Locality-constrained Low Rank Coding for face recognition

This paper presents Locality-constrained Low Rank Coding (LLRC) as a novel approach for image classification. The widely used Sparse representation based algorithms reconstruct a test sample using a sparse linear combination of training samples. But they do not consider the underlying structure of the data in the feature space. On the other hand, Low Rank representation has been recently used for clustering face images into their respective classes by taking advantage of the low rank structure of the data. LLRC first imposes a locality constraint to choose the training samples that are in the vicinity of the test sample. Then it applies the low rank constraint on these training samples to further choose a subset that belongs to a subspace corresponding to one face class. In this manner, the training samples used to reconstruct a given test sample can be chosen from just one class rather than a mixture of classes, thus enhancing the classification accuracy. We evaluate our algorithm on face image datasets. Our algorithm outperforms sparse representation based algorithms, thus showing that exploiting the structure of data is important. We further demonstrate that both locality constraint and low rank constraint are imperative to obtain superior performance.