Automated Analysis of Ancient Coins

• Automated analysis of ancient coins is an interdisciplinary field that uses computer vision, machine learning, and statistical frameworks to identify, classify, and grade numismatic artifacts.
• Recent approaches leverage advanced object detection models like CLIP and Vision Transformers to achieve robust motif recognition, unsupervised die-linkage, and semantic legend extraction.
• Integrated pipelines combine multi-modal data and tailored feature extraction methods to enhance grading accuracy and clustering performance in coin analysis.

Automated analysis of ancient coins is the application of computer vision, machine learning, and statistical frameworks to the large-scale identification, classification, semantic interpretation, grading, and die-linkage of numismatic artifacts. Recent work leverages object detection architectures such as CLIP, Vision Transformers, hierarchical convolutional networks, as well as specialized workflows in die analysis, symbol detection, and legend reading to address the complexity and diversity of ancient coinage. Automated pipelines are now capable of matching or exceeding expert annotation in motif recognition, die-grouping, and quality grading, with accelerating progress in multi-modal and weakly-supervised methods.

1. Object Detection Frameworks for Semantic Coin Analysis

Recent research has established Contrastive Language-Image Pre-training (CLIP) models and Vision Transformers (ViT) as state-of-the-art for object detection and semantic motif identification within ancient coin imagery. For CLIP-based frameworks, images are normalized (e.g., resized to 224×224 and pixel-scaled), then coin surfaces are divided into candidate patches via sliding-window or selective search. Feature extraction leverages CLIP’s image encoder $f_{\theta_I}(x)$ and text encoder $f_{\theta_T}(t)$ to produce unit-normalized d-dimensional vectors; similarity is computed via dot-product, $s(u, v) = u \cdot v$, for both image and text queries (Cabral et al., 2024).

Search and classification tasks are formalized: coins are ranked by similarity scores to reference motifs, either as patches against a target image or as semantic concept queries (e.g., “swastika stamped on aged coinage”). For classification, p-value calibration is applied to similarity scores against a null distribution from coins not featuring the inquiry object, producing robust presence/absence decisions even in degraded datasets. Larger CLIP models (ViT-H-14, ViT-L-14) reliably detect complex motifs and outperform traditional local descriptors (SIFT, ORB) except in the case of simple geometric patterns where classical algorithms remain competitive.

Vision Transformer approaches—partitioning input images into fixed-size patches—use transformer encoder layers with multi-head self-attention to capture long-range spatial dependencies and motif configurations (Reid et al., 14 Jan 2026). In motif recognition, ViT-Large models fine-tuned on coin datasets achieve higher accuracy, precision, and F1 scores than convolutional baselines (average accuracy ≈ 0.80 for ViT versus ≈ 0.75 for CNN). ViT’s global contextual modeling enables the recognition of subtle and occluded symbols on worn coins.

2. Symbol-Based and Semantic Element Classification

Bag-of-Visual-Words (BoVW) pipelines, enhanced by rotation-invariant spatial tiling, remain influential for coarse-grained symbol-based classification. Dense local descriptors (typically SIFT) are clustered by k-means to form vocabularies; features are encoded by histograms in subdivided image regions. Circular tiling ensures rotation invariance critical for unconstrained coin orientations, yielding superior accuracy to rectangular and log-polar tiling for small vocabularies (e.g., 72% vs. 65%), and convergence across tilings at large vocabularies (≈75%) (Anwar et al., 2013).

Semantic motif learning has shifted towards modeling the “content” rather than direct image matching; weakly-supervised CNNs are trained using labels extracted from multimodal metadata (e.g., auction descriptions). Each concept (cornucopia, patera, eagle, horse, shield) is mapped to a keyword set via language normalization (multi-lingual translation and stemming) and images are labeled through text-mining (Cooper et al., 2019). AlexNet-style CNNs attain per-concept test accuracies up to 0.84, with strong agreement between occlusion-based heatmaps and expert expectations of motif localization. Vision Transformers further enhance global context modeling for multi-label motif detection under noisy weak supervision, outperforming CNNs on the same multi-modal corpus (Reid et al., 14 Jan 2026).

3. Automated Die-Linkage and Unsupervised Clustering

Automated die studies—critical to quantifying ancient monetary production—are now feasible by combining fast feature extraction, robust matching, and unsupervised clustering. State-of-the-art pipelines employ deep local descriptors (XFeat, D2-Net), outlier rejection (MAGSAC++), and graph-based clustering algorithms (Label Propagation with silhouette-based hyperparameter selection) to partition large corpora without ground-truth labels (Cornet et al., 2024, Harris et al., 2023, Heinecke et al., 2021). SSIM-based global image metrics offer robust alternatives to keypoint-only methods, yielding near-perfect discrimination and clustering in coin die-link identification (Labedan et al., 3 Feb 2025, Labedan et al., 7 Nov 2025).

Bayesian microclustering models operate directly on pairwise dissimilarity matrices derived from dense keypoint matches, weighted descriptor distances, and global structure similarity (SSIM). Cluster assignments are inferred by MCMC with variation-of-information loss minimization, achieving NMI and ARI values over 0.9 on historical corpora. For 3D scanned Celtic and Greek coins, registration combines point-to-plane ICP and random-search alignment; pairwise pattern comparison by logistic regression over distance histograms yields die-link probabilities, which are thresholded in a graph to extract die clusters with ARI up to 0.86 (Horache et al., 2020).

SSIM-based scoring systems are scalable for large die-link studies; metrics and explicit thresholding define likely links, while cluster evaluation achieves F1-scores > 0.92 and ARI/NMI ≈ 1.0 over curated datasets. Hybrid human–machine interfaces now facilitate interactive verification and visualization (overlay, fade, loupe) of candidate links within coin typologies (Labedan et al., 7 Nov 2025).

4. End-to-End Hierarchical Coin Classification and Landmark Discovery

Deep convolutional models—fine-tuned on large, annotated coin datasets—support fine-grained recognition of catalog labels (RIC), emperor identities, and motif types via hierarchical classification architectures (Kim et al., 2015, Anwar et al., 2019). Over 9,000 coin images spanning 314 RIC labels and 96 emperors are processed by AlexNet networks with customized output layers; two-step pipeline predictions combine obverse-side identity with reverse-side motif. Classification accuracy reaches 76.2% (hierarchy+CNN), with substantial gains over SVM baselines.

Optimization-driven landmark discovery uses sparsity-regularized transparency masks over image patches to identify the minimal spatial regions essential for class discrimination. For image $I$, optimization is:

$$\min_{x\in[0,1]^K} \ell_c(f_I(x)) + \lambda \|x\|_1,\quad \text{subject to}\quad p(c|f_I(x)) \geq p(c|I) - \epsilon$$

Heatmaps correlate with expert motif descriptions, confirming the model's ability to locate attested features such as shields, columns, or owls. Feature attention and compact bilinear pooling (CoinNet) further enhance classification, yielding >98% accuracy on 18,000 Roman Republican coins and strong robustness to erosion, illumination variation, and style diversity (Anwar et al., 2019).

5. Automated Grading and Quality Assessment

Automated grading tasks, such as quantifying Mint State grades, require robust handling of class imbalance and limited labeled data. Hybrid systems integrating hand-engineered coin-expert features (edge-wear statistics, wedge-based spatial analysis, color clustering, luster modeling) with neural architectures outperform end-to-end CNNs, particularly in datasets of <2,000 coins (Dogra et al., 4 Dec 2025). Feature-driven artificial neural networks (ANN) achieve exact-grade matching at 86% and top-3 tolerance accuracy at 98%, whereas CNNs trained on small, skewed datasets tend to collapse predictions to the majority class. Domain knowledge encoded in engineered features is critical for rare or niche ancient coin issues.

Generalization across coin types, metals, and eras is achieved by extending color clustering to oxidation profiles (verdigris, sulfide layers), integrating multispectral imaging, and incorporating 3D relief modeling for surface wear and edge rounding. Advanced imbalance correction (SMOTE, focal loss, weighted cross-entropy) further improves minor-class sensitivity.

6. Legend Recognition and Textual Feature Extraction

Recognition of legends (inscriptions) on ancient coins challenges classical OCR due to low metal-background contrast and non-linear arrangement of engraved text. Object recognition-based pipelines employing SIFT descriptors with fixed horizontal orientation and “half-spectrum” gradient merging yield far superior rates on cropped legend images than commercial OCR engines (Kavelar et al., 2013). Each character location is scored by SVM classification, with word-level recognition implemented as a pictorial structures model minimizing an energy function balancing per-character scores and regularity of spacing. Recognition accuracy for character-level “Coin” legends reaches 75.6% versus near-zero for ABBYY OCR; the word-level pipeline enables integration of legend parsing into automated coin-issue cataloging.

7. Limitations, Challenges, and Future Directions

Performance is modulated by dataset signal quality, imbalanced samples, severe wear, lighting-induced variability, and the diversity of engraving styles. Single-modality pipelines benefit from integrating multi-modal supervision, multi-label architectures, and deep spatial attention or transformer modeling for enhanced robustness. Domain adaptation, semi-supervised learning, and feature fusion across relief, color, legend, and metal composition will likely yield further advances. Current research advocates extensions to fully multi-modal CLIP or Transformer-based systems, weakly supervised localization, and learned global representations tailored for forensic provenance (e.g., forgery detection) and automated cultural heritage mapping.

All code, datasets, and metrics (CLIP-based similarity, SSIM-based clustering, Bayesian microclustering, Procrustes alignment) are released or documented in the associated preprints, facilitating reproducible research and collaborative benchmarking for digital numismatics.

---

References:

• CLIP-based object detection and calibration in digital numismatics (Cabral et al., 2024)
• Vision Transformer motif analysis (Reid et al., 14 Jan 2026)
• BoVW symbol-based classification (Anwar et al., 2013)
• Semantic content learning from multi-modal metadata (Cooper et al., 2019)
• Deep feature fusion for Republican coins (Anwar et al., 2019)
• Unsupervised die analysis (Cornet et al., 2024, Heinecke et al., 2021, Harris et al., 2023)
• SSIM-based die link identification (Labedan et al., 3 Feb 2025, Labedan et al., 7 Nov 2025)
• CNN landmark discovery (Kim et al., 2015)
• Feature-driven grading (Dogra et al., 4 Dec 2025)
• Legend object-recognition OCR (Kavelar et al., 2013)
• 3D scan registration for die matching (Horache et al., 2020)
• Semi-supervised graph transduction framework (Aslan et al., 2018)