GEOSCAN Search Results: Fastlink


TitleApplications of end-to-end deep learning models in processing digital rock data
AuthorBizhani, MORCID logo; Ardakani, O HORCID logo; Little, EORCID logo
SourceGeoconvention 2021, abstracts; 2021 p. 1-2 Open Access logo Open Access
LinksOnline - En ligne
Alt SeriesNatural Resources Canada, Contribution Series 20210160
PublisherGeoConvention Partnership
MeetingGeoConvention 2021; September 13-15, 2021
DocumentWeb site
Mediaon-line; digital
File formatpdf
Subjectsfossil fuels; sedimentology; Science and Technology; Nature and Environment; models; petroleum resources; petroleum exploration; hydrocarbons; reservoir rocks; porosity; scanning electron microscopy; Methodology; Data processing; Artificial intelligence
Illustrationsflow diagrams; photomicrographs
ProgramGeoscience for New Energy Supply (GNES) Program Coordination
Released2021 09 01
Augmented Intelligence (AI) and Deep Learning (DL) techniques and the production of digital rock data have become powerful tools for studying and characterizing reservoir rocks properties - such as pore structure - at unprecedented resolutions. The rise of powerful imaging methods (e.g. micro-CT scanner) has enabled acquiring a substantial amount of image data of different rock properties under varying conditions. The caveats of digital rock data analysis techniques are the resolution and processing of these types of data. Image enhancement and segmentation are often time-consuming and subjective to the methods used. Deep learning as an emerging new means of image processing offers a great potential for automating image analysis tasks that can significantly speed up characterization. The research presented here demonstrates the use of deep-learning models for seamless processing of scanning electron microscopy (SEM) or micro-CT images of rock samples. Image denoising, resolution enhancement, semantic segmentation, and prediction of several properties such as porosity are shown to be successfully performed by these models. The automation of these tasks greatly reduces the time spent on each step. Additionally, as more data become available better models can be trained to push the boundaries of data characterization past physical limitations of currently available imaging devices.
We show through the use of several connected convolutional neural networks (CNN) that images acquired using SEM or micro-CT data can be directly characterized right out of the instrument using trained CNNs. In our work, a network first denoises the image, a second network takes the output of the denoiser and enhances the resolution by a factor of 4. The third network segments the super-resolved images. A final network is also attached to quickly obtain information about the pore size and porosity of the samples. The use of these models greatly reduces processing time. Additionally, it offers flexibility in terms of obtaining large lower resolution images and boosting their resolution through deep-learning models.
The novelty of this work is in the use of several state-of-the-art CNN models simultaneously to pre-process and produce quick results using raw microscopic rock images. We believe the industry, as well as the researcher, can benefit substantially by adopting these new tools in their analysis that can save time and enhance the workflow.
Summary(Plain Language Summary, not published)
Microscopy imaging is a powerful method for studying rock samples. However, images acquired using high-resolution techniques required several pre-and post-processing steps for quantitative analysis. The development of robust deep learning models along with access to computational power has transformed many industries in their quest for efficiency and speed. The same principle is applied in this work to automate processing microscopy images of rock samples for fast and accurate characterization purposes. The benefit of adopting state-of-the-art Augmented Intelligence (AI) and Deep Learning (DL) techniques for such tasks are speed, accuracy, and consistency of outcomes. Furthermore, it is possible to further improve the images beyond the physical limitations of imaging devices through deep learning methods.

Date modified: