Slip/antiblocking agents

Abstract

Dictionary learning (DL) methods are widely used for pattern recognition in recent years. In most DL methods, the l1 norm is employed to promote sparsity of the coding. However, the usage of the sparse coding based methods is limited since solving the l1 based sparse coding is very time-consuming. In this paper, a novel orthogonal collaborative dictionary learning (CDL) method is proposed for accurate and efficient face classification. In this method, several class-specific dictionaries and one common dictionary are learned jointly from the training data, where the class-specific dictionaries are used to model the appearance of the subjects and the common dictionary is used to model the facial variations. To learn these dictionaries, we introduce an orthogonality promoting term to encourage the facial variations to be independent of the appearance as much as possible, and introduce a scatter constraint term to remove the variations in the class-specific dictionaries. Since CDL can derive analytical solutions for both code learning and dictionary updating, it is much more efficient than many other DL methods in terms of training and classification. Experiments conducted on seven face databases show that CDL outperforms many state-of-the-art DL methods and coding methods in both accuracy and efficiency.