Texture segmentation of 3D x-ray micro-computed tomography images using U-NET

Recent advances in numerical methods combined with the use of 3D X-ray micro computed tomography acquisition systems improved the characterization of reservoir rocks at pore scale. This approach, known as digital rock physics (DRP), consists of simulating rock properties using 3D X-ray micro computed tomography images at pore scale. DRP has been extensively used to estimate numerically rock properties like porosity and permeability. This methodology was successful in sandstone reservoir rocks due to their relative homogeneity. Nevertheless, this approach failed in many cases when applied for carbonate reservoirs due to their heterogeneity at several length scales. In order to overcome this limitation, we propose to use the texture information in the images to identify and segment the most representative textural regions. Indeed, several studies showed that texture information is correlated to variations of physical rock properties. In recent years, the advancements made in deep learning algorithms improved largely the performance of segmentation methods. In particular, we focus on a machine learning method based on convolutional neural network called the U-NET architecture to segment 3D X-Ray micro computed tomography images in terms of textures. The challenge is to identify precisely representative textures in highly heterogeneous rocks such as carbonate rocks. We investigate the performance of the proposed segmentation method on both synthetic and real data.