基于双流形正则化的迁移稀疏算法研究

Research on Transfer Sparse Algorithm Based on Dual-Manifold Regularized

为了完善跨域图像表示模型、增强跨域图像的分类功能,研究基于双流形正则化项的迁移稀疏表示算法。在迁移稀疏编码的基础上,利用特征间局部流形信息构建拉普拉斯图,将特征拉普拉斯正则化项融入目标函数中。首先,通过k均值聚类均衡选择基向量;其次,根据特征间局部流形结构信息构建拉普拉斯图,并将该拉普拉斯图作为正则化项引入到迁移稀疏编码算法的目标函数中;同时,将跨域图像的几何流形结构信息和分布差异信息考虑进去,以保证编码的稳定性和鲁棒性。

In order to further improve the cross domain image representation model and enhance the classification function of cross domain image,a dual-manifold regularized transfer sparse concept coding is proposed. On the basis of migrating sparse coding,Laplacian graph is constructed with local manifold information between features,and the characteristic Laplacian regularization term is incorporated into the objective function. Firstly,k-means clustering method is used to balance the initial vectors;secondly,the local geometrical information of features is used to construct Laplacian graph,and the Laplacian graph is introduced as the regularization term into the objective function of transfer sparse coding. What’s more,the information of geometric features and the distribution differences are considered,which can insure a stable and robust image representation.