Research On Algorithms Of Super-Resolution Reconstruction Of Image And Video Sequences

The technique of image super-resolution reconstruction(SRR) is a technique which could reconstruct the high resolution(HR) image from one single low resolution(LR) image or some LR multi-images by using signal processing technology. It could eliminate detrimental effect of space and time resolution decline due to image discretization and image degradation, meanwhile, compensates space and time resolution decline of LR image, then gets high resolution image. Image SRR technique has a broad application prospect in many fields such as telemeter, military, medical imaging and public security.The study is carried out in the following aspects in the paper: The related algorithms of one single image SRR, image registration, video sequences SRR and image denoise preprocessing. The main contents of this paper are as following:1. Image SRR theory is introduced; the math model of image SRR and video sequences is analyzed; the prospect of image SRR technique is described.2. Image denoise preprocessing. Because image SRR technique belongs to pixel-level image fusion field, fusing blurred image directly is bound to bring image noise into fusion result. Before image SRR, image noise is preprocessed in order to gain better fusion effect. Because noises in the actual system are mainly additive noises, the filtering method of impulse noise and Gaussian noise are studied and applied.3. Utilizing the interpolation technique and the artificial neural network fusion, one single HR image is constructed. Firstly, on the basis of studying existing interpolation algorithms, double-channel interpolation is proposed, and two different kind of interpolation techniques are used for different domains of the image and at the same time computation loads is reduced. Secondly, on the basis of Gauss pyramid decoding theory, the algorithm of one single image SRR based SGNN is proposed. In this algorithm, high frequency subimage is defused by SGNN in order to get HR image. 4. Video image sequences SRR is discussed in the basis of registration and SRR algorithm. Firstly, improved Keren algorithm based on 6-paremeter affine transformation model is proposed. Utilizing the proposed algorithm, image registration is studied in order to determine location relationship of inter-frame subpixel and prepare for math modeling of SRR. Secondly, video sequences SRR is studied and LMS adaptive filtering method applied to video sequences SRR which is blurred by outliers. This algorithm eliminates blur through selecting the reasonable step. As a result, the effect of SRR is better and this algorithm is of robustness.