The Superresolution Reconstruction Based On Image Registration

Image super-resolution (SR) reconstruction can be comprehended utilizing observed low-resolution (LR) images to reconstruct a real high resolution (HR) image. It can successfully recover the information out of the cut-off frequency so that we can get more details of an image and enhance image resolution. The methods of image SR can be classified as single frame SR and multiple frame SR, and can also be classified as frequency domain and spatial domain in processing fields. SR approaches mainly include finite impulse response (FIR), fast Fourier transform (FFT), de-aliasing, etc. in frequency domain and interpolation approach of non-uniformly spaced samples, projection onto convex set (POCS), maximum a posteriori (MAP) estimate, etc. in spatial domain.Multiple frame SR can utilize different information between images and some apriority information to reconstruct a HR image. The emphasis of this paper is to research a SR reconstruction method based on image registration for which the registration algorithm is very important. The image registration firstly assumes that there is some transformation between frame and frame of images, such as translation, affine and planar surface etc. and then registers images. The algorithms of image registration are firstly introduced, which include especially frequency registration algorithm based on FT and global motion model registration algorithm based on affine transform. Whereafter, the registration results are used to SR reconstruction of image. The investigated SR method mainly is interpolation approach of non-uniformly spaced samples, which transforms the nonstandard displacement pixels into standard displacement pixels and farther realizes SR reconstruction.The experiment results are given in last chapter. The experiments are simulated by use of MATLAB. The experiments result analyses indicate that image registration plays an important role in SR reconstruction. The frequency registration algorithm based on FT can suppress frequency aliasing but cannot estimate rotation angle, and global motion model registration algorithm cannot suppress frequency aliasing, but, in which the rotation estimation is more accurate. The two methods have their own advantage and deficiency. When the...