The ubiquity of Roman emperor portraits in the ancient world, serving as visual proxies for imperial leadership, is well-documented. Traditional identification methods primarily focus on coiffure patterns, but this approach risks excluding portraits that deviate from standardized representations. This article explores the potential of facial recognition software as an alternative, aiming to provide a new empirical method for identifying Roman emperors in portraiture, overcoming limitations of the coiffure-based approach. The research question is: Can existing AI-based facial recognition methods be used to identify portraits of Roman emperors?

The existing literature highlights the limitations of relying solely on coiffure patterns for identifying Roman emperors. Scholars have noted that portraits lacking characteristic hairstyles are often excluded from studies, potentially leading to an underrepresentation of variations in emperors' appearances. The discovery of "unofficial" portraits further underscores the need for a more inclusive identification methodology. The current debate has reached a stalemate, highlighting the need for a new empirical approach.

The study uses transfer learning with deep convolutional neural networks (CNNs). A pre-trained Inception-ResNet-v1 model, initially trained on a large dataset of human faces (VGGface2), is adapted for Roman imperial portraits. A dataset of 673 images (645 of nine emperors and 28 non-emperors) was compiled from online datasets, catalogues and original photographs. The images were split into training, validation, and test sets. Data augmentation techniques (horizontal flip, shift-scale-rotate, random brightness contrast, multiplicative noise) were applied to increase the training dataset size. Two experiments were conducted: Experiment 1 retrained the final two densely connected layers of the model; Experiment 2 further fine-tuned the model by retraining the last convolutional layer of the feature extractor. The F1-score and confusion matrices were used to evaluate model performance. UMAP dimensionality reduction was used to visualize the model's embeddings.

Experiment 1, a baseline transfer learning approach, achieved an accuracy of 81.1%. Experiment 2, fine-tuning the last convolutional layer, improved accuracy to 89.2%. The weighted average F1-score for Experiment 2 was 0.95 on the validation set and 0.90 on the test set. UMAP visualization showed clear clustering of images by emperor, indicating the model's ability to distinguish between different emperors. While not reaching state-of-the-art face recognition accuracy (low 99%), the results demonstrate the potential of facial recognition for identifying Roman imperial portraits, even with a limited dataset.

The findings demonstrate that facial recognition techniques, despite the challenges posed by the characteristics of ancient portraiture (lack of texture, color, presence of damage, and variations in artistic style), can be successfully applied to identify Roman emperors. The improvement in accuracy between Experiment 1 and Experiment 2 highlights the importance of fine-tuning the model for the specific characteristics of the dataset. Some misclassifications may be attributed to inherent similarities in the portraits of some emperors due to deliberate stylistic choices ("Bildnisangleichung") and the influence of fashion trends on representation. Damage and restorations on the portraits also contributed to some misclassifications. However, the overall performance suggests that facial recognition offers a valuable tool for the study of Roman imperial portraiture.

This study successfully demonstrates the feasibility of using facial recognition techniques to identify Roman emperors from their portraits. While the accuracy is not yet at the level of state-of-the-art face recognition systems, the results are promising and provide a foundation for future research. Future work should focus on expanding the dataset, incorporating more sophisticated image preprocessing techniques to address issues such as damage and restoration, and exploring the potential of this approach to distinguish between emperors and non-emperors.

The study is limited by the relatively small size of the dataset. The availability of high-quality images of Roman imperial portraits is constrained by the condition of the surviving artifacts and their photographic documentation. The model's performance could be further improved by incorporating more sophisticated image preprocessing methods to mitigate the effects of damage and variations in artistic style. The current study primarily focuses on identifying emperors; future work should address the more challenging task of distinguishing between imperial and non-imperial portraits.