基于L_(2,1)模和图正则化的低秩迁移子空间学习

L_(2,1)-norm and graph-regularization based low-rank transfer subspace learning

本文提出一种基于L_(2,1)模和图正则化的低秩迁移子空间学习方法.首先,在低秩重构过程中通过对重构矩阵施加具有旋转不变性的L_(2,1)模约束,可在挖掘目标域数据的关键特征的同时提高算法对不同姿态图片分类的鲁棒性.其次,在目标函数中引入图结构的正则化,使得迁移时数据中的局部几何结构信息得以充分利用,进一步提高了分类性能.最后,为解决源域数据较少导致的欠完备特征空间覆盖问题,在公共子空间中利用源域数据和目标域数据联合构造字典,保证了重构的鲁棒性.在Caltech256, Office, CMU–PIE, COIL20, USPS, MNIST, VOC2007和MSRC数据库上的大量对比实验验证了本文方法的有效性和鲁棒性.

A novel L2;1-norm and graph-regularization based low-rank transfer subspace learning method was proposed in this paper. Firstly, by applying the L2;1-norm constraint on the reconstruction matrix during low-rank reconstruction, the key features embedded in the target domain can be better explored. In addition, the rotation invariant characteristic of L2;1- norm will gives the algorithm the capability of handling the images with different poses. Secondly, the graph-regularization was integrated in the object function to better utilize the local geometric information embedded in the training data. As a result, the classification performance can be further enhanced. Finally, to tackle the problem of incomplete feature space coverage problem result from the insufficient source domain data and ensure the robustness of reconstruction, we advocate grouping the target domain and source domain data to form a joint “dictionary”. Extensive experiments on Caltech256, Office, CMU–PIE, COIL20, USPS, MNIST, VOC2007 and MSRC dataset validate the effectiveness and robustness of our algorithm.