시공간 평탄화 제약 조건을 이용한 동적 비디오에서의 초해상도 영상 복원

초록

This paper proposes an efficient super resolution method based on iterative back projection (IBP). The existing super resolution methods usually have focused on the videos containing global motion only and they also have high computational complexity. Therefore, for real-time super resolution over video sequences, these methods are not suitable. However, iterative back projection method by Irani and Peleg is a stable and less demanding approach which is appropriate for video sequences. Therefore, the proposed method is based on iterative back projection. However, they still have some drawbacks, the artifacts owing to incorrect motion estimation and poor quality than reconstruction-based methods. To solve these problems, the proposed approach is explained below. In contrast to conventional iterative back projection methods, the proposed method enforces the spatio-temporal smoothness prior based on Huber Markov random field (MRF) image prior. While iterative process to estimate high resolution image is performed, the amount of the spatio-temporal constraints is updated simultaneously. As a result, the reconstructed image is much smoother and more continuous. In addition, to reduce artifacts caused by incorrect motion estimation we applied bilateral filtering to temporal edge based optical flows between adjacent frames. We also use elliptical weighted area (EWA) filter whose kernel shape and size are determined by the motion vector so that point spread function (PSF) is approximated more accurately. The proposed algorithm is applied to both synthetic and real sequences and compared with the results of the other reconstruction methods. We confirmed that the proposed method is highly robust to inaccurate motion estimation and produces high resolution images with preserved edge.

키워드