Automatic identification of rock fractures based on deep learning

Abstract

Rock fractures are one of the main factors leading to rock failure. Accurately extracting fracture characteristics is crucial for understanding the rock failure mechanism. Inspired by the latest developments in computer vision, we introduce a state-of-the-art deep learning model YOLACT++ for the automated interpretation of rock fractures. YOLACT++ inherits the basic architecture of YOLACT (You Only Look At CoefficienTs) and optimizes the backbone network, which improves segmentation accuracy while ensuring real-time performance. Based on Unmanned Aerial Vehicle multi-angled proximity photography, the dataset is collected from various rocky slopes for model training and validation. We propose performance evaluation metrics for the model, including intersection over union, precision, and recall, as well as quantitative parameters for describing fractures, including orientation, trace length, roughness, aperture, spacing, and fracture intensity. The segmentation results of YOLACT++ are compared with two other classic instance segmentation models, the Mask Region-based Convolutional Neural Network (Mask R-CNN) and the You Only Look Once (YOLO) V8. The results show that YOLACT++ has a stronger generalization ability, with more accurate segmentation results at image boundaries. With the ResNet-101 backbone network, YOLACT++ achieves 93.8 %, 87.1 % and 92.2 % for precision, intersection over union and recall, respectively. This represents improvements of 5.4 %, 3.6 %, and 8.3 % compared to Mask R-CNN, and 3.3 %, 7.8 %, and 4.2 % compared to YOLO V8. Overall, the deep learning-based YOLACT++ model proposed in this study provides an efficient and reliable approach for the automated interpretation of rock fractures. It can also be applied to crack recognition in other materials.