Digital rock physics has seen significant advances owing to improvements in micro-computed tomography (MCT) imaging techniques and computing power. These advances allow for the visualization and accurate characterization of multiphase transport in porous media. Despite such advancements, image processing and particularly the task of denoising MCT images remains less explored. As such, selection of proper denoising method is a challenging optimization exercise of balancing the tradeoffs between minimizing noise and preserving original features. Despite its importance, there are no comparative studies in the geoscience domain that assess the performance of different denoising approaches, and their effect on image-based rock and fluid property estimates. Further, the application of machine learning and deep learning-based (DL) denoising models remains under-explored. In this research, we evaluate the performance of six commonly used denoising filters and compare them to five DL-based denoising protocols, namely, noise-to-clean (N2C), residual dense network (RDN), and cycle consistent generative adversarial network (CCGAN)—which require a clean reference (ground truth), as well as noise-to-noise (N2N) and noise-to-void (N2V)—which do not require a clean reference. We also propose hybrid or semi-supervised DL denoising models which only require a fraction of clean reference images. Using these models, we investigate the optimal number of high-exposure reference images that balances data acquisition cost and accurate petrophysical characterization. The performance of each denoising approach is evaluated using two sets of metrics: (1) standard denoising evaluation metrics, including peak signal-to-noise ratio (PSNR) and contrast-to-noise ratio (CNR), and (2) the resulting image-based petrophysical properties such as porosity, saturation, pore size distribution, phase connectivity, and specific surface area (SSA). Petrophysical estimates show that most traditional filters perform well when estimating bulk properties but show large errors for pore-scale properties like phase connectivity. Meanwhile, DL-based models give mixed outcomes, where supervised methods like N2C show the best performance, and an unsupervised model like N2V shows the worst performance. N2N75, which is a newly proposed semi-supervised variation of the N2N model, where 75% of the clean reference data is used for training, shows very promising outcomes for both traditional denoising performance metrics and petrophysical properties including both bulk and pore-scale measures. Lastly, N2C is found to be the most computationally efficient, while CCGAN is found to be the least, among the DL-based models considered in this study. Overall, this investigation shows that application of sophisticated supervised and semi-supervised DL-based denoising models can significantly reduce petrophysical characterization errors introduced during the denoising step. Furthermore, with the advancement of semi-supervised DL-based models, requirement of clean reference or ground truth images for training can be reduced and deployment of fast X-ray scanning can be made possible.
All Science Journal Classification (ASJC) codes
- Water Science and Technology