Wavelet Domain Multidictionary Learning for Single Image Super-Resolution

Abstract EN

Image super-resolution (SR) aims at recoveringthe high-frequency (HF) details of a high-resolution(HR) image according to the given low-resolution (LR)image and some priors about natural images.

Learningthe relationship of the LR image and its correspondingHF details to guide the reconstruction of the HR imageis needed.

In order to alleviate the uncertainty inHF detail prediction, the HR and LR images are usuallydecomposed into 4 subbands after 1-level discretewavelet transformation (DWT), including an approximationsubband and three detail subbands.

From ourobservation, we found the approximation subbands ofthe HR image and the corresponding bicubic interpolatedimage are very similar but the respective detailsubbands are different.

Therefore, an algorithm to learn4 coupled principal component analysis (PCA) dictionariesto describe the relationship between the approximationsubband and the detail subbands is proposedin this paper.

Comparisons with various state-of-the-artmethods by experiments showed that our proposedalgorithm is superior to some related works.