Abstract: To address the problems of low efficiency, high risk, and limited automation in traditional fracture measurement methods, this study proposes a 3D characterization approach for rock fractures by integrating monocular depth estimation and instance segmentation. The improved DPT model is employed to achieve high-precision depth estimation, while the Mask R-CNN model is used for accurate segmentation of fracture regions and reconstruction of 3D point clouds. Based on the reconstructed fracture point clouds, algorithms are developed for calculating fracture length through skeleton extraction and cumulative Euclidean distance, fracture width through local principal direction projection and bidirectional nearest-neighbor search, and fracture attitude through least-squares fitting of trace point sets. Experimental results show that the mean relative errors of fracture length and width are 8.73% and 10.29%, and the mean absolute errors of dip and dip direction are 3.92° and 7.92°, respectively, meeting engineering accuracy requirements. This study realizes non-contact and automated acquisition of fracture geometrical parameters, providing an effective technical reference for digital measurement and intelligent identification of rock fractures in complex engineering environments.