Local And Non-local Prior Based Image Super Resolution

As an effective software technology, image super-resolution reconstruction canbreak through the physical limitations of imaging equipment to improve the spatialresolution. This thesis make an in-depth research on image super resolutionreconstruction based on the local and non-local prior and proposes three differentsuper-resolution algorithms.To solve the problem that it is hard to accurately estimate the motion patternbetween video frames, a video super-resolution algorithm based on gradient structuretensor and normalized convolution is proposed. The proposed algorithm needs not toperform accurate motion estimation and is capable of carrying out video up-samplingand denoising simultaneously.To address the problem that existing super-resolution algorithms can not fullymodel the local and non-local image prior due to their complexity of integration, asingle image super-resolution algorithm based on steering kernel regression andnon-local means is proposed. The proposed method extends the traditionalsuper-resolution techniques that base on steering kernel regression and non-local meansand can achieve image interpolation, denoising and deblurring simultaneously.To overcome the shortcomings that the reconstruction-based methods cannotproduce “new” high frequency information due to the lack of image library while theexample-based methods are usually specialized for a fixed zooming factor, a singleimage super-resolution algorithm with a high resolution dictionary is proposed. Theproposed method not only can perform different zooming factors with the same learneddictionary, but achieves image interpolation, denoising, and deblurring simultaneously.