Multiple graph regularized sparse coding and multiple hypergraph regularized sparse coding for image representation
Multiple graph regularized sparse coding and multiple hypergraph regularized sparse coding for image representation
复制标题
用于图像表示的多图正则化稀疏编码和多超图正则化稀疏编码
DOI：
10.1016/j.neucom.2014.11.067
复制
发表时间：
2015
期刊：
影响因子：
6
通讯作者：
Li Cuihua
中科院分区：
文献类型：
--
作者：
Jin Taisong;Yu Zhengtao;Li Lingling;Li Cuihua
流形正则化稀疏编码在各种应用中表现出良好的性能。在应用中必须考虑的关键问题是如何自适应地选择合适的图超参数在流形学习稀疏编码任务。通常，应用交叉验证，但它不一定会扩大规模，容易导致过拟合。本文提出了用于图像表示的多重图稀疏编码（MGrSc）和多重超图稀疏编码（MHGrSc）。受EngravityManifoldRegularizer的启发，我们构造了多重图和多重超图正则化器，以保证稀疏码沿着数据流形的测地线的光滑性，其特征在于融合了多个先前给定的图拉普拉斯算子或超图拉普拉斯算子。然后，提出的正则化，分别纳入到传统的稀疏编码框架，这导致两个统一的目标函数的稀疏编码。采用交替优化方法对目标函数进行优化，提出了两种新的流形正则化稀疏编码算法。提出的两种稀疏编码方法联合学习复合流形和稀疏编码，并且在流形学习中学习图超参数是全自动的。在真实的数据集上的图像聚类实验表明，本文提出的稀疏编码方法优于现有的方法，具有上级的性能。
Manifold regularized sparse coding shows promising performance for various applications. The key issue that must be considered in the application is how to adaptively select the suitable graph hyper-parameters in manifold learning for the sparse coding task. Usually, cross validation is applied, but it does not necessarily scale up and easily leads to overfitting. In this article, multiple graph sparse coding (MGrSc) and multiple Hypergraph sparse coding (MHGrSc) for image representation are proposed. Inspired by the Ensemble Manifold Regularizer, we formulate multiple graph and multiple Hypergraph regularizers to guarantee the smoothness of sparse codes along the geodesics of a data manifold, which is characterized by fusing the multiple previously given graph Laplacians or Hypergraph Laplacians. Then, the proposed regularziers, respectively, are incorporated into the traditional sparse coding framework, which results in two unified objective functions of sparse coding. Alternating optimization is used to optimize the objective functions, and two, novel manifold regularized sparse coding algorithms are presented. The proposed two sparse coding methods learn both the composite manifold and the sparse coding jointly, and it is fully automatic for learning the graph hyper-parameters in the manifold learning. Image clustering tests on real world datasets demonstrated that the proposed sparse coding methods are superior to the state-of-the-art methods.