Image Super-Resolution Based on Dictionary Learning and Anchored Neighborhood Regression with Mutual Incoherence.

Yulun Zhang,Kaiyu Gu,Yongbing Zhang,Jian Zhang,Qionghai Dai

DOI: https://doi.org/10.1109/icip.2015.7350867

2015-01-01

Abstract:In this paper, we employ unified mutual coherence between the dictionary atoms and atoms/samples when learning the dictionary and sampling anchored neighborhoods respectively for image super-resolution (SR) application algorithm. On one hand, an incoherence promoting term in dictionary learning for SR is introduced to encourage dictionary atoms, associated to different anchored regressors, to be as independent as possible, while still allowing for different regressors to share same samples. On the other hand, a unified form with mutual coherence between dictionary atoms and training samples is proposed when we group neighborhoods of samples centered on each atom and find the nearest neighbors for input samples in image super-resolution. Extensive experimental results on commonly used datasets demonstrate that our method outperforms state-of-the-art methods by obtaining compelling results with improved quality, such as sharper edges, finer textures and higher structural similarity.