Research On Ensemble Learning Based Face Recognition

With the rapid development of information technology and continuous in-depth researches in recent years, face recognition is one of the intensive research topics of image processing and analysis, human-computer interaction and pattern recognition. Because of its non-invasive, later investigated and other advantages, it becomes the most important biometric identification technology. Efficient face recognition algorithms have become an important guarantee of public safety, information security and financial security. Improving the performance of face recognition algorithm is crucial. Ensemble learning has gradually become the first of the four machine learning research since1990s. It can effectively improve the generalization of machine learning algorithms. We use ensemble learning to improve the performance of face recognition.Illumination variance can change face appearance hugely. The similarity of the two face images of the different people may be even greater than the similarity between the two images of the same person under different illuminations. Infrared face recognition which can inhibit the influence of Illumination variance attracts more and more attentions. But thermal infrared face recognition is sensitive to the temperature of the envirienment and body. Near-infrared face recognition is robust against illumination variance and temperature, so it is the major method to inhibit the influence of Illumination variance. Near-infrared face images may lose some facial features compared to visible images, which cause some loss of recognition performance. First work of this thesis is to ensemble near-infrared face recognition and visible face recognition on Score-Level. In experiment part, we use two heterogeneous face databases to verify the effectiveness of our method. The experimental results confirm that the effectiveness of this ensemble is good.Biometrics has large variety of applications. There are three kinds:civilian application, forensic applications and high security applications. Researches for biometrics in high security applications have not attracted as adequate attention as civilian or forensic applications. We give a systematic analysis and name the problem to be solved in order to meet the performance requirements for high security applications a Double Low problem. There is no general resolution for double low problem. Second work of this thesis is to propose a hybrid ensemble framework to solve this problem. The framework has two parts:serial and paiallel. In serial part, all classifier is used in order until one of them output the right result. In paiallel part, Rank-Level fusion is used to fusion all the ranks output by the classifiers to get the final result. The experimental results in multiple databases show that this framework can achieve good results in face recognition.