ThesisFace Recognition Robust to Occlusions
Face recognition is an important technology in computer vision, which often acts as an essential component in biometrics systems, HCI systems, access control systems, multimedia indexing applications, etc. In recent years, identification of subjects in non-controlled scenarios has received large amount of attentions from the biometrics research community. The deployment of real-time and robust face recognition systems can significantly reinforce the safety and security in public places or/and private residences. However, variations due to expressions/illuminations/poses/occlusions can significantly deteriorate the performance of face recognition systems in non-controlled environments. Partial occlusion, which significantly changes the appearance of part of a face, cannot only cause large performance deterioration of face recognition, but also can cause severe security issues.
In this thesis, we focus on the occlusion problem in automatic face recognition in non-controlled environments. Toward this goal, we propose a framework that consists of applying explicit occlusion analysis and processing to improve face recognition under different occlusion conditions. We demonstrate in this thesis that the proposed framework is more efficient than the methods based on non-explicit occlusion treatments from the literature. We identify two new types of facial occlusions, namely the sparse occlusion and dynamic occlusion. Solutions are presented to handle the identified occlusion problems in more advanced surveillance context. Recently, the emerging Kinect sensor has been successfully applied in many computer vision fields. We introduce this new sensor in the context of face recognition, particularly in presence of occlusions, and demonstrate its efficiency compared with traditional 2D cameras. Finally, we propose two approaches based on 2D and 3D to improve the baseline face recognition techniques. Improving the baseline methods can also have the positive impact on the recognition results when partial occlusion occurs.
In this thesis, occlusion-robust face recognitions are proposed for different scenarios, so that identification of subjects in non-controlled environments (i.e. video surveillance) becomes more robust in practical conditions.