Regularized Negative Label Relaxation Least Squares Regression for Face Recognition

摘要

Least squares regression (LSR) is widely used for pattern classification. Some variants based on it try to enlarge the margin between different classes to achieve better performance. However, the large margin classifier doesn't work well when it deals with the complex applications in the real world, such as face recognition, where images are captured with different facial expressions, lighting conditions or background. To address this problem, we propose a regularized negative label relaxation least squares regression method with the following characteristics. First, we introduce a negative ε dragging technique to relax the strict binary label matrix into a slack label matrix, which has more freedom to fit the labels and reduces the class margins at the same time. Second, we introduce manifold learning and class compactness graph to devise a regularization item to preserve the intrinsic structure of data and avoid the problem of overrating. The class compactness graph can enable samples from the same class to be kept close together after they are transformed into the slack label space. The algorithm based on L2-norm loss function is devised. The experimental results show that our algorithm achieves better classification accuracy.