Studies On Several Problems In Image Super-resolution Reconstruction

With the development of internet and information technology, a few of multimedia applications have been emerging constantly. There is a high demand for image and video technology. However, due to the limitations of imaging system, it is not always to capture an image at a desired high-resolution (HR) level, which causes many difficulties for image processing, analysis, and understanding, leading to an obstacle in correctly understanding the laws of the objective world. Improving the performance of the hardware is clearly one way to increase the resolution of the acquired images. However, this method may not be feasible due to the increased cost. Therefore, alternative methods should be proposed in order to increase the spatial resolution. The technique of super-resolution (SR) is developed under this circumstance. The aim of the SR is to improve the resolution using signal processing techniques instead of changing hardware environment. Accordingly, the SR technique has attracted broad attention from the academic world at home and abroad as soon as it was proposed. At present, the research of SR will be of great value in a lot applications such as digital entertainment, remote sensing, medical imaging, surveillance system and so on.This dissertation reviews the SR related theory and classical algorithms systematically. To address various difficulties in SR applications such as inaccurate motion estimation, reconstruction process sensitive to the deviation of assumed model, information breaking in image sequence and ambiguity during searching for HR examples and so on, this dissertation proposes a series of solutions. The major contributions are the following:(1) To improve the reliability of motion estimation in SR applications, a SR algorithm is proposed, which is based on non-local regularization and reliable motion estimation. In the motion estimation phase, by incorporating a multi-lateral filter in the proposed method, the proposed algorithm can use reliable motion estimates to correct unreliable estimates, thus providing more reliable information on motion fields for HR image reconstruction. In the reconstruction phase, the HR image reconstruction is then regularized by using a non-local (NL) similarity prior instead of the popular statistical priors. This NL prior not only reduces undesirable artifacts but also mitigates the effect of inaccurate motion estimation. Finally, the blur kernel estimated by analyzing sharp edges is used to deblur the recovered blurry HR image. By using these methods, the proposed algorithm effectively improves the quality of multi-frame SR reconstruction. (2) Though traditional SR methods usually have good performance in theoretical environment, they tend not to be robust in real-life environment, that is, once the actual situation is inconsistent with the assumed theoretical model, traditional SR algorithms often end up with unsatisfactory results. To enhance the robustness of SR algorithm, an adaptive robust multi-frame SR algorithm is studied in this paper. Some robustness-enhancing methods are adopted in both registration phase and reconstruction phase. On registration phase, the probabilistic motion estimation is introduced to reduce the sensitivity of SR algorithm to the accuracy of motion estimation. In addition, Heaviside function is adopted to implement the motion weight mapping, which improves self-adaption of the algorithm further. On reconstruction phase, a regularized estimation based on Huber norm is used to reconstruct SR image, which makes the proposed algorithm more stable to minimize the cost function while still robust against large errors.(3) Traditional sequence SR methods require high-quality low-resolution (LR) sequence. When the correlated information contained in the sequence has been destroyed, annoying artifacts often appear in the SR outcome. To solve this problem, a multi-frame SR combined with single image restoration technique is proposed in this paper. Firstly an HR image is reconstructed by using the information existed in the LR sequence. Secondly a sparse-coding-based single image restoration technique is applied to recover the ineffective regions that cannot be well-reconstructed due to lack of enough information in the sequence. The proposed algorithm combines the advantages of two types of methods. It can not only use the correlated information in the sequence itself, but also can use the information from external example-dataset to repair the ineffective regions. Therefore, the proposed method can effectively take on the SR task when the sequence contains insufficient information.(4) To address the ambiguity of HR/LR examples in single image SR, A combinatorial framework is proposed, which integrates edge-preserving and detail-extrapolating SR together On edge reconstruction phase, a "denoised" sparse coding model is employed to establish the primitive structure prior for recovering sharp edges. During dictionary learning, the proposed method eliminates the examples which exist serious ambiguity. On detail reconstruction phase, the local context is utilized to solve example-ambiguity problem and to determine whether an LR patch should be reconstructed or not by the example-based method, which help the proposed algorithm avoid the influence of reconstructing all LR patches together and improve the quality of detail reconstruction.