Tensor-based locally maximum margin classifier for image and video classification
Computer Vision and Image Understanding
| Hi-index | 0.01 |
Learning discriminative subspaces for multimodality image analysis has been an extensively discussed topic over the past several decades in interdisciplinary fields. It has attracted much attention recently due to the increasing demand for developing real-world human-computer interaction systems. A large family of subspace learning methods has been designed based on different motivations and objective functions. Although they are diversified, it is intuitive to uncover some common ideas among them. Can we unify them and formulate new algorithms to further enhance the discriminating power of features extracted by those subspaces? Stemming from this motivation, in this dissertation, a unified framework of discriminative subspace learning is presented. This framework is composed of three levels in a top-down manner. In the first level, it defines basic subspace learning strategies, global learning or local learning. In the second level, the first-level strategies are embodied in the sample space, feature space, and learning space to provide key concepts. In the third level, four particular criteria—manifold learning, Fisher graph, similarity metric, and high-order data structure—are defined to specify the top two levels for algorithm design. The concepts and criteria in the third level can be applied separately or jointly to conduct the algorithm design constrained by the upper levels. To demonstrate the effectiveness of the framework, an expert model of the query-driven locally adaptive (QDLA) method and four new subspace learning algorithms corresponding to different learning-locality criteria are presented. These four algorithms are locally embedded analysis (LEA), discriminant simplex analysis (DSA), correlation embedding analysis (CEA), and correlation tensor analysis (CTA). Extensive experiments demonstrate that applying the local manner in the sample space, feature space, and learning space can sufficiently boost the discriminating power for feature extraction by the subspace learning. As an advanced extension, a learning-locality based subspace learning algorithm for multiple/multimodality feature fusion is also developed in both unsupervised and supervised learning cases. Those methods are successfully applied to several real-world applications of facial image computing, such as face recognition, head pose estimation, realistic expression/emotion analysis, human age estimation, and lipreading.