Vision AI for cultural heritage is not a demo problem

Most Vision AI demos are clean. A model sees a clear object, draws a box and the result looks obvious.

Cultural heritage data is not like that.

When I worked on repetitive pattern recognition in ancient Peruvian pottery, the hard part was not only training a model. The hard part was using computer vision without pretending that the model understood archaeology.

The research problem

The goal was to identify repeated decorative motifs in pottery textures.

That sounds like object detection, but the context changes everything. The image is not a neutral input. It may come from a photograph, a texture extracted from a 3D object, a crop selected by a researcher or a surface affected by erosion and missing fragments.

The model is looking at a representation of an artifact. That means its output must be treated carefully.

For me, the useful question was:

Can a model help researchers inspect visual repetition faster?

Not: can the model replace interpretation?

What the model is good for

The model is useful as an attention tool.

It can scan textures and suggest areas where repeated motifs may appear. That reduces the time spent searching manually and gives researchers a starting point for review.

The most useful output is not “this is the final answer.” The most useful output is “look here first.”

That difference is important. It keeps the workflow honest.

Why this is harder than a normal dataset

Archaeological visual data has ambiguity built in.

A motif may be damaged. A line may be decorative noise or part of a figure. A repeated shape may appear at a different scale. A texture seam may create a false visual pattern.

That makes annotation a research decision, not only a labeling task.

If two experts disagree about a region, the dataset should not hide that uncertainty. The model will learn from whatever rule is written down.

How I used it

I treated the model as part of a review loop:

1. Train on annotated motifs.
2. Run predictions over texture images.
3. Review high confidence and low confidence results.
4. Identify repeated mistakes.
5. Improve annotation rules and dataset examples.

This is where the work becomes interesting. The model helps reveal weak parts of the dataset and unclear parts of the visual definition.

How other projects can use the same idea

This research pattern can transfer to many domains where experts inspect repeated visual structures:

• textile analysis
• ceramic fragment comparison
• historical document ornament review
• industrial defect grouping
• satellite texture exploration
• museum collection search

The shared idea is simple: use computer vision to surface candidates, then keep expert review in the loop.

What I recommend

Start narrow. Pick one visual pattern and make the review process useful before expanding classes.

The first milestone should be a tool that lets someone inspect predictions quickly. A model with a strong metric but a poor review workflow is not very useful.

I would prioritize:

• clear annotation rules
• examples of ambiguous cases
• review of false positives
• review of missed motifs
• visual comparison across similar artifacts

That is where research value appears.

What feels innovative

The innovation is not that a neural network can detect patterns. The innovation is turning visual recognition into a research assistant that helps organize evidence.

It can compress the search layer. It can reveal where definitions are weak. It can help compare collections. But it should not claim authority over cultural interpretation.

That balance is the part I care about.

References

• Repetitive Patterns Recognition in Textures of Ancient Peruvian Pottery
• ACM Journal on Computing and Cultural Heritage
• Ultralytics training docs
• Ultralytics detection dataset docs

The strongest version of this work is not a demo. It is a review system that makes visual evidence easier to inspect and question.