Research On Super-resolution Image Reconstruction Algorithm Based On Sparse Representation

Super-resolution image reconstruction is a technique to reconstruct ahigh-resolution image based on a single or a series of low-resolution images of the samescene, and combining priori knowledge. It can improve effectively the image resolutionand image quality without changing the existing hardware.In this thesis, the super-resolution image reconstruction algorithm based on sparserepresentation is studied. The method uses the processing strategy of image block,which is an application of compressed sensing theory in the field of imagesuper-resolution reconstruction. In other word, the sparse representation coefficients ofhigh-resolution image block can be accurately recovered by its down-sampledlow-resolution image block. A pair of high and low-resolution dictionary is acquiredthrough joint training. Then, the sparse representation coefficients of the inputlow-resolution image block are calculated based on the low-resolution dictionary. Thecoefficients are utilized together with the high-resolution dictionary to get thehigh-resolution image block corresponding to the low-resolution image block.In order to improve the matching between high and low resolution training samples,an improved dictionary training method is proposed. In the acquisition of redundantdictionary joint training samples, two methods are used to improve the quality oftraining samples:1）High frequency component of the block is used as the highresolution image samples, which fewer samples are required for dictionary training andbetter effect of super-resolution image reconstruction is got under the same condition.2）An improved multi-threshold LBP operator is used, which can effectively extract thetexture features from low-resolution image blocks, including the macro and microstructure information. It can be applied to the composition of low resolution trainingsamples. Thus, matching accuracy and quality of image reconstruction are improved.The reconstruction effect of the redundant dictionary size, regularizationcoefficient of sparse representation solving as well as the number of training samplesare analyzed in detail through experiments. Some advices on parameter selection aregiven.For the problem of high computational complexity for super-resolution imagereconstruction algorithm based on sparse representation, an adaptive fast superresolution image reconstruction method based on image block feature is proposed. According to the variance of gray-scale image block, bicubic image interpolation withlower computational complexity is adopted to the image blocks with smaller variance,which the gray-scale changes slowly. For the image blocks with bigger variance, whichhave plentiful texture, super-resolution image reconstruction algorithm based on sparserepresentation is employed to get better reconstruction results. The experimental resultsshow that the proposed algorithm can greatly decrease the image reconstruction timeand barely reduce the reconstruction precision.