Çatalhöyük Image Repository

Contact: Claudia Engel, Academic Technology Specialist for Anthropology, Stanford Libraries

Background

Team: Ian Hodder, Justine Issavi, Claudia Engel, Dominik Lucas, Chris Chute

This project focuses on the image repository of the Çatalhöyük Archaeological Project (http://www.catalhoyuk.com). The images are being submitted for long term preservation in the Stanford Digital Repository (SDR). The additional metadata generated through this project will contribute to discovery and provide a model for future images processed in a similar way.

Recently inscribed on the UNESCO World Heritage List, the 9000-year-old neolithic settlement of Çatalhöyük in central Turkey is widely recognized as one of the most important archaeological sites in the world. Since 1993 an international team of archaeologists, led by Prof. Ian Hodder (Stanford), has been carrying out excavations and research. Since its outset the Çatalhöyük project stood out as an early adopter of innovative approaches in information technology and digital recording solutions. Today, after 25 years of research, the project has accumulated close to 5TB of data, including an image repository with a total of about 150,000 images.

The images are used to identify artifacts, to document the excavated objects in their excavation contexts, and to record the excavation process, which is by nature destructive. However, extracting the wealth of relevant information from these images for research remains a challenge for several reasons. In many cases, metadata recorded with the images are incomplete and inconsistent. Secondly, researchers require access to information captured in the images that is not contained in the metadata. And lastly, the number of images cannot be reasonably processed by hand.

Research

While machine learning has been applied in archaeology (Eramian et al 2017, Maaten et al 2017, Gansell et al 2007, Maaten et al 2007, Power et al 2017, Prasomphan & Jung 2017) it has typically focused on single objects and patterns to support archaeologists in their assessment and classification of individual archaeological finds. Information about the context these objects are found in, i.e. the relation among artifacts and between artifacts and their excavators, is largely inaccessible. Context and process information are fundamental for archaeological research (Hodder 1999) and the composition of archaeological photographs can be linked to specific practices of archaeology (Carter 2015). However, neither can be captured though traditional image pattern recognition. Furthermore, already existing models, as for example the Google Vision API, are of limited use for this project, as they are trained on non-archaeological data. Our in-situ excavation images are “messy”, they typically lack contrast, so objects are difficult to detect, are taken from different viewpoints and scale, and come in different resolutions as the quality of cameras varies over time.

Data

Source data: 150,000 digitized images. The images include people working, artifacts and photos documenting the excavation.

45,000 of the images have inconsistent or very incomplete metadata. The rest have minimal metadata.

The image collection includes objects of particular archaeological relevance ("X Finds") that are photographed separately in the lab under controlled conditions and can be used as training set to detect the same objects in the excavation photos.

Desired Output

1. To label the over 45,000 of images that lack valuable metadata, we plan to use (a) a subset of already labeled images in the database, (b) a subset of images labeled manually, and (c) a subset of images that were taken with a whiteboard that contains information about the object and photograph.

2. Object recognition: In order to be able to query images for "burial hole with skeleton", “bone with stone artifacts”, “walls with painted plaster”, we also plan to identify particular archaeological objects, like postholes, fire spots, bucrania, figurines, and more.

Experiments

1. Detect images with whiteboards,.
2. Extract the textual information and parse the text. This information will be used to train a model to tag the untagged images in our database.
3. Annotate the whiteboards to isolate and machine read the handwritten text on the whiteboards.