From object detection to archaeological object detection developing a model for amphora identification of a Punic wreck site using object recognition AI

Abstract

The last years have seen an increase in the use of object detection methodologies in land archaeology during surveys, the study of archaeological assemblages, reconstruction of archaeological materials, and taphonomic studies. The advances of these methods in maritime archaeology have been more limited. This study explores how AI object detection can help identify archaeological materials underwater. It aims to explain the issues that the underwater environment presents for automated detection, to bridge the knowledge gap that exists between the practical application of this computer vision technique to maritime archaeology, and to provide a practical example of its application on the underwater assemblage of Xlendi Archaeological Park, one with which to evaluate the possibilities that the use of such methodology presents for archaeological research. We trained, classified and named a total of seventy-two detection models based on three differentiating factors. The Progressive Complexity Index (PCI) divides them into groups based on their level of complexity and the amount of archaeological information embedded in their predictive process. The Parameter of Archaeological Identification (PAI) specifies the archaeological framework used during training to teach subjective information. Finally, the models are also different in the version and size of model they use. To fulfill the goals of this project, we used these differences to interpret the results of a series of comparatives tests made on data not seen by the algorithm during training, thus recreating a real-world situation in which to evaluate the technique. The result is the division of the models into three progressively complex groups: nature models, state models and typological models. Nature models focus on the assessment of underwater archaeological assemblages by the nature of the materials to be found in them, classifying them based on them being ceramic, litter, modern elements, or part of the natural background. On their best iterations, these achieved an average precision of identification of 87.8%. State models, on the other hand, focus on the state of preservation of those materials. Their best iteration’s average precision, while lower at 75.2%, still produced very usable models on a real-world scenario. Finally, typological models focus on ceramic materials based on their typology. Their best iteration, while not being field-ready with an average precision of 61.1%, offers a lot of potential for improvement. This dissertation has demonstrated how subjective archaeological information can be integrated into YOLO models to develop detection models tailored to specific archaeological questions. By analyzing and comparing these models, it has outlined the technique’s fundamental applications, limitations, and future potential for studying underwater archaeological assemblages.