Automatic fracture recognition and steady-state quantification in forward-looking borehole videos via deep learning

Abstract Automatic recognition and quantitative characterization of rock mass fractures are critical for surrounding rock classification, stability analysis, and geological hazard risk assessment. To address inconsistencies in frame-by-frame measurements caused by variations in visible scale and local morphology in forward-looking borehole videos, as well as geometric distortion from panoramic unwrapping that limits the accuracy of full-length quantification, this study introduces a raw video-domain approach for automatic fracture recognition and steady-state parameter quantification. A joint detection–segmentation strategy is implemented to obtain pixel-level fracture masks. In the video domain, continuous cross-frame trajectories of individual fractures are established by integrating feature matching and optical-flow tracking with a consistency verification criterion. Multi-frame fusion is then performed to convert instantaneous single-frame geometric measurements into steady-state video-domain outputs. Validation experiments on a self-constructed dataset of 55 borehole videos demonstrate that, on the validation set, the detection mP@0.5 and mR@0.5 reach 90.76% and 88.20%, respectively, while the segmentation Dice score is 82.51%. The cross-frame association success rate remains stable at 94.5%, with a drift error below 2.6 pixels, enabling continuous and reliable fracture trajectories. The fused geometric parameters exhibit strong agreement with manual measurements (R 2 > 0.995), and the quantification error decreases and converges as the number of fused frames increases. Compared with three panoramic image-domain quantification methods, the proposed approach reduces mean quantification errors in length, width, and orientation by 52.9%, 58.1%, and 45.2%, respectively, effectively mitigating distortion-induced errors from panoramic unwrapping. These results indicate that the proposed method enables stable video-domain fracture recognition, consistent cross-frame modeling, and steady-state parameter output, thereby providing reliable inputs for engineering fracture quantification and safety assessment decision-making.