Image Modeling Based Regularized Multi-Frame Super-Resolution Reconstruction

Along with the gradual development of the information age and the growing popularity of image processing, image resolution has been increasingly and highly demanded in scientific research and practical applications,which gradually raises new challenges for the manufacturing technology on imaging sensors.At present,hardware programmes are,however,technically bottleneck,such as reducing the pixel size or increasing the chip size.In the meanwhile,prices of expensive high precision photographic are not applicable to wider applications.The technique of super-resolution(SR)reconstruction has become one of the hottest research topics in the world,consisting in that not only a high-resolution image can be reconstructed at a relatively low cost,but also the existing low-resolution imaging systems can be still utilized.Therefore,it is of great significance to research on the problem of super-resolution.For image denoising,image interpolation,image enhancement,motion estimation and other key issues in SR, this paper manily focuses on image modeling for edge-preservation,texture-preservation,and comer-preservation, and correspondingly proposes several regularized SR reconstruction algorithms.(1)The edge-preserving potential functions in isotropic MRF modeling are systematically discussed,and a type of non-negative robustρnorms is particularly defined for potential functions.In the meanwhile,the bilateral filtering is closely connected with the Bayesian MAP approaches,and subsequently generalized to the so-called robust bilateral filtering.Furthermore,a kind of structure-adaptive anisotropic filtering is designed based on both the luminance and the geometry of pixels,following which an anisotropic MRF-based edge-enhancing SR algorithm is proposed,capable of simultaneously estimating the high resolution image and the subpixel motion among low-resolution frames.Experimental results demonstrate the effectiveness of the proposed approach,both in the visual effect and the PSNR value.(2)To overcome the shortcomings of the continuous total variation(CTV)model,this paper respectively proposes the edge-preserving beyond digital TV(BDTV)model and the texture-preserving adaptive beyond continous TV(ABCTV)model.BDTV is particularly applicable to cartoon images,capable of simultaneously prerving edges and removing staircase artifacts on homogeneous regions.ABCTV achieves the adaptive coupling of the CTV with higher-order CTV using the principal curvatures,not only able to preserve edges but also able to preserve textures.Based on BDTV,two simultaneous SR reconstruction and motion estimation algorithms are proposed,respectively using image registration and optical flow techniques to estimate the sub-pixel motion.Both static and dynamic SR experimental results demonstrate the effectiveness of the proposed algorithms.(3)There have been proposed many magnification algorithms in the past decades,most of which concentrate on the smooth reconstruction of edge structures.Edge reconstruction,however,may usually destroy the corners,thus producing perceptually unpleasant rounded comer structures.A kind of comer shock filtering is particularly designed to enhance the corner structures,based on a new measure of corner strength and the theory of level-sets motion under curvature.Through combining the directional diffusion,the edge shock filtering.and the corner shock filtering,a regularized PDE approach for magnification is proposed to simultaneously reconstruct the edges and enhance the comers.The proposed PDE approach is also robust to random noise.Based on the regularized PDE interpolation approach,a structure enhancing PDE SR algorithm is subsequently proposed,which is actually the effective development of Capel and Kim's TV-based PDE SR algorithm.Both static and dynamic SR experimental results demonstrate the effectiveness of the proposed algorithms.(4)The paper also analyzes the filtering behavior of the structure tensor based variational PDE approaches, utilizing the edge shock filtering and the proposed corner shock filtering,Weickert's anisotropic PDE corresponds to a kind of smoothing-enhancing and corner-preserving filtering mechanism;while Tschumperle's computable trace-based PDE is the reduced version of Weickert's approach,corresponding to a kind of smoothing filtering mechanism but not capable of corner-preservation.Furthermore,comer-preserving conditions are intuitively proposed for structure tensor based variational functionals.Based on above discussions,a kind of common regularization PDE framework is proposed for different image applications.The filtering behavior on homogeneous regions,edge and corner structures can be described more intuitively and efficiently by the common framework.Experimental results of image filtering,magnification,and inpainting demonstrate the efficiency of the common PDE framework.(5)A SR-objected robust registration algorithm for the rigid image transform is proposed,through maximizing the designed structure similarity based on the robust estimation of structure orientation by the structure tensor. Experimental results demonstrate that the proposed algorithm is robust to random noise,optical blur,and spectrum aliasing,and behaves better than the commonly utilized KPB algorithm in accuracy,robustness,and applicability.