LAUSR

Abstract Recently, the anchored neighborhood regression (ANR) and the adjusted anchored neighborhood regression(A + ) have shown the state-of-the-art performance in image super-resolution (SR). However, ANR and A + only learn one group of anchors and regressors for the SR task, which are generally not enough to cover all variations of natural image structures. Therefore, it is necessary to train multiple models so as to enhance the adaptability. In this respect, the training patches are divided into several groups so that the mapping matrix in each group can be learned more stably. Meanwhile, we propose the partially supervised strategy to match the most corresponding mapping matrix, which considers the class information moderately to estimate the HR patches. The experiment results show that our method outperforms other competing methods in terms of both quantity and quality.