CoreLog Vision: Lithology from Drill-Core Imagery, with Honest Sim-to-Real Out-of-Distribution Detection
Logging the lithology of drill core, identifying the rock type down a borehole from the physical core in its trays, is slow, subjective manual work, and automating it from core-tray photographs is an obvious target for computer vision. The catch is data: labelled core photographs are scarce, so a model is often trained on synthetic or one operation's core and then asked to read another's, and it will do so silently and wrongly unless it can tell that the new core is unlike its training data. CoreLog Vision is an in-browser pipeline that classifies core-tray imagery into lithologies and, crucially, detects when it is looking at core it was not trained on. On synthetic core its convolutional classifier reaches 99.4% accuracy over six lithologies against a classical colour-and-texture baseline's 92.9%, on a leakage-safe split that holds out entire boreholes (not random patches). On real core, from the open DCID-7 drill-core image dataset, a dedicated MobileNetV3 head reaches 99.2% over seven lithologies, while the synthetic-trained features transfer to only 88.8%, the honest sim-to-real gap. The report's central result is the detector that guards that gap: nine out-of-distribution methods are benchmarked at telling real DCID-7 core from synthetic training core (both funnelled to the same 24-pixel window so no detector can cheat on resolution), and feature-space detectors succeed (the best at area-under-precision-recall 0.999, the shipped one at 0.92) while a reconstruction-based detector fails (0.28). The lesson, know what you do not know, and do it in feature space, is the report's contribution; every number and figure regenerates from the committed artifacts. Interactive tool (MIT): https://github.com/fsantibanezleal/CAOS_CoreLog ; live at https://corelog.fasl-work.com .