作　　者：卢滢宇[1] 金錱 

机构地区：[1]宁波职业技术学院,浙江宁波315800 [2]浙江大学宁波理工学院,浙江宁波315100

出　　处：《实验室研究与探索》2016年第2期128-134,共7页Research and Exploration In Laboratory

摘　　要：当前大多数图像集合分类方法对图像集合进行表示时往往做出部分先验假设,然而在许多实际应用中,这些假设可能无法成立,尤其是当集合内部存在大量复杂的数据变化时更是如此。此外,基于这些假设进行模型学习时可能会丢失部分区分性分类信息。针对这一问题,本文提出一种基于特征表示与学习的图像集合分类方法。对每个图像集合,首先将计算它的多阶统计量作为特征表示。对每阶统计量,计算一个内核矩阵来衡量两个图像集合的相似性。然后,通过利用局部多内核指标学习(LMKML)方法来学习一种距离指标,进而将不同阶统计量综合起来。最后,利用最近邻分类器进行分类。基于4种常用图像集合数据库的实验结果验证了本文算法的有效性。Most existing image set classification methods usually need to make some prior assumptions. However,in many practical applications,these assumptions may not be held,especially when there are large and complex data variations within a set. Moreover,the models learned based on these assumptions may also lose some discriminative information for classification. To solve this problem,this paper proposes an image set classification method based on feature representation and learning. For each image set,we first compute its multiple order statistics as feature representation. For each order statistic,we compute a kernel matrix to measure the pairwise similarity of two image sets.Then,we learn a distance metric by using the localized multi-kernel metric learning method to combine the different order statistics. Lastly,the nearest neighbor classifier is used for classification. Experimental results on four widely used image set datasets are presented to show the efficacy of our proposed approach.

关 键 词：图像集合分类 多阶统计量 特征表示 内核矩阵 近邻分类器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]