Image Registration And Reconstruction In Super Resolution

High-resolution (HR) images or videos are necessarily used for illustrating object's details and motion process in many digital media related fields. However what people can get in real application is usually a sequence of low-resolution (LR) images, in which the key information cannot be displayed properly to meet the requirements. Generation of super-resolution(SR) image can fit this popular demand very well. It uses the aliased information of multiple LR images with a different motion of the same scene to reconstruct the "super" resolution image which has even higher resolution than the original one. Its application fields include high-reality image and animation, HDTV, battleground simulation, information security etc., and it is quite meaningful for theoretical studies and application.The generation of super resolution image is composed of two steps: step 1 is to register each LR image using sub-pixel level image registration technology, and step 2 is to take advantage of the super resolution reconstruction algorithm to reconstruct the HR image based on the irregular sample points. The result of the SR reconstruction will be affected by the accuracy of the registration directly. Currently, frequently-used registration methods can be classified as two categories: frequency domain based method and spatial domain based method. Both types have their own advantages: frequency domain methods have strong robustness, while spatial domain methods have higher accuracy.Based on previous researches and by focusing on image registration techniques, this dissertation puts forward a combined registration method of both frequency-domain and spatial-domain method, which is also applied to SR reconstruction to improve the results.First of all, image registration algorithms are studied and the main purpose and classification of image registration are described. The different theory evidence, realization means, algorithm models and limitations are also analyzed. Further study is made on both typical frequency-domain method (by Vandewalle et al.), and spatial-domain method by Keren et al. Improved methods are also presented here. In this paper, it is the first time for us to add the pyramid model to the algorithm by Vandewalle et al. to increase the accuracy by coarse-to-fine registering layered down-sampled images. The estimation for the rotation parameter is improved as well as the robustness. In addition, another combined algorithm based on that of Vandewalle et al. and that of Keren et al. is also proposed, which combines both the advantages of frequency-domain and spatial-domain methods. In this method, frequency domain registration is first used on the images, and then the estimated parameters are used to rotate back the original images, followed by the spatial domain registration algorithm. Improved algorithm has significantly reduced the errors of the registration when the relative translations are large, which not only expands the scope of application, but also guarantees the accuracy.Secondly, a new evaluation method for the registration algorithm has been proposed in this paper, which can effectively estimate the accuracy of the algorithms. It can be applied to algorithm comparison and research.In conclusion, the analytical algorithm discussed in this dissertation is applied to the SR reconstruction process. Comparing the result of two different reconstruction algorithms, Robust SR and CFA reconstruction algorithm, we find that the CFA algorithm performs better.