摘要：The key issue facing the machine learning based super-resolution (SR) method is how to describe the relationship between the low-resolution (LR) and high-resolution (HR) images. Sparse representation techniques have provided effective tools for this task. In classical coupled dictionary models, the most important issue is how to train two dictionaries to convert the HR and LR data samples to a unified feature subspace. To address this problem, this paper presents novel coupled dictionary training approach for SR. In the proposed model, reverse sparse representation constrains are employed to train coupled dictionaries to reduce the weaknesses of the SR problem. To avoid the alternative iteration and reduce the time complexity, the HR and LR dictionaries are trained in two steps. First, the HR dictionary is trained with the traditional single dictionary training algorithm. Next, according to the HR dictionary and the HR data set, the reverse sparse representations are prepared to generate the LR atoms. Finally, the LR dictionary is generated with reverse sparse representations and the LR data set. Experimental results demonstrate that our approach outperforms 7 related approaches.
关键词：Super-resolution; sparse representation; dictionary training; non local mean regularization.
