Automated Core Fracture Characterization by Computer Vision and Image Analytics of CT Images

Abstract

Abstract Computer vision (CV) techniques were applied to X-ray computed tomographic (CT) images of reservoir cores to evaluate their potential for rapidly identifying fractures and other lithologic/geologic characteristics. The analyses utilized feature-labeled CT cross sectional images, themselves distributed submillimeter voxels of density- and atomic number-sensitive CT Numbers, as inputs to a Fully Convolutional Neural Network (FCN) for semantic segmentation of reservoir core features. In FCN, an image is interrogated using a series of sliding windows at various scales to create weighted filters to reduce error between classes in training images. These networks of filter layers were used to assign probabilities of classes, which were upscaled back to the original image dimensions resulting in probabilistic class assignments onto each pixel. FCN model accuracy, defined by its ability to replicate manually-assigned labels in the raw (unannotated) training image stack, was at least 80% and generally improved with the size of the training set. Once the labels were assigned, the underlying feature frequency, orientation, and size were measured in 3D volume reconstructions using algorithms modified from standard image analysis software. This method allowed users to endow a classification model with subject matter knowledge for further, autonomous label prediction. Thus, while initial image annotation was labor intensive, subsequent images were rapidly classified once the model was built. The classified labels were analyzed for abundance, orientation, and size of fractures were calculated to characterize spatial information of these features. FCN combined with fracture labeling improved knowledge capture and automation of fracture identification. Models trained by high quality 3D datasets can greatly reduce the time needed to describe subsequent core. The method demonstrated is not limited to fractures, other lithologic/geologic features could be trained using the same method, which may result in additional efficiencies.