Image Super Resolution Restoration Based On Learning

Image super resolution, as an important branch of digital image processing, hasbeen widely used in video surveillance, remote sensing image processing, computervision systems, medical imaging systems in these years.Learning-based reconstruction algorithm is one of the most successful superresolution image technology which gets the correlation of the high resolution andlow resolution through learning, and using this to reconstruct image, the informationacquired from learning can be used to optimizing and replenishing image. The basicidea of learning-based algorithm is that it can restore image more accurately, reserveimage feature consistently and the high robustness of noise. Based on thisframework, this thesis target to improve the image visual effects by learning-basedimage super resolution algorithm. The primary work in this thesis includes:Firstly, based on the learning-based algorithm, combine with the sparserepresentation theory, this thesis utilized K-SVD dictionary learning algorithm toconstruct dictionary which shrinks the scale of redundant dictionary withoutdecreasing the quality of image reconstruction. To using the correlation of the highresolution and low resolution to restoring the super resolution image, which not onlydecrease the complexity of algorithm but also increase the represent capability ofdictionary.Secondly, this paper has proposed image content based dictionary learningalgorithm of super resolution reconstruction. According to the difference of imagecontent, this thesis introduces the cluster concept, by using the cluster algorithm anintegrated and large scale dictionary can be divided to several different category, andthen select a specific category to restoring image. Compare to the traditional superresolution image reconstruction, this algorithm has obvious advantages both onbenchmark of SSIM and PNSR and visual effects also.Finally, this paper has provided image content based double dictionary learning algorithm of super resolution image reconstruction. After divided the training imageby image content, this thesis partitions the high frequency of image to mainfrequency and redundant frequency and train them respectively, finally a doubledictionary will be generated. Then with these dictionaries, double layers operationswill be applied to the images, and more high frequency will be added to image andconsequently a better restored image. The experiment shows this algorithm capturesmore details of image and gives better visual effects.