Automatic Classification of Roman Flagons

Archaeologists devote much time to identify their finds, a tactile / visual excercise (feeling and looking) whereas the knowledge base consists of the researcher's memory supplemented with relevant publications. There are several problems in this area of archaeological craftsmanship. What, if the find is not recognized? Do we go through the books again or do we ask a colleague? Is the find classified as 'indeterminable', or do we undertake a publication of the new type?
The problem even increases when the find actually is recognized. How well does it fit ? Maybe in another publication a better resemblance might be discovered. But who will continue to leaf through the books, once a 'fair resemblance' has been detected ?
In order to investigate the possibilities to use automated tools an experiment was conducted to automate the identification of Roman flagons.
The typology
Roman flagons are a rather dull type of ceramics, lacking fancy decorations. The shapes of the flagons, however, vary considerably. Stuart (1962, 1977) distinguishes 12 types, based on differences in shape and size. Several hundreds of complete flagons form the basis ot the typology.
The experiment
From all drawn flagons about 40 variables were measured. On top of these data a system was built to enter user data. The STUART system (Simple Technical Utility for the Advance of Roman Typologies) compares these user data with those in the database, and determines the type of the user's flagon, based on a modified version of Gower's (1971) similarity measure. The original drawings and descriptive texts by Stuart are all integrated in the system.