Problem
Geologists document drill core by hand in a core storage facility: they measure core recovery, judge rock type, color and structure, and compile a report that is then loaded into geological software. The work is slow, and two specialists often describe the same core differently. Companies also hold large core archives that nobody has capacity to re-examine.
Example from the article: a well can have several hundred core boxes. Each box holds an arbitrary number of sections (the example box has 10), may be wet or dry, may be rotated, and is not always 1 m wide.
Solution in brief
A chain of CV models restores one continuous core from the box photos, predicts a few dozen features per section, and computes RQD through a segmentator rather than by direct regression. A small quadratic program then nudges the predicted numbers until they satisfy known linear relations. A web application shows every block's output, so a geologist can correct it.
Questions this case answers
- How to automate a manual measurement and description of physical samples from photos with a chain of computer vision models that an expert can check block by block
- How to make independently predicted numeric features consistent with known linear relations
- How to automate drill core documentation from box photos
- How to estimate RQD (Rock Quality Designation) from core images
Tools and methods
- YOLOv8 (
single_cls = True) — finds core sections in a box photo - Template-like augmentation (TLA) — swaps sections between boxes to enlarge the training set
- InceptionResNetV2 / ResNet — classifiers and regressors for section features
- Segment Anything Model (SAM), DeepLabV3+, UNet — segment core pieces, exclusion areas and areas of interest
- OSQP — solves the consistency problem for regression features
- FastAPI, React, Node.js, MongoDB, PostgreSQL, S3, Docker — application stack
Google or a one-time email code. No password.