Face Modeling using Morphing-based Averaging and Low-Rank Decomposition

In video surveillance, the viewing angle of face with respect to camera, called angular occlusion (also referred to as head pose) will limit system’s ability in face recognition. In this paper, a method for angular occlusion elimination in face images is proposed, which is based on image morphing. The proposed method models a frontal face from a batch of images with different head poses belonging to a person. The frontal face is modeled through spatial interpolation of input image pixels using translation functions, linear interpolation and intensity averaging. In order to improve the modeling result, the proposed method is applied to divergent face images (face imageswithsymmetric poses). Then, low-rank decomposition is employed to align the modeled faces. Radial basis function neural network is considered for translation function. The main advantages of the proposed method is that in spite of common modeling methods, depth information, calibrated images and head pose data are not required. The algorithm’s performance on PRIMA dataset is investigated. Considering that the input face images only have variation in pose, experiment results show that the proposed method will model frontal face image with properaccuracy.