Research on Image Super Resolution Reconstruction Technique
|School||Harbin Institute of Technology|
|Course||Information and Communication Engineering|
|Keywords||super resolution regularization SIFT sparse representation redundantdictionary iterative shrinkage|
As an important digital image processing technique, image super resolution iswidely used in the field of security, personal image processing, medical imageprocessing, satellite remote sensing imaging etc. Low resolution images not onlyaffect the identification of interesting target, but also reduce the visual perception ofimages. Hence, it is necessary to do some research on the image super resolution toguarantee the image visual perception and target identification rates, and this papergo through these arguments mainly from two aspects as follow:Firstly, some research are done on multi-frame based image super resolutionreconstruction algorithms, and which generate a HR image through making full useof multiple LR images of the same scene. As to this kinds of SR algorithms, somesimuliation and analysis are done to methods of maximum a posteriori, projectiononto convex set, iterative back projection and regularization reconstruction in thispaper, and then come to a conclusion that BTV regularization reconstruction methodcan generate best HR results. As to image sequences with local motion target,traditional global geometric transformation matrix estimation methods can not meetthe requirments. Considering that scale invariant feature transform is a good tool formoving target tracking, we estimate geometric transformation matrix of movingtarget based on SIFT, then, bilateral total varation regularization reconstructionmethod is used to generate HR images. Though multi-frame based image superresolution reconstruction algorithms can achieve the most reliable HR results,actually, multi-frame with different target information are difficult to obtian inpractical applications. Thus maching learning-based super resolution methods havebecome the hot issues in SR filed.In order to achieve the goal of HR images reconstruction, machinelearning-based super resolution attempt to capture the cooccurrence prior betweenLR and HR image patches through trainning sets. As to this kinds of SR algorithms,Examples based super resolution, sparse representation based SR, iterativeshrinkage and sparse representation based SR methods were studied in this paper.Examples based super resolution can generate HR images with large magnificationfactor, while has disadvantages of that the input image must be similar to trainingsets, hence it is generally used for face reconstruction. As one of focuses in singleimage super resolution, sparse representation based SR algorithm using pairs ofredundant dictionaries method can generate HR images with better quality. Inconsideration of its high computational complexity, this paper proposed a fast superresolution method while maintaining its performance. Because of generating HR images with better quality from noisy and smooth LR images, a deep research isdone on SR method based on PCA and iterative shrinkage. Given advantage ofsparse representation using redundant dictionary, a SR method based on redundantdictionary and iterative shrinkage is proposed in this paper, which can generate HRimages with better quality rather than original algorithm.