Research On Super-resolution Reconstruction Oriented Image Registration

With the development of multimedia technique, much higher definition is needed to satisfy the increasing demand. Super-resolution restoration of the video sequence is the digital signal processing method, which is used to improve the image resolution. It is also a kind of image enhancement technique which is used to construct one or more super-resolution images from the degraded and low resolution observed sequence. It is related to image registration, image restoration, information rendering and fusion technique. Image restoration is used to restore gray-scale resolution from the observed degraded image; image fusion is to combine information from multiple images to make the target image much clearer. Image registration which is defined as mapping between two images both spatially and with respect to intensity. Its accuracy determines the performance of super-resolution.At present, the main problems of the existing image registration algorithms applied to super-resolution are: (1) Most of super-resolution methods don't pay enough attention to image registration; (2) The existing image registration algorithms are almost based on the assumption that aside from the translation components, the other parameters vary very small in the sequence. In many practical situations, however, these assumptions are almost never met. The accuracy of image registration results can't satisfy the requirements of super-resolution when the sequence images' rotation angle and the scale factor between the images to be registered change in a large scale.This thesis, according to the problems mentioned above, proposed the algorithm of the sequence image registration based on target tracking. Firstly,the interested target area is roughly located by using the method based on SIFT target tracking,then the motion parameters are calculated, and finally these parameters are refined by using the method of Lucas. Experiments demonstrate that the proposed method can get much more accurate motion parameters when components of them (Translation, scaling and rotation) vary in a large scale.As the method of Lucas registration goes into local minimum easily, and genetic algorithm has the strategy of population search superior to the traditional searching methods and a capacity of global optimization, the algorithm of image registration based on genetic algorithm is proposed. The calculated rough motion parameter is used to generate initial population, design fitness function and construct selecting operator, cross operator, mutation operator in this paper. Experiments demonstrate that this algorithm improves registration accuracy further.To make the image registration method avoid being trapped in local minum, this paper poposed a registration algorithm which unites target tracking, generic algorithm and image registration organically, and as a result, it can improve the accuracy registration of sequence images in the situation that the motion parameters vary largely.