Frontal face modeling using morphing-based averaging and Low-rank decomposition

Head pose variations are a major problem in face recognition. Many advanced methods exist for synthesizing frontal face images with head pose variations. These methods are generally categorized into three groups: 3D based methods which use 3D face models, path based methods which exploit linear transformations of small segments of the face, and deep learning methods. Frontalization methods usually use a single head pose image for frontal face synthesizing and heuristics for filling the hidden parts of the face in the reference model. This causes some artifact in the image, and reduces similarity between the synthesized and actual frontal face image. In this research, frontalization is done by a weighted averaging algorithm, as an extended version of the path based frontalization method powered by matrix rank minimization, which exploits from classical image domain transform. This preprocessing technique grants matrix decomposition methods the ability to eliminate head pose variations. Representing the actual face models and robustness to facial landmark detection errors are the advantages of this method compared to other existing methods. Quantitative and qualitative comparison results indicate superiority of the proposed method in face frontalization.