Residual dense network and N2N significantly overestimate CO2 saturation. Despite the additional insight obtained from inspecting the gray-scale intensity histograms, histograms alone cannot be used to infer image quality. This transformation is based on the gray-scale intensity of each pixel and its neighboring pixels (Machado et al., 2013). The STANDS4 Network. We also compare computational resource requirements and ease of implementation of the different methods and provide recommendations. The neighborhood is pre-specified by the user as 6 faces, 18 edges, or 26 vertices. 1. IEEE Signal Process. Contrast-to-noise ratio (CNR) is a reference-less metric which yields a higher value for higher contrasts and lower values for indistinguishable foreground and background regions. For example, a denoising method that can offer accurate bulk properties might not be able to provide accurate pore-scale properties and vice versa. (a) Example regions of interest (ROI 15) considered for bulk and edge CNR calculations are shown in red boxes. Commonly used denoising filters in the digital rock physics literature, namely NLM and AD, show reasonable performance in terms of traditional denoising metrics like PSNR and SSIM. There are several other forms of noise including fractal noises like Perlin noise, periodic noise from helical sampling and finally random noise that may occur due to the change in X-ray quanta for multiple subject densities. In each of the above architectures, a state-of-the-art model has been identified and used for comparison against the traditional filter-based denoising techniques. Adv Water Resour. The following conclusions are derived based on the analysis of our results. Blind/reference-less image spatial quality evaluator (BRISQUE) helps understand the presence of noise in an image without the use of a reference (Mittal et al., 2011). The primary reason for the success of GANs, however, is its robust loss function which inherently accommodates several noise removal objectives like L1 distance, distribution matching etc. The purpose of a DAE is to remove noise. This calls for a more localized inspection of the images to assess how voxel spatial distribution impacts the accuracy of petrophysical property estimates. We propose a combination of the losses to reflect our semi-supervised task in training. Use idealized simulated or synthetic ground truth reference images and test the effect of different types and levels of noise on the performance of each of the denoising methods, as well as combinations of the denoising methods. (1979). Finally, the unsupervised models (N2V and N2N) require similar memory requirements, but N2V is approximately four times slower than N2N. (2019) trained a 20-layer feedforward neural network using synthetic data generated by the stochastic cross-correlation-based simulation (CCSIM) algorithm and found that neural networks perform better compared to bi-cubic interpolation for image super-resolution while the synthetic data further improved the model's generalizability. x ^ = arg max x P x y = arg max x P y x P x P y. Pet. The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Figure 6 shows example segmented cross sections from the HQ and LQ datasets. Comput. Machado, A. C. S., Machado, A. C. S., and Lopes, R. T. (2013). These advances allow for the visualization and accurate characterization of multiphase transport in porous media. The accuracy with which we can estimate these pore-scale properties affect our ability to explain and predict multiphase fluid flow in porous media. Vis. 37, 13481357. 40, 112. Med. Further, the application of machine learning and deep learning-based (DL) denoising models remains under-explored. 496, 5159. SPE 138591. doi: 10.1109/ISBI.2011.5872394. doi: 10.3997/2214-4609.201900074, Tian, C., Fei, L., Zheng, W., Xu, Y., Zuo, W., and Lin, C. W. (2020). Mach. The NLM filter was implemented using the adaptive-manifolds-based approach (Gastal and Oliveira, 2012). Data generation is achieved using a learnt, parametrized reverse process that performs iterative denoising, starting from pure random noise (see figure above). Although temporal accumulation introduces temporal lag, it doesnt produce blurriness. The use of MCT imaging is becoming increasingly indispensable in several disciplines, including geotechnical and petrophysical characterization and understanding multiphase flow in porous media. Al-Menhali, A. S., Menke, H. P., Blunt, M. J., and Krevor, S. C. (2016). The high pore pressure and temperature enabled the maintenance of scCO2 conditions. in Geology, Ghent University. doi: 10.1016/j.cageo.2012.09.008, Andrew, M., Bijeljic, B., and Blunt, M. J. The first one is the choice of an appropriate denoising method and its adaptation to ToF data, the second one is the issue of the optimal positioning of the denoising step within the processing pipeline between acquisition of raw data of the sensor and the final output of the depth map. Available online at: https://www.corelab.com/cli/core-holders/x-ray-core-holder-fch-series (accessed January 9, 2021). This can be explained by how a median filter works, where it involves replacing the gray-scale values of voxels within a user-specified window by their median value. The accuracy of the model predictions thus depends on the accuracy of processing the images that feed into these models. Solutions like NVIDIA Real-Time Denoisers (NRD) make denoising techniques more accessible for developers to integrate into pipelines. Computed Tomography: Principles, Design, Artifacts, and Recent Advances, 3rd Edn. doi: 10.1021/acs.est.6b03111, PubMed Abstract | CrossRef Full Text | Google Scholar, AlRatrout, A., Raeini, A. Q., Bijeljic, B., and Blunt, M. J. Rep. 7, 5192. doi: 10.1038/s41598-017-05204-4. doi: 10.1109/TPAMI.2020.2968521, Keywords: image processing, micro-computed tomography, deep learning, denoising, image enhancement, digital rock physics, carbon capture utilization and storage (CCUS), enhanced oil recovery, Citation: Tawfik MS, Adishesha AS, Hsi Y, Purswani P, Johns RT, Shokouhi P, Huang X and Karpyn ZT (2022) Comparative Study of Traditional and Deep-Learning Denoising Approaches for Image-Based Petrophysical Characterization of Porous Media. Intell. Automatic method for estimation of in situ effective contact angle from X-ray micro tomography images of two-phase flow in porous media. ImageNet classification with deep convolutional neural networks. Additionally, the error is increasing toward the bottom of the sample in most denoising methods, especially Gaussian, RDN, and N2V methods. Though these factors influence the choice of a model individually, we list combinations of those factors to offer model recommendations for different scenarios. 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. Scale-space and edge detection using anisotropic diffusion. 110, 157169. Unsupervised DL-based methods (N2N and N2V) generally exhibit poor performance for most properties compared to other denoising methods including some of the traditional filters. Transp. PyBRISQUE-1.0 software package (2020). Micro-Computed Tomography (Micro-CT) in Medicine and Engineering. The proposed algorithm that combines a detailed analysis of the noise statistics with an innovative autoencoder architecture is a promising path to denoise radio . It ranges from 0 to 1, where higher SSIM indicates a cleaner image. Lett. Experimental Investigation of Pressure Dependence of Contact Angle in CO2-Brine Systems. ABBREVIATIONS; ANAGRAMS; BIOGRAPHIES; CALCULATORS; CONVERSIONS; DEFINITIONS; GRAMMAR; LITERATURE . Each evaluation metric points to unique capabilities of the models. Master Thesis, Pennsylvania State University. Automatic measurement of contact angle in pore-space images. A comparative study of new and current methods for dental micro-CT image denoising. As such, the objective is to find the parameters that act on the sparse noisy data to obtain the closest representation of the clean signal. Connected pathway relative permeability from pore-scale imaging of imbibition. Image denoising aims to recover a clean image from a noisy observation. In this paper, we address these research gaps and compare the performance of traditional (user-based) denoising methods against more sophisticated DL-based denoising methods. The selection of an optimum denoising model cannot solely depend on the visual quality of the denoised image, or even on the standard denoising evaluation metrics. 22, 114. Get full control over your acoustic environment The only glass-related product successful in mitigating low-frequency noise like airplanes, trains, cars, etc. The supervised DL-based methods, except for N2C, show weak performance while the semi-supervised DL-based methods show slightly better performance. When denoising an image, its also important to keep visual details and components such as edges, corners, textures and other sharp structures. On the challenges of measuring interfacial characteristics of three-phase fluid flow with x-ray microtomography. Experimental investigation of residual saturation in mixed-wet porous media using a pore-scale approach. Indeed, 10 years ago, these achievements have led some researchers to suspect that "Denoising is Dead", in the . However, other supervised models like RDN and CCGAN and semi-supervised models like N2N50 and N2N25 showed weaknesses either on traditional metrics, petrophysical estimates, or computational requirements. Thesis, PhD. See our cookie policy for further details on how we use cookies and how to change your cookie settings. (2021) reviewed various DL applications in pore-scale imaging and modeling, including image segmentation, image super-resolution, petrophysical property prediction, flow simulation, as well as common convolutional neural network (CNN) architectures and various types of GANs for image generation. Temporal accumulation reuses data from the previous frame to determine if there are any artifacts or visual anomalies in the current frame that can be corrected. Improvements in the DL denoising field result from a clearer understanding of the specific noise forms and improvements in the architecture of neural networks. J. A steeper slope indicates a sharper boundary, which makes it more easily identifiable during image segmentation and petrophysical characterization. Spatial filtering selectively alters parts of an image by reusing similar neighboring pixels. https://doi.org/10.17612/A1QA-2A25. No use, distribution or reproduction is permitted which does not comply with these terms. Brief review of image denoising techniques - PMC - National Center for Larger error is seen at the top of the sample for all denoising methods except N2C and N2N75. doi: 10.1103/PhysRevE.94.043113. 1. The prominent ones include photon starvation, detector saturation also known as ghosting, central rotation artifact, cone-beam effect, metal artifact, cupping artifact, streaks and dark bands, under-sampling, poor contrast, beam hardening, scatter, and ring artifacts (Boas and Fleischmann, 2012). Imaging. Similarly, the intensity standard deviation (which determines the relationship between how fast the similarity value decreases based on voxel intensities) was set to 0.2. To perform the petrophysical characterization, we first segmented the datasets. The red region shows that some small solid features can be inaccurately characterized as fluid, resulting in porosity estimation errors, while the green region demonstrates a blurry fluid-fluid interface in the LQ image, which can lead to errors in estimating fluid saturation and petrophysical properties pertaining to fluidfluid interfaces. They characterized the 3D pore structure in a glass bead pack and three Berea sandstone samples to determine whether topological properties such as pore connectivity and phase features of individual fluid phases such as saturation can be resolved. Denoising provides users with immediate visual feedback, so they can see and interact with graphics and designs. We have also marked regions to highlight spatial feature differences between the HQ and LQ images and associated implications for petrophysical property estimation. Finally, we discuss the state-of-the-art methods for image denoising, . The model then learns to estimate the mean of the signal through several iterations. Figure 1 shows top view cross-sections from the HQ and the LQ data sets. Noise2void-learning denoising from single noisy images, in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Long Beach, CA), 21242132. (2018). Effect of fluid topology on residual nonwetting phase trapping: implications for geologic CO2 sequestration. Tawfik, M. S (2020). Transp. Denoising is necessary in real-time ray tracing because of the relatively low ray counts to maintain interactive performance. Noise-to-noise ratio had the lowest values across all the metrics. N2C (fully supervised) and N2N75 (semi-supervised with 75% HQ data) overall showed the most favorable outcomes. 60, 8490. Image denoising is to remove noise from a noisy image, so as to restore the true image. Ilastik: interactive learning and segmentation toolkit. Noise sources Test the hypothesis of whether the sequential use of N2N25 can improve image quality to a point where accurate results are achievable. Symp. Therefore, image denoising plays an important role in a wide range of applications such as image restoration, visual tracking, image registration, image segmentation, and image classification, where obtaining the original image content is crucial for strong performance. ACM Commun. Specifically, Gaussian, impulse, salt, pepper, and speckle noise are complicated sources of noise in imaging. We also proposed new semi-supervised denoising models (N2N75, N2N50, and N2N25) to assess the value of information and explore whether faster, lower exposure MCT images can partially substitute high-exposure datasets, which can be costly and can also hinder our ability to capture phenomena that occur at smaller time scales. We pick the epoch with the lowest validation PSNR in order to perform testing. Noise-to-noise ratio is a unique model which tries to estimate the unvarying signal behind varying noise realizations. Deep 2nd-order residual block for image denoising | SpringerLink A Review on Deep Learning Approaches for Low-Dose Computed Tomography Restoration. doi: 10.1016/j.jcis.2006.08.048, Purswani, P., Karpyn, Z. T., Enab, K., Xue, Y., and Huang, X. Those filters can be classified into linear and non-linear filters. Note the emphasis on the word . Sensors 21, 117. It is measured in decibels (dB) and the higher the PSNR, the better the image. Anyone whos taken a photo with a digital camera is likely familiar with a noisy image: discolored spots that make the photo lose clarity and sharpness. ROI 0 represents the background ROI against which contrast is calculated. Speech enhancement is an important task and it is used as a preprocessing step in various applications such as audio/video calls, hearing aids, Automatic Speech Recognition (ASR), and speaker recognition. Compare the performance of the different denoising methods using more complex petrophysical properties like surface roughness, tortuosity, interfacial curvature, in-situ contact angles, as well as fluid flow parameters such as absolute permeability, relative permeability, and capillary pressure. Micro-computed tomography has since been successfully used for quantifying a wide range of petrophysical properties such as volume fraction for porosity and saturation quantification, specific surface area (SSA), pore- and blob-size distributions, in-situ contact angles, interface curvatures for local capillary pressures, grain sphericity, angularity, roughness as well as phase connectivity (Sharma and Yortsos, 1987; Prodanovi and Bryant, 2006; Karpyn et al., 2010; Herring et al., 2013; Landry et al., 2014; Larpudomlert et al., 2014; Berg et al., 2016; Klise et al., 2016; Scanziani et al., 2017; Chen et al., 2018; Tawfik et al., 2019; McClure et al., 2020). (A) Porosity, (B) scCO2 saturation, and (C) brine phase fraction percent error (relative to HQ) along the length of the core for the different denoising methods. We also evaluated the denoising methods using physics-based metrics, or more specificallypetrophysical properties. The median filter on the other hand shows an opposite result where it exhibits better performance when estimating porosity and poorer performance in estimating scCO2 saturation. (2014). IEEE Trans. Reservoir condition imaging of reactive transport in heterogeneous carbonates using fast synchrotron tomography - effect of initial pore structure and flow conditions. 375, 187192. Denoising Basics of Image Processing - GitHub Pages In this paper, we present a comprehensive comparison of the performance of various image denoising methods which are broadly categorized as traditional (user-based) non-learnable denoising filters and DL-based methods. Image Denoising Using Autoencoders in Deep Learning - Omdena This comparison can enable optimum image processing workflow selection for creating accurate digital rocks that can be used for multiphase flow prediction and explanation. 118, 310320. Eng. Beyond Darcy's law: the role of phase topology and ganglion dynamics for two-fluid flow. The implementation of BI is detailed in the Appendix in Supplementary Material. Proc. A Micro-CT investigation into pore-Scale CO2 Sequestration Processes in Fratured Reservoir Rocks.
Carlsbad Cavemen Basketball Roster, 11419 Brithon Drive Orlando, Fl, Articles D