The segmentation procedure in a digital rock physics workflow is often very challenging and time consuming. Here, we present an alternative method for quickly segmenting digital rock physics images that utilizes machine learning. Elemental SEM images of a core sample serve as inputs into a neural network. The network then outputs the probability of a pixel belonging to a certain class. Segmentation is implemented by choosing the class with the highest probability. This process allows for the uncertainty to be quantified in mineral phase identification. After training the algorithm, the rest of the image and subsequent images can be quickly segmented. We demonstrate the segmentation process on a shale sample with six different phases.
Presentation Date: Wednesday, September 27, 2017
Start Time: 3:05 PM
Location: Exhibit Hall C, E-P Station 4
Presentation Type: EPOSTER