Research And Implementation Of VIS/NIR Face Recognition

Face recognition technology has received more and more attention in the fields of computer vision and pattern recognition, and gradually become a hot topic. Now, most face images are taken under visible light conditions, and in visible light environments, lighting conditions are usually complex and variable. The performance of face recognition is affected by the variable illumination, thus overcoming the influence of illumination changes has become an important issue in face recognition. As the near-infrared light image is robust to the illumination changes, the near-infrared light imaging technology is widely used to solve the problem. Enrolled and query face images both are captured under near-infrared lighting, but in the practical application, most face images are taken under visible light conditions, such as ID photos and so on. Therefore, a new problem, cross recognition between near-infrared and visible lighting face images, has emerged. Because of their different imaging methods, visible lighting and near-infrared face images from the same person is significantly different in appearance, but they should be recognized as the same person in the view of human cognition, which means that there are underlying correspondences between near-infrared and visible lighting face images from same person.In this paper, there are two key points:1. Feature fusion. After reading lots of literature and making some experiments, we find that there are three kinds of features which have a good performance in visible/ near-infrared face recognition, namely SIFT, LBP and HOG features. There are two existing techniques of feature combination:serial and parallel combination. In our method, we get a new representation which fuses three features with weighted serial feature combination method. According to the result of experiments, we have a better performance with the new representation than single feature.2. Multitask clustering algorithm based on ELM. Extreme Learning Machine is a simple and effective Single-hidden layer feed forward neural network learning algorithm. Instead of learning with one single task, multitask learning is a machine learning algorithm, which has more than one task. We get the inherent commonality between different tasks to avoid poor fitting caused by less training and improve the generalization performance. In this paper, a method called Multitask Clustering based ELM is proposed. Firstly, face image data is projected to low-dimensional space with ELM mapping method, then the data is recognized with multitask clustering method. Finally, verifing its performance of the method through a series of experiments.