Automatic fracture characterization in CT images of rocks using an ensemble deep learning approach

Abstract

The presence of fractures in a rock mass can have a substantial influence on its mechanical and hydraulic properties. For many years, computed tomography (CT) scan has been effectively utilized to investigate the internal structures of rock. However, quantitative characterization of fracture based on CT data remains a challenging endeavor due to the inevitable blurry appearance of fractures in CT images, complexity in fracture patterns and the heterogeneity of host rock. In this study, a deep learning-based method is presented for automatically and accurately segmenting rock fractures in CT images, with special consideration for directional ambiguous fractures embedded in heterogeneous backgrounds. Our method involves two stages: (1) fracture detection in the form of minimal bounding boxes using the Faster R-CNN deep learning algorithm, and (2) fracture segmentation inside the detected bounding boxes using the U-Net deep learning algorithm. The detection stage aims to establish spatial constraints for the segmentation process, thereby significantly minimizing undesirable noise and artifacts existing in the background from consideration and consequently improving the segmentation result. In addition, we also develop a system that can automatically extract fracture properties including orientation, length, aperture, and wall roughness of fractures based on detection and segmentation results from our method. Experiments on CT images of fractured coarse-grained granite, fine-grained sandstone, and shale core samples were carried out in an attempt to confirm the applicability of our approach. The comparison between fracture segmentation results and the manually-annotated ground truths shows that our approach can achieve a competitive dice score of up to 0.942 and also outperform other state-of-the-art deep learning approaches including Mask R-CNN and U-Net alone.