Singular value decomposition based sample diversity and adaptive weighted fusion for face recognition

The performance and robustness of face recognition are largely determined by the data samples used for model training. To obtain more representative samples of a face, this paper proposes a novel approach to acquire two groups of virtual samples from the right singular vectors as well as from left singular vectors via singular value decomposition (SVD) for each class of training samples. The generated virtual images not only enrich training samples but also obtain more representative information of faces, therefore higher accuracy of face recognition is achieved. Furthermore, we propose a simple and effective method that automatically determines adaptive weight without any manual intervention for three groups of scores, including the original samples and two groups of virtual samples. The weighted score fusion scheme is able to offer more supplementary information from multiple sources and obtain better performance in face recognition. Experiments on three benchmark datasets demonstrate that our proposed method is robust and obtains better accuracy for face recognition compared with previous methods.