A Restoration Network as an Implicit Prior

Authors: Yuyang Hu, Mauricio Delbracio, Peyman Milanfar, Ulugbek Kamilov

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical results show that the method using a super-resolution prior achieves state-of-the-art performance both quantitatively and qualitatively.
Researcher Affiliation Collaboration 1Washington University in St. Louis, 2Google Research {h.yuyang, kamilov}@wustl.edu, {mdelbra,milanfar}@google.com
Pseudocode Yes Algorithm 1 Deep Restoration Priors (DRP) 1: input: Initial value x0 Rn and parameters γ, τ > 0 2: for k = 1, 2, 3, . . . do 3: zk xk 1 γτG(xk 1) where G(x) := x R(Hx) 4: xk sproxγg(zk) 5: end for
Open Source Code Yes We publicly share our implementation that shows the potential of using restoration models to achieve SOTA performance. We have provided the anonymous source code in the supplementary materials. The included README.md file contains detailed instructions on how to run the code and reproduce the results reported in the paper.
Open Datasets Yes Our training dataset comprised both the DIV2K (Agustsson & Timofte, 2017) and Flick2K (Lim et al., 2017) dataset, containing 3450 color images in total.
Dataset Splits No The paper mentions using DIV2K and Flick2K for training the Swin IR model, and Set3c, Set5, CBSD68, and Mc Master datasets for evaluation. It states that hyperparameters were fine-tuned on the Set5 dataset, implying its use as a validation set, but no explicit percentages, counts, or citations to specific predefined splits are provided for any of these datasets.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments, such as GPU models, CPU specifications, or cloud computing resources.
Software Dependencies No The paper mentions software like Swin IR, Dn CNN, IRCNN, DRUNet, and Pytorch, but does not provide specific version numbers for these software components, which is required for reproducible description of ancillary software.
Experiment Setup Yes During training, we applied q bicubic downsampling to the input images with AWGN characterized by standard deviation σ randomly chosen in [0, 10/255]. We used three Swin IR SR models, each trained for different down-sampling factors: 2 , 3 and 4 . We consider image deblurring using two 25 25 Gaussian kernels (with the standard deviations 1.6 and 2) used in (Zhang et al., 2019), and the AWGN vector e corresponding to noise level of 2.55/255. We fine-turned the hyper-parameter γ, τ and SR restoration prior rate q to achieve the highest PSNR value on the Set5 dataset and then apply the same configuration to the other three datasets. In each DRP iteration, we run three steps of a CG solver, starting from a warm initialization from the previous DRP iteration.