Research On Image Super-resolution Reconstruction

Since CCD and CMOS image sensors have been widely used, people already have got high quality images. However, due to hardware cost and fabrication complexity limitations, how to increase the current resolution level by applying low cost software tools has become a topic of very great interest. On the one hand, in the traditional single image restoration problem only a single input image is available for processing, which result in the limited capability of resolution improving. Then a new technique, super-resolution reconstruction (SRR) is required. On the other hand, a low cost and easy used approach is indispensable to obtain high quality images in many practical cases, including safety monitoring, health diagnosis and military surveillance.This paper focuses on the algorithm robustness to model errors. The paper mainly studies the approach to obtain a high resolution (HR) image from observed multiple warped, blurred, decimated and noisy low resolution (LR) images. The main contributions of the paper are as follows:On the research of observation model, two different kinds of models are first categorized and summarized, followed by a detailed analysis and evaluation of them. Based on the Elad model, the frame of SRR is presented. The analytic results indicate the relation between model errors and the quality of reconstructed image.On the research of image registration, different kinds of subpixel registration techniques are discussed. Based on the rigid transformation, an optic flow method on a Gaussian pyramid structure is utilized. Experimental results show that this method do obtain subpixel accuracy.On the research of reconstructed algorithm, a method for automatically estimating regularization parameter is applied to L1-norm based SRR framework. Besides, a novel initial guess of HR image is proposed, which result in good robustness and effective convergence. A novel explanation of L1-norm based bilateral total variation (BTV) term is presented according to pixel differences statistics combined with probability prior model, and the superiority of BTV is validated using Kullback-Leibler distance. Then a new method based on mixture prior distribution and L1-norm is proposed. Experimental results indicate the proposed approach outperform the old one.