Super-resolution Reconstruction for Multi-resolution Image Sequence

Abstract

A blind super resolution (SR) image reconstruction algorithm based on scale invariant feature transform (SIFT) and image registration is proposed for multi-resolution image sequence taken in various focal lengths. First, SIFT keypoints in images are extracted. Then keypoint descriptors are matched initially under the criterion of vectorial angle cosine and outliers of matches are eliminated by random sample consensus (RANSAC) algorithm to improve registration accuracy. And registered low-resolution (LR) images are mapped onto a high-resolution (HR) grid according to their transform parameters. Finally, space pixels are filled in by a pixel reliability weighted algorithm to reconstruct the image with a higher resolution. Experimental results show that the proposed algorithm can estimate scaling factors accurately and it is effective in affine transformation and is robust to registration errors within a certain range. The algorithm can essentially improve the resolution of multi-resolution image sequence with relatively satisfactory reconstruction result especially under the condition when the number of low-resolution image frames is too small and available information for reconstruction is seriously insufficient.