When deep denoising meets iterative phase retrieval
Authors: Yaotian Wang, Xiaohang Sun, Jason Fleischer
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare Deep-ITA-F/S with other widely used algorithms on FPR, namely HIO (Fienup, 1982), Oversampling Smoothness (OSS) (Rodriguez et al., 2013), Dn CNNADMM (Venkatakrishnan et al., 2013; Heide et al., 2016; Chan et al., 2017) and pr Deep (Metzler et al., 2018). ... Results from two experimental setups are reported here. ... Table 1. PSNRs and SSIMs of reconstructions initialized with random noise with varying noise level in the measurements. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering, Princeton University, Princeton, New Jersey, USA. Correspondence to: Jason W. Fleischer <jasonf@princeton.edu>. |
| Pseudocode | Yes | Algorithm 1 RED-ITA-F; Algorithm 2 RED-ITA-S |
| Open Source Code | No | The paper does not provide explicit statements or links indicating the availability of open-source code for the methodology described. |
| Open Datasets | No | The test images used in the simulations, shown in Figure 1, consist of 6 commonly used natural images and 6 unnatural ones. ... Dn CNN models are trained on patches of natural images with mean-squared-error as the loss function, using Adam as the optimizer (Kingma & Ba, 2014). The paper refers to 'commonly used natural test images' and a prior work for Dn CNN training, but does not provide concrete access information (link, DOI, specific repository, or formal citation with author/year) for the specific datasets used in *their* experiments. |
| Dataset Splits | No | The paper does not provide specific dataset split information (percentages, sample counts, or detailed methodology) for training, validation, or testing. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments (e.g., GPU/CPU models, memory, or cloud instances). |
| Software Dependencies | No | The paper mentions software components like 'Dn CNN' and 'Adam' but does not specify their version numbers or any other software dependencies with specific versions. |
| Experiment Setup | Yes | The parameters in the algorithms were as follows: for HIO and OSS, β = 0.9. The regularization parameter λ is found best set as 0.01 σ2 for Dn CNN-ADMM, 0.025 σ2 for both Deep-ITA-S/F, and 0.05 σ2 for pr Deep, where σ is the standard deviation of noise in the Fourier amplitude (or set to 0.1 if no noise is added). Similar to the practice in (Metzler et al., 2018), pr Deep and Deep-ITAs sequentially use Dn CNN models that are trained with noise standard deviations of 60, 40, 20 and 10, each with 300 iterations for a total of 1200 iterations. The penalty parameter ρ used in Deep-ITAs is set to 1/2λ. |