Discriminative Dictionary Learning Based on Sample Diversity for Face Recognition

摘要

Dictionary learning algorithms applied to face recognition always suffer from the following problems. It is difficult for face databases to provide sufficient training samples, which hinders algorithms from extracting reliable atoms. Then, facial images are susceptible to external conditions. And even the facial images from the same individual vary with facial poses and expressions. That is the reason why most of the dictionary algorithms are not robust. Moreover, the discrimination of algorithm is limited due to ignoring the locality characteristics. We proposed a novel discriminative dictionary learning algorithm framework based on sample diversity to solve the above problems. First of all, the skillfully generated virtual face images can enrich sample diversity and alleviate the small sample size problem. Secondly, new error constraints are added to the elaborate objective function to restrain outliers from the training samples and make algorithm robust. Thirdly, the local information of atoms is preserved via the graph Laplacian matrix of dictionary instead of directly using the training samples, which aims to enhance the discrimination of the dictionary and reduce the influence of noise. Experimental results show that the proposed dictionary learning algorithm framework can achieve higher performance level than some previous state-of-the-art algorithms.