ABSTRACT
Support Vector Regression (SVR), a typical learning tool, is considered as an extension of Support Vector Machine (SVM), which exhibits excellent generalization ability in predicting functional outputs without any prior knowledge or assumption on the training data. In general, SVR is capable of forecasting data via either linear or nonlinear mapping, and it is widely used in applications of data mining, financial forecasting, etc. Previously, SVR has been shown to address SR problems (Glasner. 2009, Ni. 2006, Ni. 2007, An. 2011). In our method, first, we do not aim at estimating the high frequency component of the LR image to be super-resolved. Instead, the prediction of our algorithm is the pixel value itself. Second, we extract smaller patches in the process of training the model, and regression is more exact. Third, differing from the work of Ni (2006) in the DCT domain, we add the average pixel value of each block after the DCT transform, and it makes the model more
1 INTRODUCTION
With the rapid development of image processing, algorithms applied to surveillance systems, medical image and video sequences are gradually extensive, especially in the field of super-resolution image. It is a process to produce a High-Resolution (HR) image from one or several Low-Resolution (LR) images. It is desirable to obtain as much HR image information as possible, because a large amount of information had been lost in the process of acquiring low resolution version. The imaging process of sensors can be modeled by
y = DBx + n (1)
where x is the high-resolution image that undergoes Blurring (B) and Downscaling (D) procedures with additive noise n. The output is a low-resolution image y that we often observe.
