Machine Vision

Training computational attention on the physical object itself

Every system is measured against documented ground truth before its judgment counts. Machine vision at VeraCorpus is not about replacing the connoisseur's eye — it is about giving that eye a memory that never degrades and a consistency that never wavers.

The Problem with Human Vision

For centuries, attribution in the fine arts has relied on the trained eye of the connoisseur — a scholar who has spent decades examining surfaces, brushstrokes, pigment layers, and canvas weaves. This expertise is irreplaceable, but it is also fragile. It cannot be transmitted in full from one generation to the next, it is subject to fatigue and bias, and it cannot scale to the hundreds of thousands of unattributed works circulating in the global market.

Machine vision does not replace this expertise. It encodes it. Every model we train begins with a corpus of works whose attributions have been verified through traditional scholarship — documented, debated, and settled over decades. The machine learns not from assumptions, but from conclusions that have already survived scrutiny.

What the Machine Sees

Our systems are trained to attend to the physical signatures of an object at multiple scales simultaneously:

Macro-structural analysis — composition, spatial arrangement, and proportional relationships that characterise a workshop or period.
Surface texture mapping — the micro-topography of paint application, tool marks, and material deposition patterns invisible to the naked eye.
Chromatic fingerprinting — spectral analysis of pigment combinations and layering sequences that can date a work or identify a regional tradition.
Degradation pattern recognition — how a work has aged reveals as much about its origin as the work itself; craquelure, foxing, and patina follow predictable material pathways.

Ground Truth as Foundation

No model enters production without validation against a controlled dataset of works with established attributions. We maintain strict separation between training and evaluation corpora, and every system publishes its precision and recall metrics against documented benchmarks. A model that performs well on training data but cannot replicate expert consensus on unseen works is not deployed — it is retrained or discarded.

This discipline distinguishes research-grade machine vision from commercial image matching. We are not building a search engine for visual similarity; we are building a diagnostic instrument whose conclusions must withstand the same peer review as any scholarly attribution.

Current Research Directions

Our active programmes include cross-medium transfer learning, where models trained on panel paintings are adapted to works on paper and canvas; multi-spectral integration, combining visible-light analysis with infrared reflectography and X-ray data; and uncertainty quantification, ensuring that every machine judgment carries a calibrated confidence interval rather than a binary verdict.

We are also investigating the forensic potential of high-resolution imaging for detecting later interventions — overpainting, inpainting, and restorations that may alter the apparent authorship of a work.

← Return to Our Research