Research On Algorithms Of Sparse Representation-based Image Super Resolution

Theory of image sparse representation becomes the latest area of signalprocessing in recent years. With its mathematical foundation more and more perfectand successful application in compressed sensing area, the practical outlook ofimage sparse representation has been reflected and image sparse representation hasbecome an important researching area in signal processing area.For the moment,researchers successfully apply this theory into related image area, including imagedenoising and restoration, classification and searching, transmission in a high speedand storage.Especially in the area of super-resolution image analysis, sparserepresentation has its unique advantage. With using the prior image knowledge torealize better reconstructed image,image super-resolution via sparse representationhas a broad applicative prospect. This paper uses image super-resolution methodmaking two applactions, including MRI super-resolution and remote sensing image,with important practical and theoretical value.By researching on the basic theoryand framework of image super-resolution via sparse representation,This paperintends to build a complete parameter analysis framework of image super-resolutionvia sparse representation. At the same time using the proposed novel trainingredundant co-dictionary technique, this paper solve the problem of pratical imagesuper-resolution anaylsis. The innovative work of this paper includes,1. Aiming at the problem of sparse representation-based super-resolutionimage model parameters selection, this paper proposes a complete framework ofsuper-resolution image analysis via sparse representation. According to the givenevaluation criteria of image quality and the results of synthetical emulationalexperiments, the paper synthetically analyses the effect of sparse representation-based algorithm parameters on the image quality.2. According to existing problem of training dictionary used for super-resolution image analysis, this paper proposes a method of training redundant co-dictionary. After analysing the existing problem of the conventional method ofdictionary training, this paper co-constructs high and low resolution image-patchdictionary to realize high resolution image-patch corresponding to the low version.By taking advantage of the mapping relationship between sparse coefficients, themothod build the mapping relationship from high patches to low version which are obtained by downsampling high patches, in order to reconstruct high resolutionimage. The experiments show that this method can efficiently realize highresolution image restruction and the quality of the restructed image is competitiveto other convnetional methods.3. As to the problem of practical application of image super-resolution viasparse represention, the paper proposes medical image and remote image super-resolution analysis application,considering the features of practical medical imageand remote image. The paper performs experiments with real MRI images andsatellite remote images. The results of experiments show that the method proposedby this paper can efficiently take advantage of prior knowledge of training samples,and reconstructs high resolution with competitive properties.