Research For Image Super Resolution Reconstruction Algorithm

Super-resolution image reconstruction is the use of digital signal processing technology from a number of frequency domain aliasing, blurred and additive noise effect lower resolution images to get high-frequency information and more pixel values. So first of all, we should have a series of lower resolution images that just for one scene and exist a little difference among them. The super resolution would use this different information to improve resolution of image that we wanted.In this paper, we will introduce the basic imaging model of charge couple device and sub-pixel image registration for rigid transformation in details, then analysis and discussion the regulation algorithm of super resolution reconstruction. There are two points that we should considered in image registration: computational complexity and accuracy of registration parameters. In this paper we use the combination of Frequency domain registration and the registration that base on rectangular pixels method to reduce the computation and to achieve sub-pixel registration.In image reconstruction, when the iterative method that based on the Maximum a posteriori (MAP) algorithm being used, we need a priori knowledge on images reconstruction to get a stable solution. In this paper we use steering kernel regression method as a regular term for super resolution reconstruction. It will depend on the direction of their own data information and select the variable and regular window function to get the better results than the algorithms that base on the bilateral filter and the Laplacian as a regular term.In this paper we discuss the step size and initial value in details. Simulation results show that when low resolution images of the noise variance is larger, the algorithm proposed in this paper has a results than algorithms that based on the Laplacian or bilateral filter operator. When we choose iterative step size, the L-curve method has advantage than the variable step size methods and method based on the noise variance. The reconstruction has considerable effect at the same time, faster convergence.