Facial recognition as a tool to identify Roman emperors: towards a new methodology

The study addresses the recognisability and identification of Roman imperial portraits, traditionally determined by coiffure patterns. This prevailing method risks excluding portraits that deviate from standardized imperial imagery, potentially obscuring variation and misrepresenting the diversity of imperial representation. The authors propose exploring facial recognition software as an alternative tool to identify emperors from sculpted portraits. The central research question is whether existing AI-based facial recognition methods can be adapted to identify Roman emperors from their portraits. The context includes widespread imperial imagery across coinage and sculpture, production via prototypes leading to standardized features, and scholarly debates about pitfalls of hair-based identification. The purpose is to provide a new empirical foothold that does not rely on predetermined criteria like coiffure, and the importance lies in potentially expanding and refining corpora of imperial portraits by recognizing faces despite damage, restorations, and stylistic variation.

Art historical methodology has long identified Roman emperors using coiffure patterns and standardized features derived from prototypes (Fittschen; Pfanner), acknowledging widespread replication in imperial portrait production. Critics (e.g., Riccardi) highlight pitfalls whereby portraits that do not fit canonical hair patterns are excluded, leading to perceived uniformity in imperial imagery. Broader scholarship emphasizes the omnipresence and representational roles of imperial portraits (Ando; Stewart; Fejfer; Lahusen). Digital and technological approaches to ancient art have increased (Schofield et al.; Pollini), but facial recognition has not been systematically applied to Roman imperial portraiture. The study builds on deep learning advances (CNNs; transfer learning) and pre-trained face recognition systems while acknowledging the challenges of applying methods designed for real faces to idealized, damaged, and variably photographed sculptures.

The authors propose a transfer learning approach using pre-trained deep convolutional neural networks adapted to imperial portrait images. Pipeline: (1) Face detection and alignment using MTCNN (Zhang et al., 2016a) to generate bounding boxes and landmarks; portraits are cropped, aligned, and resized to 256×256 pixels (aspect ratio retained). (2) Classification via Inception-ResNet-v1 (Szegedy et al., 2017) with transfer learning. Dataset: 673 images total—645 images across nine emperor classes (Augustus, Trajan, Hadrian, Antoninus Pius, Marcus Aurelius, Lucius Verus, Commodus, Septimius Severus, Caracalla) plus 28 non-emperor images as a distractor class. Images came from web databases (e.g., Arachne), book scans, archival holdings, and authors’ own photography. Data splits: training 351 images and validation 174 images using stratified sampling to preserve class proportions; an independent test set of 148 imperial portrait images (no non-emperors included due to scarcity) selected at random without stratification. Data augmentation (Albumentations) increases training variability and effective dataset size by approximately sixfold: horizontal flip (p=0.5), shift-scale-rotate (shift limit 0.02, scale limit 0.2, rotate ±10°, p=0.8), random crop (224×224, p=0.8), random brightness/contrast (limits 0.2, p=0.8), and multiplicative noise (multiplier 0.8–1.0 per channel, p=0.8). Training details: images normalized to [-1,1], cropped to 224×224 80% of the time, then resized to 160×160 (original model input size). Experiments: (1) Replace and retrain the last two dense layers of Inception-ResNet-v1 to adapt from 1000 original classes to 10 study classes and to focus on features relevant to sculpture rather than skin/texture. (2) Fine-tune the last 2D convolutional layer of the feature extractor (starting from Experiment 1) to learn more dataset-specific features for imperial portraits. Hyperparameters: dropout 0.4, 40 epochs, batch size 16 (train) and 8 (validation), cross-entropy loss with softmax, Adam optimizer, multi-step learning rate with initial 0.005 and steps at epochs 5 and 35. Evaluation metrics: per-class F1-scores and confusion matrices on validation and test sets; weighted average F1 by class image counts. Additionally, UMAP projections visualize 512-D embeddings (penultimate dense layer) for cluster separability.

- Transfer learning enabled recognition of Roman emperors from sculpted portraits despite lack of texture/color, varying lighting/pose, and damage/restorations. - Experiment 1 (dense layers retrained): weighted average F1 = 0.92 (validation, n=174) and 0.81 (test, n=148); overall accuracy reported 81.1%. - Experiment 2 (fine-tuned last conv layer): improved performance with weighted average F1 = 0.95 (validation) and 0.90 (test); overall accuracy 89.2%. - Class-level test F1 improvements from Exp 1 to Exp 2 include: Antoninus Pius 0.59→0.84; Lucius Verus 0.76→0.97; Septimius Severus 0.44→0.77; many others remained high (e.g., Augustus 0.93→0.91; Hadrian 0.85→0.90; Marcus Aurelius 0.82→0.92; Commodus 0.76→0.85; Caracalla 0.80→0.84; Trajan 0.86→0.88). - UMAP visualizations show well-formed, well-separated clusters for emperor classes in training and largely preserved clustering in validation; the non-emperor distractor class forms a dispersed, overlapping group reflecting its heterogeneity. - Misclassifications frequently involved emperors with historically assimilated portrait types (e.g., Marcus Aurelius vs. Commodus; Hadrian vs. Antoninus Pius) and images with damage or modern restorations; in many such cases the correct emperor appeared as the second-ranked prediction.

The findings demonstrate that modern face recognition models, adapted via transfer learning and limited fine-tuning, can effectively distinguish sculpted imperial portraits, offering a complementary tool to traditional hair-pattern-based identification. Although overall accuracy (81–89%) is below state-of-the-art human-face FR (>99%), the performance is notable given the limited and heterogeneous data and the distinct domain shift from living faces to idealized, damaged marble/bronze portraits. Misclassifications align with known art-historical phenomena—Bildnisangleichung (intentional assimilation of imperial images), dynastic stylistic convergence (Zeitstil), and the difficulty of distinguishing emperors from private individuals when fashion trends dominate—indicating that some errors may reflect genuine ambiguities rather than model shortcomings. Damage and restorations reduce distinctive features and can mislead the model, yet second-choice predictions often include the correct emperor, suggesting usable ranked outputs. Clustering analyses support that the model learns discriminative embeddings for emperor identities. The approach provides an empirical basis that avoids hard-coded criteria like coiffure patterns, potentially expanding recognized corpora and informing attribution debates.

Existing AI face recognition methods, when adapted through transfer learning, can identify Roman emperors from sculpted portraits with high F1-scores (0.95 validation; 0.90 test), providing a new empirical methodology alongside traditional art-historical approaches. The study contributes a proof-of-concept pipeline (detection/alignment with MTCNN, classification with fine-tuned Inception-ResNet-v1), curated datasets, and quantitative evaluations, demonstrating meaningful clustering of emperor classes. Future work will extend beyond recognizing among known emperors to discriminating imperial portraits from those of private individuals, expand datasets, refine domain-specific training, and further address challenges from damage, restorations, and stylistic assimilation.

- Limited dataset size and class imbalance across emperors constrain training from scratch and generalization; reliance on augmentations and transfer learning is necessary. - Domain shift from real human faces to idealized, monochrome sculptures lacking skin texture/color may limit feature transfer. - Variability in pose, illumination, photographic conditions, damage, and modern restorations introduces noise and can cause misclassification. - Historical assimilation of portrait types (Bildnisangleichung) and broader fashion trends (Zeitstil) may render some images genuinely ambiguous even for experts and algorithms. - Non-emperor distractor class is heterogeneous and was excluded from the test set due to scarcity, limiting evaluation of emperor vs non-emperor discrimination. - Reported accuracies (81–89%) are below state-of-the-art FR for living faces; results may not generalize beyond the studied emperors or to other media without further domain adaptation. - Dataset access is restricted due to copyright; full public reproducibility is limited though available on request via the corresponding author.