Multispectral Face Image Registration Based on T-Distribution Mixture Model

In order to enhance the accuracy and robustness of multispectral face registration results suffering from non -rigid deformation, noise, and outliers, a multispectral face registration method based on the spatial geometrical structure and local shape features of feature points is proposed. On the one hand, we use inner-distance shape context as the local shape feature of the point set, and create the similarity measure function between visible and infrared images. On the other hand, a Student' s-T mixture model is used to represent the transformation model estimation in non -rigid point set registration process, and the model can be solved by using the expectation maximization algorithm. The simulation results show that the proposed method can realize exactly registration of point sets with deformation, noise, and outliers. The visible and infrared real image databases demonstrate that the matching error and computing efficiency of the proposed method outperform those of the comparison methods. As a result, the multispectral face images after registration and fusion will improve the performances of follow-up face detection and recognition.