Noise2Noise: Learning Image Restoration without Clean Data
Authors: Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, Timo Aila
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We use three well-known datasets: BSD300 (Martin et al., 2001), SET14 (Zeyde et al., 2010), and KODAK1. As summarized in Table 1, the behavior is qualitatively similar in all three sets, and thus we discuss the averages. When trained using the standard way with clean targets (Equation 1), RED30 achieves 31.63 0.02 d B with σ = 25. The confidence interval was computed by sampling five random initializations. The widely used benchmark denoiser BM3D (Dabov et al., 2007) gives 0.7 d B worse results. |
| Researcher Affiliation | Collaboration | 1NVIDIA 2Aalto University 3MIT CSAIL. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide a specific repository link, explicit code release statement, or indicate code in supplementary materials. |
| Open Datasets | Yes | We train the network using 256 256-pixel crops drawn from the 50k images in the IMAGENET validation set. We use three well-known datasets: BSD300 (Martin et al., 2001), SET14 (Zeyde et al., 2010), and KODAK1. (1http://r0k.us/graphics/kodak/). We perform experiments on 2D slices extracted from the IXI brain scan MRI dataset.4 (4http://brain-development.org/ixi-dataset T1 images.) |
| Dataset Splits | Yes | Our training set consisted of 860 architectural images, and the validation was done using 34 images from a different set of scenes. The training set contained 4936 images in 256 256 resolution from 50 subjects, and for validation we chose 500 random images from 10 different subjects. |
| Hardware Specification | Yes | training took 12 hours with a single NVIDIA Tesla P100 GPU. On an NVIDIA Titan V GPU, path tracing a single 512 512 pixel image with 8 spp took 190 ms... Training took 13 hours on an NVIDIA Tesla P100 GPU. high-end graphics server with 8 NVIDIA Tesla P100 GPUs and a 40-core Intel Xeon CPU. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4). |
| Experiment Setup | Yes | We train the network using 256 256-pixel crops drawn from the 50k images in the IMAGENET validation set. We furthermore randomize the noise standard deviation σ [0, 50] separately for each training example... The probability of corrupted pixels is denoted with p; in our training we vary p [0.0, 0.95] and during testing p = 0.5. For approximate mode seeking, we use an annealed version of the L0 loss function defined as (|fθ(ˆx) ˆy| + ϵ)γ, where ϵ = 10 8, where γ is annealed linearly from 2 to 0 during training. The network trained for 2000 epochs using clean target images... A single network training iteration with a random 256 256 pixel crop took 11.25 ms and we performed eight of them per frame. |