Discriminant Manifold Learning via Sparse Coding for Robust Feature Extraction

Most off-the-shelf subspace learning methods directly calculate the statistical characteristics of the original input images, while ignoring different contributions of different image components. In fact, to extract efficient features for image analysis, the noise or trivial structure in images should have little contribution and the intrinsic structure should be uncovered. Motivated by this observation, we propose a new subspace learning method, namely, discriminant manifold learning via sparse coding (DML-SC) for robust feature extraction. Specifically, we first decompose each input image into several components via dictionary learning, and then regroup the components into a more important part (MIP) and a less important part (LIP). The MIP can be considered as the clean portion of the image residing on a low-dimensional submanifold, while the LIP as noise or trivial structure within the image. Finally, the MIP and LIP are incorporated into manifold learning to learn a desired discriminative subspace. The proposed method is general for both cases with and without class labels, hence generating supervised DML-SC (SDML-SC) and unsupervised DML-SC (UDML-SC). Experimental results on four benchmark data sets demonstrate the efficacy of the proposed DML-SCs on both image recognition and clustering tasks.