Image Super Resolution Algorithms Based On Correlated Images

Image super resolution is always the hot problem. The resolution of imagesreceived is always low as the limited of the video super supervision, the net band,and the low amount radiation, then we always hope to resore the high resolutionimage based on the low resolution image. Therefore, many image super resolutionalgrorithm based on one or multiple low resolution images are proposed catered tothis kind of requirement,for example, the super resolution based oninterpolation,multiple low resolution inputs and machine lerning. The image superresolution is an ill-posed problem as a low input corresponds to multiple hignresolution images. With the improvement of the social networking services and thebig data, the images in the net is more abundant, which will influence the traditionalsuper resolution seriously.In this paper, a scheme for image super resolution based on correlated images isproposed. Given a low resolution input,then the correlated images are obtainedthrough retrieving.We consider the similar images as assistant information to restorethe missing datails in the high resolution images.Although the content of the inputand its correlated images is similar, they are different in illumination, perspective,shooting time or the position of the same objects caused by motion,therefore theretrieved images can not be ultilized immediately.In this paper, we align thecorrelated images to the input with RANSAC algorithm before reconstructing theimage. What is important is that we propose Error Reduction Iteration to furtherenhance the quality of reconstructed image.In summary, the content of our paper is as follow:(1) We propose the framework for the super resolution based on correlatedimages. Given a low resolution input, we will obtain an initial high res olution imagethrough up sampling. Then we will retrieve the correlated images which will bealigned to the initial image. Then we will reconstruct the initial image, and thenrestore the reconstructed image.(2) We choose MSD to be the measurement of the similarity between twoimage patches according to analyse the differences among different measurements.(3) During the restoration, algorithm Error Reducing Iteration based onminimizing gradient is proposed to deblur and get rid of the block effects efficiently.What is more, the proposed framework can be extended to many traditionalfields such as deblurring.