Answers are grounded in this paper and cite their sources.
Assesses automated coin grading for diverse collections, indicating improved accuracy and efficiency.
The assessment and grading of coins based on preservation is a fundamental task in numismatics and automated museum technologies. Traditional grading relies on visual inspection, which is an inherently subjective measurement approach due to the dynamic and variable nature of coin conditions within the Sheldon scale. While prior research has explored automated grading, existing studies remain limited in functionality, often focusing on specific coin types or historical periods, which restricts their practical feasibility. To address these gaps, we propose a computer vision-based approach for automated coin grading, combining a dual deep-learning model with a grade-shift algorithm. The approach integrates modified ResNet-D and Normalized Free Residual Network (NFResNet) architectures to leverage their complementary strengths, overcoming the limitations of single-model approaches. The grade-shift algorithm determines the final grade by evaluating the probability distributions from both models, computing confidence distances, and incorporating the domain knowledge enabling robust classification even for diverse coin datasets. We evaluate the proposed solution on the dataset provided by a grading company. The dataset encompasses multiple U.S. coin types and Mint years across a broad historical period, introducing substantial intra-dataset diversity compared to prior efforts. Experiments demonstrate classification into three major grade categories: Circulated (CR), Almost Uncirculated (AU), and Mint State (MS) , achieving accuracy 0.894 and F1-score 0.851. Gradient Weighted Class Activation Mapping (Grad-CAM) visualizations validate that the model learns numismatically meaningful features, providing interpretable evidence for its decision-making process. This work advances automated grading by establishing a scalable and objective framework for heterogeneous coin collections, while laying a foundation for future research in coins pattern recognition for numismatics and museum technologies.
No takes yet. Share an insight, caveat, or question.
Getahun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: