A fast video super resolution for facial image

Multi-frame super resolution has found various applications in different domains of machine vision such as remote sensing, object recognition, and security applications for the last two decades. Classic super resolution methods are not able to handle real word videos where different parts of scene undergo different motions. Most of recent methods in the literature address this problem but they suffer from time complexity. In this paper, we propose a fast method for super resolution of facial videos. Our proposed method provides less computational complexity in addition to handling videos having general motion patterns. These two benefits make the proposed method suitable for security purposes. In the proposed method, first we extract a number of key points from face in each video frame. Then, for each pixel in the reference frame the corresponding pixels in other frames are determined using triangular patches. Subsequently, the obtained solution is improved by minimizing an energy function considering both appearance and pixel displacements. Super resolved facial image is finally obtained by using information available in a small window around approximated location of the low resolution frames. The effectiveness of the proposed method has been demonstrated for real video sequences.