RA3I

Rock Art Analysis; Artificial Intelligence; Computer Vision; Archeological Documentation; Automated Tracing; Pattern Recognition; Cultural Heritage; Explaninable AI

Ra3i-Rock Art Analysis with Artificial Intelligence

Short tagline: From photograph to classified motif.

One-line summary:
Ra3i is a working AI-powered pipeline that converts rock art photographs into archaeologically usable tracings, and automatically detects and classifies motifs, developed by Techframe in consortium with Instituto Politécnico de Tomar (IPT) and Instituto Terra e Memória (ITM).

Product Description

Rock art documentation has traditionally relied on manual tracing and classification, slow, labour-intensive work, vulnerable to inconsistency between methods. Ra3i is a working system, ready to use and continuously evolving, that automates this process while keeping every decision transparent and auditable.

The pipeline moves through three integrated modules:

• Feature Mining — extracts pigment signatures, deep visual features, regions of interest, and surface/morphological descriptors from the original photograph, producing a structured context package.
• Automated Processing — a retrieval-based engine selects and sequences the right combination from a library of 40 processing techniques (enhancements, tracing, noise cleaning, overlay) for each image, learning from past successful cases and generating a fully auditable process dossier for every output.
• Detection & Classification — two complementary YOLO-based architectures work in parallel across original images and tracings, reconciled in a model agreement layer, classifying motifs into five classes (anthropomorph, zoomorph, geometry, lines, other) with confidence scores.

Why it matters

• It collapses the laboratory bottleneck. Tracing that required hours of specialist work is ready in seconds, returning time to the field, where the record is actually captured, particularly urgent for sites at risk of loss or degradation.
• It makes methodologies reproducible and citable. A processing plan can be shared between researchers and re-run on different datasets, and its reference can be placed in a paper, taking an authenticated reader straight to the shared results.
• It preserves authorship and controls sharing. Work is signed, provenance is kept, and sharing between peers, institutions and the public happens only by the author's permission, with an audit trail.
• Every output is explainable. Intermediate images, confidence scores and a full process dossier accompany each result, so the specialist verifies and adjusts rather than accepting a black box. The AI assists; it does not decide, and no automatic classification becomes an authoritative record without human review.
• It turns documentation into leverage. Ra3i does not preserve sites, no software does. By documenting them with auditable precision, it gives researchers and institutions the material to argue for preservation with those who decide.

Metrics

• Motif segmentation: 0.873 mask mAP@50
• Pigment detection (4 classes): 0.900 macro-F1
• Median processing time: ~18 seconds per panel
• 40+ distinct automated processing techniques in the library

All models were trained on domain datasets built specifically for this purpose; no pre-existing rock art benchmarks were used. Figures refer to rock paintings. The system also processes engravings, painted and engraved surfaces are distinguished at intake, although engraving support is a capability in development, with dedicated training on the roadmap.

Who it's for

Research centres, heritage foundations, archaeologists and academic researchers working with rock art documentation at any scale, from single-panel studies to large regional datasets.

Consortium

Developed by Techframe – Sistemas de Informação SA, in partnership with Instituto Politécnico de Tomar (IPT) and Instituto Terra e Memória (ITM). Co-financed by COMPETE 2030, Portugal 2030, and the European Union.

Learn more: ra3i.org