Investigation Into Deep Learning Methods for a Formal Analysis of Ancient Sculptures

The formal analysis of ancient sculptures plays a crucial role in understanding historical contexts, cultures, and societies. In-depth archaeological studies often rely on mostly manual, time-consuming and comprehensive visual analysis conducted by specialists in the field. This work explores the integration of deep learning methods, based on neural network architectures, for segmenting facial components and classifying portraits of the first four Roman emperors of the Julio-Claudian dynasty, with the aim of supporting archaeologists in their work. The segmentation is evaluated objectively using the Intersection over Union metric (IoU), with scores around 0.85, in combination with qualitative visual analysis via image masks. For the classification task, F1-scores of approximately 0.78 are achieved for emperor recognition, and 0.79 for portrait type classification. Furthermore, highlighting the decisive regions used by the network for classification offers a new perspective on the importance of facial regions compared to the traditionally considered hair patterns.