Self-learning super-resolution (SLSR) algorithms have the advantage of being independent of an external training database. This paper proposes an SLSR algorithm that uses convolutional principal component analysis (CPCA) and random matching. The technologies of CPCA and random matching greatly improve the efficiency of self-learning. There are two main steps in this algorithm: forming the training and testing the data sets and patch matching. In the data set forming step, we propose the CPCA to extract the low-dimensional features of the data set. The CPCA uses a convolutional method to quickly extract the principal component analysis (PCA) features of each image patch in every training and testing image. In the patch matching step, we propose a two-step random oscillation accompanied with propagation to accelerate the matching process. This patch matching method avoids exhaustive searching by utilizing the local similarity prior of natural images. The two-step random oscillation first performs a coarse patch matching using the variance feature and then performs a detailed matching using the PCA feature, which is useful to find reliable matching patches. The propagation strategy enables patches to propagate the good matching patches to their neighbors. The experimental results demonstrate that the proposed algorithm has a substantially lower time cost than that of many existing self-learning algorithms, leading to better reconstruction quality.

In the learning based super-resolution (SR), one of the most important issue is how to learn the relationship between the high resolution (HR) and low resolution (LR) images. Sparse representation has provided dictionary learning methods to describe the relationship. This work presents a coupled dictionary training algorithm named coupled K-singular value decomposition (K-SVD) for SR problem. In this algorithm, the best low-rank approximation provided by singular value decomposition (SVD) is utilized to update the LR and HR dictionaries. Experiments demonstrate that our algorithm converges stably and achieves superior SR results.