| 英文摘要 | As an important branch of biometrics, face recognition has broad application prospects in national public security and information industry，and at the same time has become one of the research hotspots in computer vision and pattern recognition. In real world scenarios, the performance of face recognition is generally affected by uncontrolled factors, such as illumination, expression, pose and acquisition conditions etc, limiting its application in practical situations. Hence, it is extremely urgent to improve the robustness of face recognition system towards environment changes. In this thesis, we study on the issues of unconstrained face recognition. We focus on several typical applications and problems, and start a research from the perspective of feature learning. The main work and contributions include: (1) For the application of unconstrained face verification, we propose a novel feature learning method by exploiting modular Deep Neural Networks (DNN). We employ DNN as the basic model, and assemble all the DNNs from each local region to form a modular DNN group on the face image. In order to get proper model parameters, we first employ the layerwise pre-training scheme to get the initial network weights. Then we combine deep learning with side information constraints in face verification, and design an appropriate optimization objective to fine-tune the whole network. Experimental results show that our method can obtain high-level representations, and overcome the variations of facial images caused by many unconstrained factors effectively. (2) For the application of Still-to-Video (S2V) face recognition, we propose a Point-Manifold Discriminant Analysis (PMDA) framework. In real S2V scenario, only a few high resolution images are registered for each subject, while the probe is video clips of complex variations. In order to model the data distribution in different scenarios effectively, the algorithm learns separate mappings for different scenario patterns (still, video). Concretely, by modeling each video as a manifold and each image as point data, we form the scenario-oriented mapping learning as a PMDA based optimization framework. Finally, the scenario-oriented mappings pursue a common discriminant space where samples from different scenarios have good clustering property. (3) For the application of cross-view face recognition, we propose an associate appearance manifolds model to learn the connection of faces under different views, by employ... |
修改评论