AI-supported analysis of archaeological structures with computer vision

Starting point

Archaeological investigations increasingly produce digital image and measurement data that can contain structures, layers and boundaries relevant to expert interpretation. Identifying these features reliably and preparing them geometrically often involves considerable manual work: experts spot relevant areas, trace them, check them and prepare them for further processing. With large or complex datasets, this becomes a time-consuming process.

The central question of the project was therefore: how can methods from computer vision and machine learning support the detection and preparation of archaeological structures without replacing human expert interpretation?

Challenge

The task went beyond classic image recognition, because usable geometric structures had to be derived from visual information. Several sub-problems had to be connected for this.

First, the varying image conditions: images differ in contrast, surface, capture situation and state of preservation, so rigid thresholds and simple filters quickly reach their limits. Second, the step from pixels to structures, because even detected areas do not yet make usable geometry: pixels and edges have to be connected into coherent structures and cleaned up. Third, the connection to existing expert processes: the analysis must not end in the model but has to deliver results in formats that existing scientific and technical tools can work with. And finally, expert control, because in a scientific context no result may be accepted unchecked as a black box; experts must be able to review, assess and correct detected structures.

So it was clear from the start: the goal was not a fully automated replacement for archaeological expertise but an AI-supported assistance system.

Solution: computer vision meets geometric processing

We developed a multi-stage approach that combines computer vision, machine learning and geometric data processing into one continuous chain: from the image data through structure detection and geometric processing to expert review and export.

1. Detect structures. Image segmentation and edge detection make relevant areas and boundaries visible in the images and produce a first machine interpretation of the image data.
2. Merge areas. The results, initially available as pixels, edges or points, are grouped into logically related areas using methods from clustering and classic data analysis.
3. Reconstruct geometries. Usable geometric objects are derived from the detected structures. This is the decisive transition from an analysis result in the image to a digital structure that can be processed further.
4. Review interactively. The methods come together in an interactive prototype in which domain users visualise results, follow intermediate steps and check detected structures. The expert assessment stays with people.
5. Export to domain formats. The prepared structures are converted into formats that can be reused in existing domain applications, depending on the use case for example Shapefile, DXF or PLY. The solution therefore does not end with a visualisation but connects to the existing workflow.

Technical approach

One special feature of the project is that we did not try to solve the whole problem with a single model. Instead, methods were combined according to their respective strengths. Deep learning approaches were used for detection and segmentation, including U-Net for pixel-based segmentation and HED (Holistically-Nested Edge Detection) for complex edges and boundary structures. Classic data analysis and geometric methods handled the further preparation, for example DBSCAN for grouping spatially related points and alpha shapes for reconstructing complex outlines.

The result is a hybrid processing chain: deep learning recognises visual patterns, classic algorithms structure the results, and geometric methods make them usable for domain systems. The real added value lies precisely where a pure AI model recognises patterns but the result still has to be brought into a usable form.

Prototyping instead of a black box

In research and innovation projects, it is rarely clear at the outset which method will work best under real conditions. We therefore deliberately did not build a rigid application around a single algorithm, but a prototype that makes different methods testable and comparable, keeps parameters adjustable, leaves intermediate steps visible and feeds domain users' feedback directly into further development. The prototype is thus a tool for development, evaluation and expert collaboration at the same time.

Outcome

The project shows how modern computer vision methods can be integrated into a highly specialised scientific workflow. The benefit arises on several levels. Machine methods detect and prepare relevant structures, so experts do not have to carry out every step by hand. An algorithmic analysis process applies the same rules reproducibly to different datasets, creating a more consistent starting point while the final assessment stays with the expert. The prototype environment shortens the path from a technical idea to a sound assessment. And because the detected information is prepared geometrically and exported to common domain formats, the result is not an isolated model output but a usable part of the real workflow.

This also reflects a principle that fits our approach: not every task should be fully automated. In scientific interpretation, domain knowledge has a value that a model does not replace. The technology takes over the repetitive, data-intensive steps, while the expert remains responsible for interpretation, validation and decisions. The result is not a replacement for expertise but a tool that lets expertise be used more efficiently.

Transferability to other industries

As specialised as the use case seems, the underlying problem is very general: relevant structures have to be detected in complex visual data, reconstructed geometrically and made usable for a downstream expert process. This class of problem appears in many fields:

• Industrial quality inspection: examining components and surfaces for cracks, defects or contour deviations.
• Structural analysis: detecting and documenting damage, geometric structures and changes in buildings.
• Geology and earth sciences: deriving layers and formations from image and measurement data.
• Materials science: segmenting microscopic images and measuring or classifying structures.
• Surveying and geodata: extracting objects and geometries from image, scan or point data and reusing them in GIS or CAD systems.

What this use case shows

The real proof of competence lies not in archaeology but in the ability to build a complete processing chain: understanding raw data, choosing suitable methods, combining ML and classic methods, preparing results geometrically, involving experts and integrating the results into existing systems. That last step is decisive, because a machine learning model alone does not solve a research or business problem. Value is only created when the model output becomes a usable part of the actual workflow.

Is there untapped information in your image or sensor data too?

Many companies and research institutions have image, measurement or sensor data whose analysis is still largely manual today. In a free initial consultation, we look together at which data is available, which analysis problem needs solving, whether computer vision or machine learning is suitable, which methods lend themselves to a first prototype and how the results can be integrated into existing processes and domain systems. The first step does not have to be a finished AI system; a targeted proof of concept is often the better way in.

Book a free initial consultation →