| Over the past several decades, face recognition has been the most important topicin the field of computer vision and artificial intelligence, which always draws muchattention of numerous researchers. With the increasing popularity of such application asvideo surveillance and sharing, it is more and more attractive to the technique of facevideo-based recognition, which can learn more space-time information within videos toimprove the accuracy rate and has been one of the hottest topics in face recognition area.Aware of the superiority of manifold learning on face recognition, the richer cognitivefeature in video face sequences than in static face images and the high dimension offace data, we naturally realize that manifold based algorithms can play a better role inface recognition for video. Based on the research to the optimization of manifoldlearning, it is of great use to find an effective way for manifold based methods to solvethe specific problems of manifold learning in face recognition for video. In this paper,we mainly propose two algorithms for video-based face recognition, one is graphembedding algorithm with its sparse variance and another is multi-manifold analysisalgorithm.Due to the various appearances, resulting from changing illumination, partiallyocclusion and so on, face images show multi-modal structure, which means there areseveral sub-clusters within one class. How to make use of this information to overcomethe difficulty existed in video-based face recognition is the key point. Traditionalsupervised graph embedding algorithms neglected the congener correlation betweenmulti-modal sub-clusters in the same class and inaptly incorporated discriminativeinformation between different classes. So we design a novel graph embedding algorithmto overcome this problem. Not only can it uncover the intrinsic manifold information offace data, but also make use of the semantic space-time information hidden in the videosequence. Furthermore, a sparse embedding is achieved by using L2;1-norm in LDAframework, so the proposed algorithm can select relevant features and learntransformation simultaneously. Moreover, considering that face data possesses multi-manifold structure, whiletraditional multi-manifold learning algorithms aim to learning one subspace, we proposea novel multi-manifold algorithm to learn multiple subspaces to preserve moremulti-manifold information and make the different manifolds more discriminative.After multiple manifold subspaces learned, a metric is applied to compute the similarityof different manifold subspaces.Experiments demonstrate on UCSD/Honda video database that the algorithms canachieve better performance than the state-of-arts reported in the recent literature. |
