Superresolved Image Synthesis for Multiple-Plane Scenes Captured with an Uncalibrated Moving Camera

摘要

In this paper, the authors propose a method to take a video sequence photographed with an uncalibrated moving camera and use the projected relationship between each frame in the sequence to merge all the frames together and generate a superresolved image, that is, one with higher resolution than the source images. By assuming that the scene can be approximated with multiple plane patches, the images are merged using a homography derived from a fundamental matrix estimated across each frame for each plane patch. In this method, blurring occurs in the generated image as a result of positional errors in the images being synthesized: this is turned to advantage in an optimization process that maximizes the energy of the high-frequency components in the generated image; the correspondence among plane-patch vertex frames, needed to derive the homography, is used to estimate subpixel precision. To prove the effectiveness of this method, the authors conducted experiments with actual images and were able to improve resolution beyond the level of the original image.