Researches On Sequence Image Super-resolution Reconstruction Algorithm

Super-resolution (SR) image reconstruction algorithm produces a high-resolutionimage from one or a set of low-resolution images of the desired scene. With the rapiddevelopment of image and video processing technology, the demand forhigh-resolution image and video is increasingly strong. High-resolution image is veryimportant for image analysis and processing, because it can provide more detailedinformation of the target. The obtained image or video resolution is often low, becauseof the interference of environmental factors, the limitations of the imaging systemhardware conditions, the channel bandwidth restrictions, and image compression andother factors. It is technical difficult to improve image resolution through theimprovement of the system hardware, and the cost investment is too big. Therefore,super-resolution image reconstruction has become a hot research topic. In this paper,we conducted in-depth research of super-resolution reconstruction algorithm, the maincontent as follows:(1) On the basis of the existing image super-resolution reconstruction algorithms,this paper proposes a novel two-stage super-resolution algorithm framework combinedsparse signal representation with the projection onto convex sets (POCS). Throughlearning and training sample images, sparse representation method would get thecommon primitives of image local structure, and then the correspondinghigh-frequency details of high-resolution images are obtained. The sparserepresentation method is to take full advantage of self-information and the prioriinformation of natural images. The POCS method can easily add the priori informationto images, in order to keep high-resolution image edges, details and other information.The POCS method is to take full advantage of the redundant information betweenimages and the priori information of images. We combine the advantages of sparserepresentation and POCS, and then propose a two-stage super-resolution reconstructionalgorithm.(2) This paper studies image sparse representation theory and single dictionarydesign algorithm, and then designs a joint dictionary design algorithm based onK-SVD algorithm. This algorithm gets a high-resolution learning dictionary and acorresponding low-resolution learning dictionary through learning and traininghigh-resolution and low-resolution image matches. Then it combines the corresponding relations of the two dictionaries and sparse coding of low-resolution image matches tozoom resolution. Feature extraction is added into low-resolution image matches,because it can obtain more efficient coding and reduce the coding complexity.(3) With studying the theory and the newest technology in existing image motionestimation algorithms, an improved Keren motion estimation algorithm based onTaylor series motion estimation method is proposed. Keren algorithm is based on smallangle Taylor series, and takes rigid body transformation model, but its estimation erroris very large when the image rotation angle is greater than6degrees. In order to solvethis problem, this paper transforms the traditional rigid body model into affinetransformation model, and adopts three layers Gaussian pyramid. This methodovercomes the shortcoming of Keren algorithm, and gets more precise motionestimation information.