Low-rank decomposition and Laplacian group sparse coding for image classification

This paper presents a novel image classification framework (referred to as LR-LGSC) by leveraging the low-rank matrix decomposition and Laplacian group sparse coding. First, motivated by the observation that local features (such as SIFT) extracted from neighboring patches in an image usually contain correlated (or common) items and specific (or noisy) items, we construct a structured dictionary based on the low-rank and sparse components of local features. This dictionary has more powerful representation capability. Then, we investigate group generation for group sparse coding and introduce a Laplacian constraint to take into account the interrelation among groups, which can maintain low reconstruction errors while prompting similar samples to have similar codes. Finally, linear SVM classifier is used for the classification. The proposed method is tested on Caltech-101, UIUC-sports and Scene 15 dataset, and achieves competitive or better results than the state-of-the-art methods.