Muti-frame Image Super Resolution Based On Modified Random Search Patch Match And Adaptive Kernel Regression

Multi-frame image super-resolution reconstruction refers to reconstruct a corresponding high-resolution image from the existing low-resolution image sequences in which a bunch of redundant information exist. This technology can surmount the intrinsic drawbacks of hardware, and improve the image super-resolution to improve visual effect and to prepare for the subsequent image processing only by the image processing method of software without improve hardware conditions of the imaging device.This paper focuses on the application of random search patch match algorithm and adaptive kernel regression algorithm in multi-frame image super resolution reconstruction. In the first two chapters, we introduced the history and current situation, as well as basic principles and key technologies of super resolution reconstruction. In chapter three, we introduce the basic theory and applications of block matching algorithm, and we introduce random search patch match algorithm in detail and verify the validity of this approach through experiments. Given the fact that some inherent disadvantages exist in traditional random search patch match method, we made some improvements to make it more suitable for applying in super resolution reconstruction. Those modification includes: use registration parameters as initialization prior, constraint the search area size, scale and rotation range restriction, store the k nearest neighbors and add multi-neighbor propagation. In chapter four, we introduce the theory of classical kernel regression and analysis its’ property by experiments. We then apply the modified version on classical multi-frame image super resolution reconstruction algorithm based adaptive kernel regression, and successfully make the reconstruction process accelerated. We also propose a kind of robust kernel regression based on curvature difference when there exist outliers in data set.In those part where experimental results are presented of each chapters, we compare the experimental outcome of different kinds of block matching algorithm to emphasis the effectiveness of random search patch match algorithm and modified version at first. Then we accelerate the adaptive kernel regression based super resolution algorithm using the modified RSPM algorithm. We improve effect of super resolution by adding curvature weights at the end. Finally, Experimental results on simulated LR image sequences demonstrate the effectiveness and robustness of the proposed method.