Image Super-Resolution Based on Locally Regularized Anchored Neighborhood and Cascade Frame

The purpose of learning-based image super-resolution (SR) is to generate a reasonable and visually pleasing high-resolution (HR) image from a low-resolution (LR) input. In learning-based image super-resolution algorithms, Anchored Neighborhood Regression (ANR) algorithm has attracted widespread attention due to its superior reconstruction speed and quality. An improved super-resolution algorithm based on Adjusted Anchored Neighborhood Regression (A+) is proposed in this paper. In order to obtain accurate reconstruction weights and projection matrices, local constrained regression is used instead of ridge regression in A+. The nonlocal means filter is incorporated to improve the super-resolved results. In addition, the collaborative representation cascade architecture is used to gradually upgrade and improve LR input by extracting features from intermediate recovered images. The experimental results on commonly used image test sets show that our algorithm can achieve better visual effects and higher evaluation indicators compared with other state-of-the-art super-resolution algorithms. The reconstruction performance of the proposed algorithm is even better than that of Super-Resolution Convolutional Neural Network (SRCNN) algorithm.