Linear Subspace Face Recognition Algorithms And Attitude Research

In the past few decades,automatic face recognition has attracted great interest and wide attention in the research communities covering image processing, computer vision, neuroscience, statistics, pattern recognition, ect.. With more and more emerging technologies, automatic face recognition has been gradually extended to the applications of commercial and public safety where it plays a significant role in static matching of controlled format face image, e.g. passport, ID, drive license, and also face recognition based on real-time surveillance. At present, when the adopted face image is obtained in a controlled manner, e.g. high resolution, frontal, or proper illuminance, the achieved recognition performance is acceptable. However, when the adopted face image is obtained from uncontrolled condition, e.g. nonfrontal, low resolution, bad illuminant condition, the recognition performance will be severely reduced, sometimes even lower than 30%. Thus, we can see that there are still many problems with face recognition yet to be solved, and these problems are the first to be faced in practical uses.This paper mainly focuses on two problems in the field of face recognition: face recognition algorithm based on linear subspace learning, pose invariant face recognition.This paper first presents an introduction of face recognition algorithms based on linear subspace learning, where several methods are explained in detail, e.g. Principal Component Analysis, Linear Discriminant Analysis, Canonical Correlation Analysis and etc. On top of this, a new face recognition algorithm based on linear subspace learning, i.e. Two-Dimensional Discriminant Canonical Correlation Analysis, is proposed. This algorithm incorporates the concept of second order tensor into the original Canonical Correlation Analysis, which efficiently avoids the singularity problem of covariance matrix and at the same time greatly reduces the computational complexity.Pose invariant face recognition is one of the main problems in the field of automatic face recognition. This paper presents a brief introduction of the face recognition algorithms corresponding to different pose changes. Due to the nonlinear transforms of face image under pose changes, algorithms based on patches become a better choice for the face recognition problems under pose changes. On top of the method of locally linear regression and Gaussian probability model, a new face recognition algorithm based on patches is proposed, i.e. Weighted Similarity of Patches for Pose Invariant Face Recognition. This algorithm uses the method of locally linear regression to generate virtual frontal face images, and uses PSNC to evaluate the quality of the generated images. Then by employing the Bayesian posteriori probability, the probability that every patch pair comes from the same person is calculated using Gaussian probability model. The discriminative power of every patch is defined such that the probability that all patches come from the same person can be derived. The obtained sum probability denotes the probability that the whole images come from the same person, which determines the final decision.