Study On The Hybrid Image Super-resolution Reconstruction Algorithms

Digital image has been extensively used as an important way of informationexchange in modern society. The space resolution of digital image has gained muchattention as image quality assessment. However, there are very limited numbers of waysto increase image resolution due to the limitation of the cost of imaging device andmany other environmental factors. So the improvement of image resolution is always anaim for engineers and researchers. In recent years, image super-resolution (SR)reconstruction methods have been discussed in the field of image processing andcomputer vision, which are based on software processing, instead of hardware.Firstly, by incorporating the concepts of multi-frame SR and single-frame SR, ahybrid image SR reconstruction method has been proposed in this paper. Under itsframework, the input low-resolution (LR) image sequence is processed in the first stage,multi-frame SR, to obtain an initial high-resolution (HR) image, which will beprocessed further in the successive stage, single-frame SR, to get more details andimprove the image quality.In the work of multi-frame SR, the research contributions are as follows:The proposal of hierarchical iterative sub-pixel registration method for thepurpose of highly accurate registration.Normally, in the process of sub-pixel registration, the reference frame will beinterpolated at first. And the magnification factor will be large if the sub-pixelregistration accuracy is high. Then this leads to too many estimated pixels in theinterpolated reference frame, which will degrade the registration accuracy. So theproposed hierarchical structure will be used to obtain half-pixel registration and highaccuracy registration will be carried out in multiple stages. Moreover, in one singlestage of registration, to make the results precise, iteration is adopted into it. This makesthe half-pixel registration with few registration errors.Next, the image fusion methods have been examined. And in the condition ofmultiple pixels aligned in the HR grids with random shifts, five image fusion methodshave been proposed, they are:Neighbor Search Nearest Interpolation method (NSNI)This method is based on the assumption that the unknown pixel has closest relationwith its nearest neighbor pixel. So the unknown pixel will be estimated from its nearestneighbor pixel.Fixed-distance Search Nearest Interpolation method (FDSNI) In the neighbor of unknown pixel, the nearest known pixel will be found bysearching on discrete distance from near to far.Neighbor Combination Interpolation methodFrom the viewpoint of unknown pixel, its estimated is obtained by weightedsummation of known pixels in its neighbor.Neighbor Extension Interpolation methodFrom the viewpoint of known pixel, the unknown pixels around the known areestimated by extending known pixels with consideration of their distance.Bi-direction Linear Interpolation methodBased on linear interpolation concept, the two-dimensional interpolation isdecomposed into horizontal and vertical linear interpolations. And the final results areobtained by combining the previous two interpolations.On the other hand, the work in single-frame SR has following contributions:The proposal of a scheme of dynamic dictionary designThe dynamic dictionary consists of two parts: fixed part, containing generic imagepatch pairs, and dynamic part, containing image patch pairs from input LR frames.Under this structure, the dictionary contains not only universal information from genericimages, but also texture information specific to input LR frames. Thus, thereconstruction results will be better.The adoption of a simple image patch matching strategy inside the dictionaryInstead of obtaining directly the spares decomposition coefficients, in this paper,Euclidean distance is used to measure the distance between the input image patch andthe one in the dictionary. With the help of sparse co-occurrence prior, image patch in DLwill be found easily and the corresponding patch in DHwill also be found toreconstruction output image.From the above, in the work of multi-frame SR, the hierarchical iterative sub-pixelregistration method and image fusion method have been proposed. And in single-frameSR, the dynamic dictionary construction and a simple image patch matching strategy aredesigned. Furthermore, the framework of hybrid SR method is presented on the abovetwo aspects and leads to better performance in term of magnification factor. Thesimulation results from both still image and video frames show its superiority over otherclassic SR methods.