IBISA (Image-Based Identification/Search for Archaeology) was a research project supported by the French CNRS.  The resulting software tool manages databases of digital images of archaeological objects, and allows the user to perform searches by examples.

The objects are only required to be quasi flat (approximately two-dimensional) and produced from matrices via some striking / stamping / casting process.  The original matrices are generally lost now, but many objects with their prints can still be found, with many similarities among them.

The IBISA system was originally designed to work with ancient (greek, roman) coins.  In this numismatic context, the matrix is then a die.

By taking advantage of the relations (die links, style similarities) among large finds (coin hoards), one could determine the original production process, chronology, geography, etc.  The difficulty is to handle a large number of objects, very similar at first sight for a non-specialist, and to compare them.  This task is time-consuming, quite exhausting, and thus error-prone.

IBISA was designed to help the user decide, from their images, if two objects are either the same, come from the same matrix, share resemblance in style, or are completely different.

It uses computer vision methods to make this decision while getting rid of the viewing conditions when searching for similarities in the databases.  The method works on the image in its globality, and does not rely on specific key-points.

To avoid any colorimetric issue, only the luminance (intensity) of the pixels is considered.  This enables also the comparison of coins struck using the same dies but with different metals (such as denarii and aurei of the Roman Empire).

To get rid of the background context, a segmentation method based on active contours first extracts the useful part of each image.

To cancel any translation, rotation, or zoom inherent to the photography, the system uses a registration method based on the Fourier-Mellin transform, which can estimate the parameters of such rigid transforms.  These transforms can then be inverted, to get the best match for the images.

For now, the similarity measure is given by the cosine similarity among the target image and the registered version of each candidate in the database.

The last viewing condition problem is the light source position. Indeed, the shading can change the cosine similarity by about twenty percents.

The solution proposed here is to consider a three-dimensional model, the elevation map replacing the luminance information.

The present work gives an overview of the original IBISA system, characterizes the errors in the estimation of the rigid transform by the registration method, as well as the impact of the light source direction on this estimation and on the similarity.  Then the extension of the system to the three-dimensional model of the coins is presented, yielding a more reliable estimation and similarity.

The problem is now to be able to get the model from images.  A simple yet efficient solution to this shape-from-shading problem is also proposed.