A study on fast learning-based super-resolution utilizing TV regularization for HDTV

Abstract

In this paper, we propose a fast learning-based superresolution image reconstruction utilizing the Total Variation (TV) regularization method by eliminating redundancy of the reference database. We have achieved 114 times faster computational time compared with that of an ordinary learning-based method. It has been generally considered that the learning-based approach is difficult to be applied to the motion pictures because of its large computational time. We have implemented our system on the CELL processor, and studied a feasibility of applying our system to HDTV receivers. The computational speed we have obtained on the CELL processor is 202 times faster than that of the standard PC. This result indicates a possibility of applying our learning-based superresolution system to HDTV receivers.

Image quality improvement for learning-based super-resolution with PCA

“…We obtained excellent result in terms of image quality improvement, but this method had a long processing time because of the learning-based method. Therefore, we proposed a database reduction technique for learning-based super-resolution [5] where similar patches are removed from the database, which produced highspeed processing. However, there was a problem where noise appeared in reconstructed images because of the excessive reduction of database redundancy.…”