Abstract:L1-norm maximization based Discriminant Locality Preserving Projection (DLPP-L1) is shown to be effective and robust to the outliers in given data, but DLPP-L1 is based on the vector space, so it has to convert those 2D matrices into high-dimensional 1D vectorized representations when handing images. But such transformation usually destroys the topology structures of images pixels, which can decrease performance. We therefore propose to extend DLPP-L1 to the 2D matrix space. A two-dimensional DLPP-L1, termed 2D-DLPP-L1, is technically proposed for image feature extraction. Compared with DLPP-L1 for representation, our proposed 2D-DLPP-L1 can effectively preserve the topology structures among image pixels in addition to inheriting the robustness property against noise and outliers. Extensive simulations on real-world image datasets show that our 2D-DLPP-L1 can deliver enhanced performance over other state-of-the-arts for recognition.

Robust L1-Norm Matrixed Locality Preserving Projection For Discriminative Subspace Learning