Abstract
Acquisition of joint attitudes is vital in mine geology but often constrained by underground conditions, while manual cataloging remains inefficient and subjective. To overcome these issues, we propose a mobile phone photography and deep learning-based method. Rock joint images are collected with smartphones, augmented by cutting and rotation, and enhanced using CLAHE. After labeling with Labelme, a dataset is built for training. A ResNet residual module and CBAM attention are integrated into a U-Net architecture, forming the RC-Unet model for accurate semantic segmentation of joints. Post-processing with OpenCV enables contour extraction, and the PCP three-point localization algorithm rapidly calculates joint attitudes. A practical engineering case verifies that intelligent joint identification can replace manual cataloging in relatively simple underground environments. This approach improves efficiency, reduces subjectivity, and provides a rapid, low-cost, and easily storable means for geological information acquisition, highlighting its potential as an effective tool and supplementary method for mine surveys.