Study Of Video Super Resolution Reconstruction Technology

In order to improve image’s resolution and restore high-frequency information,super-resolution reconstruction technology estimates a high-resolution image usingmulti-frame low-resolution images. Due to the displacements between differentframes of image sequences, some complementary information exists between thosedifferent images. Actually super-resolution technology reconstructs high-resolutionimages with the complementary information.This paper first discusses the current research and future developments of videosuper resolution techniques. Then basic theories of super resolution and useful imagedegradation models are introduced. Meanwhile, several popular super resolutionalgorithms are described. Besides that, several existing motion estimation methods areintroduced, since high accuracy motion estimation results play a vital role in superresolution technique. Therefore, a best matching point extraction method based oncorrelative constraint matching and projective constraint is proposed in this paper. Inorder to get rid of unstable and false corners, firstly the candidate matching points areobtained with the correlative constraint matching. Secondly, images are initiallymatched based on the whole stability of corners in the process of affine transformation.Finally, by setting the condition of coplanar point projective constraint, the relativehighest accurate matching points are selected as the best matching corners.Experiments show that the best matching points selected by the proposed method arerobust to added noise and interference, and ideal registration results can be achievedeven for real images with serious added noise and interference. In addition, a noveloptical flow motion estimation method is also introduced in this paper, which utilizesweighted gradient constancy assumption and incorporates with photometric invariantassumption in HSV color spaces. Reliable and precise motion estimation results canbe obtained even with large displacements and strongly varying illumination in imageframes. Experimental results show that the optical flow method can achieve motionestimation results with sub-pixel accuracy.In the final part of the super resolution method, a novel super-resolutionframework is proposed to protect flat regions and edges simultaneously withLorentzian stochastic estimation and gradient constrain. In order to remove the outliers and maintain the smoothness of the reconstructed image, the Lorentzianstochastic estimation is used for measuring the difference between the estimatedhigh-resolution image and each low-resolution one. Moreover, this paper proposes anew regularization item, termed as Lorentzian gradient constraint, which can keep flatregions and enhance edges simultaneously by incorporating with the bilateral totalvariation.