Automatic Fracture Segmentation and Detection From Image Logging Using Mask R-CNN

Abstract

Image logs provide a wide range of information about petrophysical properties and geological features of reservoirs. The identification of fractures by image logging is very important for the precise prediction of production and the accurate evaluation of oil and gas. However, the interpretations of underlying features from fracture occurrences, which could be crucial for experts in fields, are relatively rare. Nowadays, deep learning networks, used to learn representations of image with diverse levels of abstraction, could perform well for understanding the intrinsic features of image log data. In this study, we proposed a deep learning method called Mask R-CNN to recognize the features of fractures based on the datasets of image logs. This deep net detects and segments each fracture individually by focusing on local information of image logs. It provides a novel way for experts and researchers to identify and quantify the fractures precisely and then calculate parameters of fractures efficiently. The applied model contains two parallel branches to recognize and segment fractures respectively. The first workflow, following the idea of Faster R-CNN, is used to track the positions of fracture through the Region Proposal Networks (RPN) and two regression networks. The other branch performs a Fully Connected Network (FCN) to implement up-sampling and output the mask of fractures from image log data. These branches both accept inputs which are based on the same feature maps via the modified Feature Pyramid Networks (Feature Pyramid Networks). The FPN is used to extract features with various scales. To obtain dataset with high quality, we annotated the fracture by manual and implemented data augmentations. All kinds of labeled fractures are marked as mask images in which the pixels 0, 1, 2 and 3 stand for background, induced fractures natural fractures and bedding separately. By the mask image with pixel-wise labels, the dataset with 518 images was used in this paper. Overall, the proposed method in this paper achieves ideal performance to detect the fractures and beddings with the average precision of over 75%. Based on the identification result, we calculate parameters of fractures, such as dip angle. As a consequence, the method in this work shows its potential for identifying all the significant information in borehole through image log data.