结合子空间Laplacian正则项的迁移联合匹配方法

Transfer joint matching method with subspace Laplacian regularization term

基于核主成分分析的迁移联合匹配方法在处理高维数据的降维问题时,存在难以获得更具判别力的特征这一缺陷,由此,提出了一种结合子空间Laplacian正则项的迁移联合匹配方法。该法将原始数据映射到高维子空间,在高维子空间中提取Laplacian正则项和MMD分布适配项,以进一步获取更具判别力的特征,在此基础上重新调整实例权重以保证源域与目标域的一致性。最后在通用迁移数据集上对所提方法的有效性和鲁棒性进行测试,结果表明所提方法能提升跨域图像的平均分类性能并具有很好的鲁棒性。

The transfer joint matching method based on kernel principal component analysis has the defect that it is difficult to obtain more discriminative features when dealing with the dimensionality reduction of high-dimensional data.Therefore,this paper proposed a transfer joint matching method combined with subspace Laplacian regularization term.In this method,the original data is mapped to the high-dimensional subspace,and the Laplacian regularization term and MMD distribution adaptation term are extracted from the high-dimensional subspace to further obtain more discriminative features.On this basis,the weights for cases are re-adjusted to ensure the consistency between the source domain and the target domain.Finally,the effectiveness and robustness of the proposed method are tested on the general transfer dataset.The results show that the proposed method can improve the average classification performance of cross-domain images and has good robustness.