Single Image Super-resolution Based On Regression And Image Self-similarity

Super resolution(SR)reconstruction is to get a high resolution image from one low resolution image or multiple low resolution images of the same scene taken at sub-pixel misalignment via image model and degraded prior etc.It plays an important role in many applications domains,such as: remote sensing,military,medical imaging,public security and so on.Super resolution reconstruction is an ill posed inverse problem in mathematics,and its research has not been interrupted.At present,there are three methods based on interpolation,reconstruction and learning for solving SR.This paper focuses on the study of single image SR reconstruction method based on self-similarity of image and regression.The main contributions of this paper include the following:(1)We propose a new single image SR reconstruction framework,which is based on image self-similarity and support vector regression(SVR)that is good at fitting data via nonlinear mapping.The proposed method is without external training database,instead,the image pyramid of the input low resolution(LR)images is established and the set of low /high resolution image patch pairs is training set.We assume that the map relationship between these similarity LR image patches and the pixel value of the center of the corresponding high resolution(HR)images is the same.The support vector regression model is used to study the mapping relationship.the proposed method is tested on numerical experiments and compared with several kinds of classical super-resolution methods.The experiment results show that the proposed method can obtain good super resolution effect,especially for the image with high similar texture.(2)Another method based on direct sparse kernel regression(DSKR)for single image super-resolution is proposed to overcome the shortcomings of the above proposed method based on support vector regression.This method aims at looking for "better" support vector to being closer to the weight distribution of similar image patch.The experimental results show that the super resolution effect is indeed improved.