Research On Algorithms Of Super-resolution Image Reconstruction

Image resolution is one of the most important parameters that affected imagequality. An image with higher resolution can improve the precision of targetrecognition and interpretation. Generally the acquired images cannot meet researchdemands due to the influence of factors such as shooting condition and imagingequipment. Many researchers have proposed super-resolution algorithms to increaseoriginal image resolution.Image super-resolution is an algorithm that generates one or morehigh-resolution images from one or more low-resolution observations of the samescenario. This paper mainly focused on super-resolution algorithms for natural imagesand remote sensing images. The main works and innovations are introduced asfollows:(1) The multi-frame super-resolution algorithm based on natural images: Theregistration algorithm of hybrid frequency-spatial domain was studied and improvedin the paper. In the traditional image registration algorithms, the iterations were usedto improve accuracy, whose number was affected by the relative parameters. Based onthe traditional algorithms，the registration evaluation parameter (REP) was proposedin the improved algorithm, which can automatically find the optimal iteration number.The experimental results demonstrated that the improved algorithm achieved betterregistration images. In addition, four common image reconstruction algorithmsincluding points onto convex sets (POCS), papoulis-gerchberg (PG), iterated backprojection (IBP) and structured-adaptive normalized convolution (SANC) werestudied and compared. SANC was proved to obtain the highest reconstruction qualityin simulation experiment.(2) The multi-frame super-resolution algorithm based on MODIS remote sensingimages: Considering the low spatial resolution, MODIS (moderate resolution imaging spectroradiometer) images were processed by the proposed super-resolution algorithm.First of all, histogram matching method was employed to reduce the irradiationdifference between these MODIS images and solve the interference stripe problem.The results demonstrated that the proposed algorithm effectively improved theoriginal image resolution from500m to250m approximately.(3) The single frame super-resolution algorithm based on MODIS remote sensingimage: A super-resolution algorithm for MODIS image based on both waveletpreprocessing and sparse representation was proposed. The dictionaries in the waveletdomain (including low-frequency, horizontal and vertical high-frequency subbands)were firstly obtained by applying the k-singular value decomposition (K-SVD)dictionary training algorithm. Subsequently through these dictionaries including therelationship between Landsat ETM+and MODIS images, a high-resolution imagewas reconstructed from a given single image using the orthogonal matching pursuit(OMP) algorithm. The experiments indicated that more image details were obtainedusing the proposed algorithm. Moreover, the computational efficiency andreconstruction quality using the algorithm can be further optimized effectively whengiven a suitable relative parameter.