Blind Super-Resolution Kernel Estimation using an Internal-GAN

Authors: Sefi Bell-Kligler, Assaf Shocher, Michal Irani

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluated our method on real LR images, as well as on non-ideal synthetically generated LR images with ground-truth (both ground-truth HR images, as well as the true SR-kernels). The latter is done both visually (see fig. 1a and supplementary material), and empirically on the synthetic dataset analyzing PSNR and SSIM measurements (Table 1).
Researcher Affiliation Academia SefiBell-Kligler Assaf Shocher Michal Irani Dept. of Computer Science and Applied Math The Weizmann Institute of Science, Israel
Pseudocode No The paper includes architectural diagrams (Figures 2, 3, and 4) but no explicit pseudocode or algorithm blocks.
Open Source Code Yes Project website: http://www.wisdom.weizmann.ac.il/ vision/kernelgan and Data and reproduction code are available on project website.
Open Datasets Yes Using the validation set (100 images) from the widely used DIV2K [1] dataset, we blurred and subsampled each image with a different, randomly generated kernel.
Dataset Splits No While the paper states that its evaluation dataset, DIV2KRK, is generated from the 'validation set (100 images) from the widely used DIV2K [1] dataset', it does not provide explicit training/validation/test splits for this generated dataset as typically defined for model training or evaluation. The model itself trains on the LR test image at test time.
Hardware Specification Yes Runtime is 61 or 102 seconds per image on a single Tesla V-100 or Tesla K-80 GPU, respectively.
Software Dependencies No The paper mentions the use of the ADAM optimizer but does not specify any software names with version numbers (e.g., specific deep learning frameworks like PyTorch or TensorFlow, or other libraries).
Experiment Setup Yes The GAN trains for 3,000 iterations, alternating single optimization steps of G and D, with the ADAM optimizer (β1 = 0.5, β2 = 0.999). Learning rate is 2e 4, decaying 0.1 every 750 iters. [...] We use 5 hidden convolutional layers with 64 channels each. The first 3 filters are 7 7, 5 5, 3 3 and the rest are 1 1. [...] R = αLsum_to_1 + βLboundaries + γLsparse + δLcenter (3) where α = 0.5, β = 0.5, γ = 5, δ = 1.