Research On Regression-based Super-Resolution Image Reconstruction Technique

Image super-resolution reconstruction refers to the technique that reconstructs a high resolution (HR) image from a single low resolution (LR) image or a few LR images. It has wide prospects in areas of security monitoring, military affairs, and medical science and so on.First, this paper describes the general image observation mode l, the fundamental principle of the super-resolution reconstruction and the classical algorithms in spatial domain, and introduces the regression theory briefly.Then, this paper has a discussion about the image interpolation methods often used in super-resolution reconstruction and focuses on the soft-decision adaptive interpolation (SAI) algorithm. With SAI and the non-local spatial constraint, an improved image interpolation method is proposed. Experimental results show that the proposed method preserves the edge structure well and improves the texture interpolation quality of some images.After that, the iterative back-projection (IBP) super-resolution algorithm is depicted in detail and a statistical empirical approach is proposed to determine the times and step lengths of the iterations. Due to the reasons that back-projection kernel is isotropic and the image edges are anisotropic, there exists jaggy artifact in the reconstructed image, especially around the edges. According to the artifact above, this paper proposed two improved methods.The first one is to exploit the structure information of the observed LR image to direct the back-projection kernel anisotropically. And the structure information is obtained by using the local autoregressive model.The other one is using the bilateral filter to enhance the edges of the image, which is reconstructed by traditional IBP method. With another IBP, the details filtered by the bilateral filter can be restored. Considering the property of bilateral filter and the real-time requirement, an edge extraction method is proposed. On the premise that the image quality will not be degraded, bilateral filter is used only to the edges to achieve fast adaptive reconstruction. Both methods can improve the reconstruction quality by effectively eliminating the jaggy artifact.Finally, the paper is concluded with future work.