scCO2-Brine-Glass Dataset for Comparing Image Denoising Algorithms

Description

This project contains micro-CT images of a super critical CO2-brine-glass system, at 41.7 C and 1180 psia. The datasets include high exposure high quality (HQ) and low exposure low quality (LQ) scans of the two-phase saturated core. It also contains denoised versions of the LQ dataset, using: 1- Traditional denoising filters: Gaussian filter, Median filter, Non-local means (NLM) filter, Anisotropic diffusion (AD) filter, Bilateral filter, and Symmetric nearest neighbor (SNN) filter. 2- Supervised deep learning denoising models: Noise-to-clean (N2C), Residual dense network (RDN), and Continuous conditional generative adversarial network (CCGAN) 4- Unsupervised or reference-less deep learning denoising models (in cases where HQ datasets are not available as a reference): Noise-to-void (N2V), and Noise-to-noise (N2N) 5- Hybrid (or semi-supervised) deep learning denoising models that we propose, in cases where only limited HQ long exposure datasets are available (exploring the possibility of reducing scanning costs while preserving the accuracy of quantitative analyses): Noise-to-noise – 75% (N2N75), Noise-to-noise – 50% (N2N50), Noise-to-noise – 25% (N2N25)