Abstract:

In view of the defects that the projection matrix learned from LatLRR ( Latent Low Rank Representation) can not explain the importance of the extracted features and preserve the local geometry of data, a novel method named LLRSP (Latent Low-Rank and Sparse Projection) with dual neighborhood preserving and feature selection is proposed. The algorithm first combines low-rank constraint and orthogonal reconstruction to hold the main energy of the original data, and then applies a row sparse constraint to the projection matrix for feature selection, which makes the features to be more compact and interpretable. Furthermore, a l^2,1 norm is introduced to regularize the error component to make the model more robust to noise. Finally, neighborhood preserving regularization is applied on the low dimensional data and low-rank representation matrix to preserve the local manifold geometrical structure of data. Datasets results of extensive experimental on various benchmark show that this method can obtain better performance than other state-of-the-art methods.

Key words: feature extraction, feature selection, dimensionality reduction, latent low-rank representation, image classification