《基于重构的超分辨算法的理论极限》青少年教育文学丛书.pdfVIP

On the Fundamental Limits of Reconstruction-Based Super-resolution Algorithms Zhouchen Lin Heung-Yeung Shum Microsoft Research, China fzhoulin jhshumg@ Abstract In practice people have found that the improvement in Super-resolution is a technique that produces higher reso- the resolution is quite limited. Current algorithms, not lim- lution images from low resolution images (LRIs). In prac- ited to RBAs, can only produce overly smooth images or tice, people have found that the improvement in resolution images with undesirable details when the magnification fac- is limited. The aim of this paper is to address the problem tor becomes large. “do fundamental limits exist for super-resolution?”. Specif- It is important to address the problem “do fundamental ically, this paper provides explicit limits for a major class limits exist for super-resolution?” so that practitioners can of super-resolution algorithms, called the reconstruction- avoid trying those unduly large magnification factors. Fur- based algorithms, under both real and synthetic conditions. thermore, a better understanding would help people choose Our analysis is based on perturbation theory of linear sys- good magnification factors under the limits so as to make tems. We also show that a sufficient number of LRIs can better use of system resources. Towards this problem, Baker be determined to reach the limit. Both real and synthetic and Kanade [2] showed that both the condition number of experiments are carried out to verify our analysis. the system and the volume of solutions grow fast w