TY - GEN
T1 - Blind Single Image Super-Resolution via Iterated Shared Prior Learning
AU - Pinetz, Thomas
AU - Kobler, Erich
AU - Pock, Thomas
AU - Effland, Alexander
N1 - Funding Information:
Funding. This work was supported by the German Research Foundation under Germany’s Excellence Strategy - EXC-2047/1 – 390685813 and – EXC2151 – 390873048.
Publisher Copyright:
© 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2022
Y1 - 2022
N2 - In this paper, we adapt shared prior learning for blind single image super-resolution (SISR). From a variational perspective, we are aiming at minimizing an energy functional consisting of a learned data fidelity term and a data-driven prior, where the learnable parameters are computed in a mean-field optimal control problem. In the associated loss functional, we combine a supervised loss evaluated on synthesized observations and an unsupervised Wasserstein loss for real observations, in which local statistics of images with different resolutions are compared. In shared prior learning, only the parameters of the prior are shared among both loss functions. The kernel estimate is updated iteratively after each step of shared prior learning. In numerous numerical experiments, we achieve state-of-the-art results for blind SISR with a low number of learnable parameters and small training sets to account for real applications.
AB - In this paper, we adapt shared prior learning for blind single image super-resolution (SISR). From a variational perspective, we are aiming at minimizing an energy functional consisting of a learned data fidelity term and a data-driven prior, where the learnable parameters are computed in a mean-field optimal control problem. In the associated loss functional, we combine a supervised loss evaluated on synthesized observations and an unsupervised Wasserstein loss for real observations, in which local statistics of images with different resolutions are compared. In shared prior learning, only the parameters of the prior are shared among both loss functions. The kernel estimate is updated iteratively after each step of shared prior learning. In numerous numerical experiments, we achieve state-of-the-art results for blind SISR with a low number of learnable parameters and small training sets to account for real applications.
UR - http://www.scopus.com/inward/record.url?scp=85140477906&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-16788-1_10
DO - 10.1007/978-3-031-16788-1_10
M3 - Conference paper
AN - SCOPUS:85140477906
SN - 9783031167874
T3 - Lecture Notes in Computer Science
SP - 151
EP - 165
BT - Pattern Recognition - 44th DAGM German Conference, DAGM GCPR 2022, Proceedings
A2 - Andres, Björn
A2 - Bernard, Florian
A2 - Cremers, Daniel
A2 - Frintrop, Simone
A2 - Goldlücke, Bastian
A2 - Ihrke, Ivo
PB - Springer Science and Business Media Deutschland GmbH
T2 - 44th DAGM German Conference on Pattern Recognition
Y2 - 27 September 2022 through 30 September 2022
ER -