Multi-phase Segmentation of Digital Rock Images Using Convolution Neural Network: Training Dataset Generation, Model Training and Result Visualization

Abstract:

This paper applied the most successful convolution neural network (CNN) model in image segmentation, U-net, to segment X-ray computed tomography (CT) images of a Berea sandstone. The Berea CT images consists of 500 two-dimensional slices of 981×981 pixels obtained at a resolution of 2.36 µm; first when the sandstone was saturated with water, and then displaced by super-critical CO₂. A workflow including pre-processing, training dataset generation, CNN model explanation, result visualization, and final segmentation results is shown in this paper.

In recent years, machine learning has provided novel solutions to current challenging problems in different fields. Among them, CNN has shown significant successes in image segmentation. The primary advantage of CNN compared to conventional image segmentation methods is that CNN can learn from the textures and patterns of the training images, rather than the grayscale value only. However, the application of CNN in Digital Rock Physics (DRP) is not that extensive and CNN is just a black box for most DRP researchers. For a better understanding of CNN, this paper shows a detailed explanation of components in U-net and clear visualizations of layer results. Also, the regular U-net suffers from the problems of compressing phase boundary information and making small objects invisible. To delineate the phase boundary more accurately, we added additional feature maps, including an information entropy map and gray gradient map, to the input data. To improve accuracy, we added one extra training step for adjacent top and bottom slices to account for the vertical connectivity of the core sample. Another novelty of this work is the application of an entropy-assisted kriging method to generate training datasets. The training datasets are the key to the performance of the CNN model because the highest possible accuracy of the CNN model is determined by the accuracy of the training datasets. However, access to high-quality training datasets is usually limited. The steps of the entropy-assisted kriging method included entropy-based masking, Gaussian mixture fitting, thresholding value determination, and indicator kriging estimation. This method segments the images using global information and local spatial correlation; it has satisfactory accuracy but was computationally expensive.