CAJ | 학술논문

为改善强降质图像的分辨率水平，提出了一种正则化恢复联合稀疏表示的单帧图像超分辨率重构框架。为同时放大图像并抑制模糊及噪声，首先根据退化估计正则化平衡极小问题的逼近项和先验项，然后基于初步的锐利清晰图像和预先建立的图像超完备稀疏表示字典实现边缘保持的图像分辨率放大。正则化恢复的输出改善了传统学习法图像超分辨中低频分量的双立方插值版本，同时对降质的有效抑制降低了字典原子对退化信息的依赖性。实验结果表明，本方法可对模糊含噪的低分辨率图像实现有效的超分辨率重构。
위개선강강질도상적분변솔수평，제출료일충정칙화회복연합희소표시적단정도상초분변솔중구광가。위동시방대도상병억제모호급조성，수선근거퇴화고계정칙화평형겁소문제적핍근항화선험항，연후기우초보적예리청석도상화예선건립적도상초완비희소표시자전실현변연보지적도상분변솔방대。정칙화회복적수출개선료전통학습법도상초분변중저빈분량적쌍립방삽치판본，동시대강질적유효억제강저료자전원자대퇴화신식적의뢰성。실험결과표명，본방법가대모호함조적저분변솔도상실현유효적초분변솔중구。
In order to improve resolution of single frame image with severe degradation, we propose a novel super-resolution reconstruction framework via regularization restoration combined with learning-based sparse representation enhancement. To achieve enlargement and suppression of blurring and noise simultaneously, we carefully balance the data fidelity and the prior item using regularization parameter on the basis of verisimilar estimation of degradation. Based on the acquired relatively clean image and pre-constructed over-complete sparse representation dictionary, image resolution zooming with characteristic of edge-preserving can then be realized. Fundamentally, the output of preceding regularization reversion remarkably betters low-frequency quality of bicubic interpolation version in conventional learning-based super-resolution. Furthermore, the ridding of blur and noise can favorably weaken dependency of atoms to degraded information. Consequently, their combination of two techniques can remarkably eliminate blur and noise, and meanwhile, remove annoying edge artifacts of enlarged image. Experiment results demonstrate that the addressed approach produces visually pleasing magnification for blurry and noisy low-resolution image.