Intelligent image-based identification and 3-D reconstruction of rock fractures: Implementation and application

This paper develops a method for intelligent identification and semi-automated extraction of 3D fracture parameters based on images in rock slope or underground engineering. Firstly, the exposed fracture trace is automatically identified by the enhanced deeplabv3 + model. Using transfer training methods, rock fracture images of multi-location and multi-type were used to train the model to improve the generalization ability. Secondly, rock sparse point cloud is generated using Structure from Motion (SfM) algorithm and dense reconstruction of rock is realized by Multi-View Stereo (MVS) method, which serve to transform the 2D fracture features to 3D. Finally, the coordinates of fracture feature points are obtained by fitting the planar features using algorithm. The methodology is validated through three types of data fields: (1) a simple and small-size laboratory installation is made up of two regular-shaped concrete blocks; (2) a slope with outcrop fractures in practical engineering; (3) a palm face from a tunnel project in Xinjiang, China. The error of the discontinuity parameters obtained using this method is about 1 %20 %, indicating that the proposed method is effective for parameter extraction and analysis of rock masses.