CAA Conference has ended
Wednesday, March 28 • 9:00am - 1:15pm
5 - Automatically Recognizing the Legends of Ancient Roman Imperial Coins

Sign up or log in to save this to your schedule, view media, leave feedback and see who's attending!

The classification of coins plays a crucial role in numismatics. While the classification of modern coins has been researched successfully in the last few years, very few attempts have been taken to solve this problem for ancient coins. Due to their large inter-class variability and low intra-class variability owed to the manual production process and abrasions which occurred over the centuries, classifying them poses a far more challenging task to numismatics than the classification of modern coins. Thus, a system that automatically recommends the best matching classes for a given coin image could facilitate the work of numismatists and speed up the overall classification process. A robust classification system relies on discriminative features of the objects of interest. The legend, i.e., the textual inscription, of a coin is considered a highly discriminative feature in the manual classification process performed by numismatists today. Hence, a coin classification system could greatly benefit from automatic legend recognition. We propose a method for automatic legend recognition in 2D coin images that accounts for the challenging properties of ancient coins. Standard OCR algorithms rely on binarization to separate text from background. However, binarization is a very error-prone process for ancient coin legends due to the complex shading effects on the highly specular coin surfaces. Therefore, the proposed method employs local features for the description of characters which are used for the training of a Support Vector Machine (SVM) for every character class. Legend recognition is then performed by shifting a sliding window across a coin image and computing the local descriptor for every location. These descriptors are then tested with every SVM to produce a probability map for each character and location. In combination with a lexicon containing the set of possible words, the best matching ones are computed using Pictorial Structures. The performance of our proposed method is evaluated on a set of ancient Roman coins.


Martin Kampel

Vienna University of Technology

Albert Kavelar

Vienna University of Technology

Sebastian Zambanini

Vienna University of Technology

Wednesday March 28, 2012 9:00am - 1:15pm BST
Building 65, 1177 Streamed into room 1095

Attendees (0)