True-scale mapping of rock discontinuities from single images without calibration.

Digital photogrammetry is widespread in site investigation of rock discontinuities. However, existing methods often rely on multi-frame synthesis techniques and calibrating references to achieve three-dimensional (3D) modeling and orientation analysis of rock discontinuities, which limits the rapid application of digital photogrammetry techniques onsite. To address this challenge, this paper introduces a hybrid machine learning architecture called OneShot-DisconNet. This method first utilizes deep neural networks trained on the MegaDepth dataset to predict single-layer depth map from monocular images. Subsequently, multi-viewpoint depth fields and textures is reconstructed on the layered depth images (LDIs) using Partial Convolutions (PConv), enabling point cloud modeling of rock surfaces. Additionally, a novel uncertainty-based multi-population genetic algorithm-driven backpropagation (UB-MPGA-BP) neural network is developed based on the MegaDepth dataset to enable true-scale mapping of 3D point clouds without references. Finally, by addressing the limitations of the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm in handling outliers and coplanar point clouds, the robustness of the structural analysis results has been enhanced. Case studies in the Beiyi Section of Shilu Iron Mine and the Rockbench database demonstrate that the proposed method is comparable in performance to Lidar-based digital photogrammetry technology, enabling true-scale mapping and orientation analysis of rock discontinuities in various scenarios. [ABSTRACT FROM AUTHOR]