| Super resolution (SR) is a challenging and hot research issue in computer vision, which aims to transfer several low resolution (LR) images to one high resolution (HR) image. It not only improves the resolution of digital imaging sensor, but also breaks the diffraction limit of systems, and some SR techniques even recover the details which do not exist in LR images. It is necessary to use SR algorithms for the obtaining of digital image becomes more difficult or expensive.In recent years, SR techniques have been benefited from the development of the computer vision technologies as well as the machine learning algorithms. Conventional SR techniques can be roughly categorized into prior-based methods and learning-based methods according to different problem-specific solutions. The former ones learn one prior by some image statistics methods, which is to primely regularize the solution space of image processing model, and the reconstruction constraint ensures the HR images to be more reasonable during the iteratively refined process; the latter ones learn HR/LR patches pairs to get a dictionary from some training sets which consists of extra HR images. When the algorithm runs, the LR patches are going to be replaced by most wanted HR patches in the dictionary.The natural scene is considered as an unlimited model, which could arbitrarily zoom in or out.HR image and LR image is a sampling process to the model. They are both infinitesimal fixed value, and HR image is an infinitesimal of lower order of LR image. In this paper, we propose two novel SR methods on the basis of the model, which include vision for digital image characteristics prior based method (VDIC) and single scene multi examples learning based method (SSME). VDIC adopts the prior from the human vision for digital image, whose contributions include (i) a more de-tailed edge classification is proposed and (ii) a un-uniform iterative method is adopted to speed up the algorithm. SSME uses a specific and actual meaningful image ex-ample as training sets, which combines the advantages of extra images learning and self-example learning. It picks up suitable example images at the beginning by image classification, then builds up two dictionaries by sparse representation, which makes the output SR image conforms to the style of the training set and be more real. At the end, we have used above-mentioned two SR methods on several image sets, and analyze the advantages and disadvantages of our methods by comparing with others. |
