Deep learning for automated rock core image analysis: Weathering classification, fracture detection, and 3D geological modelling

Abstract

Rock core images are a primary source of subsurface information, but conventional manual logging is slow, labour-intensive and prone to subjective error. This study presents CoreVision, an end-to-end data-driven workflow for automated interpretation and digitisation of drilling-core imagery in linear underground infrastructure projects. The workflow adapts established computer-vision and OCR backbones, including YOLOv5, DINOv2, and Tesseract, while introducing task-specific advances for heterogeneous rock-core imagery, continuous RQD reconstruction, and fracture morphometry. At the macroscopic scale, Core_YOLO modifies the YOLOv5 FPN + PAN architecture by adding a coarser feature map tailored to elongated core segments, improving the detection of lithological and structural variability in rock-core imagery. On a benchmark of 50 core-box images representative of Melbourne Silurian sedimentary rocks, Core_YOLO raises the mean Average Precision from 0.710 (YOLOv5) to 0.889. The detected bounding boxes feed an automated spatial-reconstruction algorithm that converts 2D detections into continuous Rock Quality Designation (RQD) logs, achieving R 2 = 0.84, RMSE = 10.67 and MAPE = 18.59% against expert-derived ground truth, with 76% interval-classification accuracy across the four standard RQD bands. At the microscopic scale, a DINOv2-based few-shot segmentation module, trained on only ten annotated prototypes, extracts fracture trace lengths and apparent orientations at the pixel level, removing the need for exhaustive pixel-wise labelling. In parallel, an optimised Tesseract-based OCR pipeline with region-of-interest cropping, binarisation and a rule-based error-recognition stage digitises core-box metadata, lifting overall recognition accuracy from 95.2% to 98.0% and depth-position recognition from 89.2% to 95.5%. The integrated outputs, comprising continuous RQD profiles, fracture geometries, and georeferenced depth records, provide a quantifiable multi-scale data foundation for 3D geological modelling, tunnel design, and digital-twin development.