Abstract
In this paper, we propose a method to improve the performance of Super-Resolution for video sequence that contains moving objects by realizing Dense Motion Estimation (DME) virtually. The performance of Super-Resolution heavily depends on the accuracy of motion estimation. Therefore, if accurate DME that allocates motion vector to each pixel can be achieved, the performance of super-resolution must be remarkably improved. However, it is difficult to realize pixel-level DME accurately on the enlarged images, because observed images lost information necessary for motion estimation via degradation process. In this paper, we first apply DME using Hierarchical Block Matching (HBM), which is one of the conventional methods, to Super-Resolution, and verify that the accuracy of DME using HBM depends on the quality of observed images. Precisely, DME using HBM works effective only for high quality observed images, but the effectiveness seriously deteriorates for low quality observed images. Then, we propose a method that realizes virtual DME using Overlapped Block Matching (OBM), which is often used to reduce block noise in motion picture coding. While this method does not improve the accuracy of individual motion vector, it apparently increases the number of reference frames by allocating multiple motion vectors to a small region, which improves the performance of Super-Resolution. Through computer simulation, we show that the proposed virtual DME is robustly effective for Super-Resolution irrelevant to the image quality of observed images.
