TL;DR
This paper challenges the long-standing practice of ignoring color in ancient coin analysis, demonstrating that color features, when properly represented, can significantly improve denomination classification. The authors introduce two novel color representations and a new dataset, showing surprisingly high accuracy that suggests color is a crucial, yet overlooked, modality for this field.
What Problem Does This Paper Address?
The automatic analysis of ancient coins, particularly for identifying their type and origin, is a challenging area in computer vision and machine learning. While significant research has focused on visual features like text and imagery, a critical aspect has been consistently overlooked: color. Historically, coin images have been treated as grayscale, with color information being discarded. This is often justified by the idea that a coin's type is determined by its iconography and inscriptions, not its color. However, this paper argues that this is a missed opportunity, especially for initial coarse-grained classification tasks like determining a coin's denomination. The sheer number of ancient coin types (tens of thousands for Roman Imperial coins alone) makes direct classification extremely difficult. The authors propose that color can serve as a powerful initial filter, narrowing down the possibilities by leveraging the fact that different denominations were often made from distinct materials with characteristic colors.
Methods
The paper introduces two novel color-based representations for ancient coins:
- Hue Histogram: This method leverages the Hue channel from the Hue-Saturation-Value (HSV) color space. Hue captures the pure color information, and the authors argue it's less affected by lighting and surface conditions than saturation or value. They create an 8-bit bin histogram of hue values, filtering pixels with low saturation and value to reduce noise. This histogram is then normalized for scale and resolution invariance.
- Histogram of Colour Words: Inspired by the "Bag-of-Visual-Words" approach in general computer vision, this method treats colors as "words." Pixels from a training set are clustered using k-means to define a dictionary of "color words." An image is then represented as a histogram of how many pixels fall into each color word cluster. The optimal number of clusters (k) is determined using the minimum description length principle.
For classification, the authors employ a random forest classifier, choosing it for its interpretability and suitability for this task, rather than more complex deep learning methods. They also explore score-level fusion of the two proposed representations.
Datasets
The authors introduce a new dataset specifically curated for denomination-based categorization of ancient coins. This dataset comprises 400 coin images, with 100 exemplars for each of the four denominations studied: ases, dupondii, sestertii, and denarii. A key feature of this dataset is that all coins are imaged against a uniform background to minimize confounding factors like clutter and variable backgrounds, allowing for a focused evaluation of color features.
Key Insights and Ideas
- Color is a forgotten but valuable feature: The most significant insight is that color, despite being readily available and often indicative of material composition, has been systematically ignored in ancient coin analysis.
- Color as an initial coarse filter: The paper strongly advocates for using color as a first-pass classification step to drastically reduce the search space for coin identification, especially given the vast number of coin types.
- Material-color correlation: Different Roman coin denominations were often made from distinct metals (e.g., copper for ases, silver for denarii), leading to inherent color differences that can be exploited.
- Color representation matters: The effectiveness of color features is highly dependent on how they are represented, motivating the development of novel, domain-specific representations.
- Hue is robust: Focusing on the Hue channel proves effective, as it captures intrinsic color information less susceptible to imaging variations.
Key Results
- Hue Histogram Performance: The hue-histogram-based representation achieved surprisingly high accuracy, with denarii and ases classification exceeding 90%. Even for dupondii and sestertii, which have similar compositions, performance was better than expected.
- Color Words Improvement: The "color words" representation showed a notable improvement in differentiating between dupondii and sestertii, halving the error rate compared to the hue histogram.
- Fusion Benefits: Combining the two representations through score-level fusion further boosted performance, achieving the highest accuracy across all denominations, with ases and denarii reaching 96%.
- Unexpectedly High Information Content: The results convincingly demonstrated that color carries significantly more relevant information for denomination classification than the authors initially anticipated.
Limitations and Future Work
- Classifier Choice: The authors acknowledge that they used a simple classifier (random forest) to focus on the color modality and representations. They suggest that more complex machine learning approaches, such as deep learning, could be explored for further improvements.
- Hue Histogram Refinement: While effective, the hue histogram could potentially be improved by incorporating the semantic relationship between neighboring bins, perhaps through techniques like Parzen window estimation to reduce noise.
- Feature-Level Fusion: The paper explored score-level fusion. Future work could investigate feature-level fusion, which is considered a more principled approach and might yield better practical results given the low dimensionality of the problem.
- Broader Application: The findings suggest that color analysis could be beneficial in other areas of computer vision where material properties influence visual appearance.
Summary updated