CobWeb - a toolbox for automatic tomographic image analysis based on machine learning techniques: application and examples

TL;DR

CobWeb 1.0 is a user-friendly MATLAB-based toolbox that automates tomographic image segmentation and analysis of digital rock images using machine learning techniques, enabling efficient characterization of geomaterials.

Contribution

Introduces CobWeb 1.0, a GUI tool integrating multiple machine learning segmentation methods for digital rock image analysis, with validation on real datasets.

Findings

Successful segmentation of phases in complex datasets

Effective normalization of edge artefacts

Versatile export options for further analysis

Abstract

In this study, we introduce CobWeb 1.0 which is a graphical user interface tailored explicitly for accurate image segmentation and representative elementary volume analysis of digital rock images derived from high resolution tomography. The CobWeb code is a work package deployed as a series of windows executable binaries which use image processing and machine learning libraries of MATLAB. The user-friendly interface enables image segmentation and cross-validation employing K-means, Fuzzy C-means, least square support vector machine, and ensemble classification (bragging and boosting) segmentation techniques. A quick region of interest analysis including relative porosity trends, pore size distribution, and volume fraction of different phases can be performed on different geomaterials. Data can be exported to ParaView, DSI Studio (.fib), Microsoft Excel and MATLAB for further visualisation and statistical analysis. The efficiency of the new tool was verified using gas hydrate-bearing sediment samples and Berea sandstone, both from synchrotron tomography datasets, as well as Grosmont carbonate rock X-ray micro-tomographic dataset. Despite its high sub-micrometer resolution, the gas hydrate dataset was suffering from edge enhancement artefacts. These artefacts were primarily normalized by the dual filtering approach using both non-local means and anisotropic diffusion filtering. The desired automatic segmentation of the phases (brine, sand, and gas hydrate) was thus successfully achieved using the dual clustering approach.