Super-Resolution Image Reconstruction using Sparse Representation

Abstract

This thesis addresses theigeneration andireconstruction of theihigh resolution (HR) imageiby using theisingle low resolution (LR) imageiand the linear coalitioniof sparse coefficientsifrom a suitablyichosen over-complete dictionary. The studyiof compressiveisensing shows thatiunder vague conditionsithe sparseirepresentation of a signalican be effectivelyirecovered from the down-samplediversion of the originalisignal. By training bothiLR and HRiimage patches simultaneouslyiby coupled dictionaryilearning, we areienforcing the similarityibetween the sparseirepresentation(SR) of LR and HRiimage patch pairsiwith respective toitheir LR and HR dictionaries. Literature surveyisuggests that differentiextracted features areiused to computeithe coefficients toiboost the predictioniaccuracy of the HRiimage patch reconstruction. Aiset of firstiand second orderi filters hasibeen employed toiextract useful features fromithe LR patch. As theisuper resolution is aniill posed problem, in this thesis weihave considered it asian optimization problemifor getting theisparsest representation ofiimage patches usingilinear regression regularizediwith L1 norm, known as a LASSOiin statistics. Our method isifound to beioutperforming the other previousistate of art methodsiin both quantitativeiand qualitative analysis. The resultsireveal that proposedimethod shows promisingiresults in reconstructing the imageitextures andiedges.