Hallucination of super-resolved face images

Reconstruction-based super-resolution algorithms are widely employed for enhancing the quality of low-resolution face images. However, these algorithms are very sensitive to the registration errors of their input images. The registration errors aggravate when working with face images coming from video sequences. The longer the video the bigger is the registration error (due to the motion of the subject). Furthermore, the improvement factor of these algorithms is limited by factors smaller than two. The proposed system in this paper deals with these two problems. In order to restrict the registration errors of the system a fuzzy-based face quality assessment is employed. To cope with the second problem, a hierarchy of different types of super-resolution algorithms is used to reach an improvement factor of four. The proposed system has been tested using real video sequences of different longs and the experimental results are promising.