Efficient face recognition using tensor subspace regression

For face recognition, traditional appearance-based methods represented raw face images as vectors and employed dimensionality reduction techniques to capture the structure of the face space. The tensor subspace analysis (TSA) algorithm uses second order tensor to represent face images and assumes the face images reside on or close to a submanifold embedded in the tensor space. The effectiveness of TSA has been demonstrated. However, TSA is time consuming since it needs to solve a series of eigen-problems. In this paper we propose a novel efficient appearance-based face recognition method called tensor subspace regression (TSR). Like TSA, we also represent face images in tensor spaces. The difference is that we cast the facial subspace learning (i.e. dimensionality reduction) problem in a regression framework which avoids the high computational eigen-step. We show the efficiency of our algorithm by analytically and empirically comparing it with TSA. Finally, experimental results on three popular facial databases show that our algorithm can also achieve acceptable performance for face classification and clustering.