Chauhan, Swaroop (2019)
Phase Segmentation and Analysis of Tomographic Rock Images Using Machine Learning Techniques.
Technische Universität Darmstadt
Ph.D. Thesis, Primary publication
|
Automated Segmentation of X-ray Tomographic Imaging using Machine Learning Techniques -
Text
(PDF)
Dissertation_Chauhan_10052019.pdf - Published Version Copyright Information: CC BY-SA 4.0 International - Creative Commons, Attribution ShareAlike. Download (15MB) | Preview |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Phase Segmentation and Analysis of Tomographic Rock Images Using Machine Learning Techniques | ||||
Language: | English | ||||
Referees: | Sass, Dr.rer.nat Ingo ; Kersten, Dr.-Ing Michael ; Schill, Dr Eva ; Schüth, Dr Christoph | ||||
Date: | 2019 | ||||
Place of Publication: | Darmstadt | ||||
Date of oral examination: | 2019 | ||||
Abstract: | The topic of this thesis is the characterization of different phases and estimation of the geometrical parameters from the digital rock images, which are generated using high resolution X-ray computer tomography (XCT) experiments. High resolution X-ray computer tomography (XCT) is a well-established, long-standing experimental approach used in the rock physics community to study transport of the energy―momentum relationship inside porous- matrix domain. The accuracy and the appropriateness of the continuum based or topology based model prediction relies extensively on the resolution and phase segmentation of the XCT images. The current technology, used in nano tomography and micro tomography is able to generate high resolution image compared to the last decade, but new adaptive and flexible algorithm are urgently needed for accurate image analysis. Within the framework of this thesis different categories (supervised, unsupervised and ensemble classifiers) of machine learning (ML) techniques in combination with different image filtering techniques were investigated and tested for accurate XCT image segmentation and analysis. This lead to the investigation of seven different ML algorithm K-means, Fuzzy C-means (FCM), Self-Organized Map (SOM), Feed Forward Artificial Neural Networks (FFANN), Least Square Support Vector Machine (LSSVM), Bragging type ensemble classification tree (Bragging) and Boost (Boosting) type ensemble classification tree. Their respective clustering and classification performance and accuracy was compared and cross-validated. Thereafter, a robust workflow was developed to predict geometrical parameters such as porosity, volume fraction of different phases (pore, matrix, mineral) and pore size distribution. Further, a (standalone) grapical user interface (GUI) “CobWeb” was developed. The current version of CobWeb is capable to read and process (reconstructed) XCT files in tiff and raw format. Tools to zoom in, zoom out, cropping, color and scale, assist in the visualization and interpretation of XCT 2D and 3D stack data. Noise filters such as non-local means, anisotropic diffusion, median and contrast adjustments are implemented to increase signal to noise ratio. The user can chose from five segmentation algorithms, namely K-means, Fuzzy C-means (unsupervised), Support Vector Machine (supervised), Bragging and Boosting (enable classifiers) for accurate segmentation and cross-validation. Material properties like relative porosities, pore size distribution, volume fraction (pore, matrix, mineral phases) can be quantified and visualized. The data can be exported into different file formats such as Microsoft® Excel (xlsx), MATLAB® (mat), ParaView (vkt) and DSI studio (fib). The current version is supported for Micosoft® Windows and runs stable on Windows® 7 to Windows® 10. As ML techniques offer us high quality and accuracy w.r.t XCT segmentation. The future research should focus on comparing numerical simulation based on analytical modelling and molecular level approaches, such as pore network modelling and Lattice-Gas or Boltzmann methods respectively. CobWeb, has further scope of integrate different modules of point cloud data from LIDAR measurements, ultrasound data and acoustic emission data. Volume rendering plugin would be an important step forward good visualization. |
||||
Alternative Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-87869 | ||||
Classification DDC: | 500 Science and mathematics > 500 Science 500 Science and mathematics > 550 Earth sciences and geology |
||||
Divisions: | 11 Department of Materials and Earth Sciences > Earth Science > Geothermal Science and Technology | ||||
Date Deposited: | 17 Jun 2019 12:00 | ||||
Last Modified: | 09 Jul 2020 02:38 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/8786 | ||||
PPN: | 449952738 | ||||
Export: |
View Item |