Interpretable Unsupervised Diversity Denoising and Artefact Removal
Authors: Mangal Prakash, Mauricio Delbracio, Peyman Milanfar, Florian Jug
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method achieves stateof-the-art results on twelve benchmark image denoising datasets while providing access to a whole distribution of sensibly restored solutions. and We compare HDN on all datasets against piq 7 unsupervised baseline methods... and piiq against 2 supervised methods... and Table 1: Quantitative evaluation of pixel-noise removal on several datasets. |
| Researcher Affiliation | Collaboration | Mangal Prakash1,2 Mauricio Delbracio3 Peyman Milanfar3 Florian Jug1,2,4 1Center for Systems Biology Dresden 2Max Planck Institute (CBG) 3Google Research 4Fondazione Human Technopole |
| Pseudocode | No | The paper refers to “A schematic of our fully convolutional network architecture is shown in Appendix A.1.1.” and Appendix A.1.4 contains “Figure 5: HIERARCHICAL DIVNOISING (HDN) network architecture.” This figure is a network diagram, not pseudocode or an algorithm block. |
| Open Source Code | No | The paper states “The noisy train, validation and test images for Kanji and MNIST data will be released publicly.” and “We will release the noisy versions of these datasets publicly.” These statements refer to data, not the open-sourcing of their method’s code. |
| Open Datasets | Yes | We used 12 publicly available denoising datasets from different image domains... and then lists specific datasets with citations, e.g., FU-PN2V Convallaria dataset from Krull et al. (2020); Prakash et al. (2019b). |
| Dataset Splits | Yes | The train and validation splits for these datasets are as described in the respective publication. For the FU-PN2V Convallaria dataset, the test set consists of 100 images of size 512ˆ512, for the FU-PN2V Mouse Actin dataset, the test set consists of 100 images of size 1024 ˆ 512, for the FU-PN2V Mouse Nuclei dataset the test set has 200 images of size 512 ˆ 256 while the test set for Denoi Seg Flywing dataset consists of 42 images of size 512 ˆ 512. and for Kanji: we randomly select 119360 images for training and 14042 images for validation. The test set contains 100 randomly selected images not present in either training or validation sets. |
| Hardware Specification | Yes | on a common 6 GB Tesla P100 GPU training still only takes about 1 day and The training for all models presented in this work were performed on a single Nvidia Tesla P100 GPU requiring only 6 GB of GPU memory. |
| Software Dependencies | No | The paper mentions using “Adamax (Kingma & Ba, 2015)” as an optimizer, but does not list specific software libraries or frameworks with version numbers (e.g., PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | We use an initial learning rate of 0.0003 and always train for 200000 steps using Adamax (Kingma & Ba, 2015). During training, we extract random patches from the training data (128 ˆ 128 patches for Bio ID Faces and natural image datasets, 256 ˆ 256 patches for Celeb A HQ, 28 ˆ 28 patches for MNIST, and 64 ˆ 64 patches for all other datasets). A batch size of 16 was used for Bio ID Faces and BSDD68 while a batch size of 4 was used for the Celeb A HQ dataset . For all other datasets, a batch size of 64 was used. To prevent KL vanishing (Bowman et al., 2015), we use the free bits approach as described in Kingma et al. (2016); Chen et al. (2016). For experiments on all datasets other than the MNIST dataset, we set the value of free bits to 1.0, while for the MNIST dataset a value of 0.5 was used. |