Video-based face recognition is an interesting problem in recent years. In this thesis, we mainly discuss face recognition from video to video. In other words, both the training data set and the testing data set are consecutive face image sequences. The contributions of this thesis mainly include the following two aspects. 1.We do some work on online learning of face appearance models. We propose two novel online learning methods, namely K-Eigenspace Learning and T-Eigenspace Learning. For K-Eigenspace Learning, the number of eigenspace models corresponding to each person is fixed. The eigenspace models are dynamically adjusted using IPCA (Incremental PCA), eigenspace merging or eigenspace splitting. We exploit the temporal information which is embedded in a face image sequence by maintaining a transition matrix. For T-Eigenspace Learning, the number of eigenspace models corresponding to each person is variable. The eigenspace models are dynamically adjusted using IPCA. In the process of online learning, more and more samples are adding to each eigenspace except one eigenspace which contains the least number of samples. Both methods can learn appearance models completely online. They both try to learn a few eigenspace models to approximately construct the face appearance manifold for each person in the training data set. Experimental results show that both methods can achieve good recognition performance. 2. We propose a fast computing algorithm for PCA (Principal Component Analysis) dealing with large scale high-dimensional data sets. A large data set is firstly divided into several small data sets. Then traditional PCA method is applied on each small data set and then several eigenspace models are obtained. At last, these eigenspace models are merged pairwise into one eigenspace model which contains the PCA result of the original data set. We also analyze the behavior of our approach with respect to the error introduced in the computing process. We apply this algorithm on good frame selection for video-based face recognition. The recognition performance is rather good, while the computing time is much less than that of traditional PCA method.
修改评论