Image And Video Super Resolution Based On Couple Dictionaries Learning

Image Super-resolution (SR) refers to use one or more low-resolution images toreconstruct a high resolution image with an appropriate SR algorithm. Traditional superresolution algorithms including Bilinear, Bicubic, Iterative reverse projection andConvex set projection which are fast, simple, easy to implement, and have been widelyapplied to the super resolution reconstruction. But at the same time the model of abovealgorithms are too simple to reconstruct a high quality results, therefore a large numberof researchers are constantly looking for a better performance super resolution method.In recent years, there is a learning based method for image super resolution has beenproposed which realize the image super resolution through learning the relationshipbetween low resolution images and high resolution images. The method has highprecision and strong robustness, and has become an important direction in superresolution studies. This paper is proposed on the framework of learning based methodwhich combined external sample learning and internal sample learning method, andresearched the image and video super resolution reconstruction. The main works asfollows:(1) An image super resolution reconstruction based on couple dictionaries learningmethod is proposed on the characteristic of extremely sensitive to the training samplesand the unstable reconstructed results. The method can reach multiple results throughexternal samples and internal samples, and fuse the results with low rank decompositionmethod. Compared with single dictionary strategy, the method can obtain a much morestable result, also have an improvement on the details of reconstructed effect.Experimental results validate that the performance of the proposed approach in bothevaluation indexes and visual quality were much better than single dictionary method.(2) For video super resolution problems, starting from video background modeling,we proposed a video background modeling method based on information compensation.This method introduced information compensation strategies into low rank backgroundmodeling way, and obtained a much better result. Experiments show that the methodcan get a background image with much richer details than low rank method.(3) A video super resolution reconstructed method based on neighborhood coupledictionaries learning algorithm is proposed. Compared with single image, videocontains much more redundant information. This method brings in the neighbor frame information into couple dictionary learning model, and applied it in process of videosuper resolution reconstruction. Experimental results show that the proposed methodcan achieve much high definition videos which contrains much more details than coupledictionaries method.