Research On Example-Based Single Image Super-Resolution

With the application of internet technology and development of multimedia technique, image information is playing a more and more important role in our daily work, study and life. In the digital image processing area, super-resolution (SR) technology is highly valued, in order to obtain a high quality image.So far, the researching emphasis of SR is laid on spatial algorithm. However, the disadvantage of spatial algorithm makes its flexibility limited and at the same time makes its computing complicated. In conclusion, the restriction of the space and model is considered as bottleneck of its stability and veracity. In this thesis, we focus on single image example-based algorithm on the basis of traditional algorithms'analysis. We are supposed to improve SR's capability, to make the computing easier and faster, and to deliver a better result.Main work of this thesis: Summarize and compare the advantage and disadvantage of several traditional SR algorithms. Example-based algorithm turns out to be very necessary and essential. After analyzing several different dimension-reduction methods, we come to the conclusion that the LLE (Locally Linear Embedding) is the theoretical basis of single image SR algorithm. It makes the Example-based algorithm impossible. Successfully put the speedy one-pass algorithm into practice through markov network and achieve ideal experimental result. To evaluate the performance of the algorithm objectively, a normalized correlation evaluation function is developed. Experimental results are provided to illustrate the performance of the proposed algorithm done on the Visual C++ 6.0. Analysis and experimental verifications are carried out on some factors effect on super-resolution performance, especially the amount of training setting and K number of neighborhood. This analysis lays a foundation for studying new super-resolution restoration algorithms and for improving the existing algorithms.