3D Image Processing Techniques Enable Study of Ancient Fingerprints

By Lina Sorg

Ancient fingerprints are present in various antique artifacts, especially pottery pieces made of ceramic and clay. These preserved prints can survive for millennia and are often visible to the naked eye. Historians study them to identify the creator, comprehend the creation process, and understand various cultural and artisanal aspects of the society in which the pottery pieces originated, including division of labor (between men and women or young and old citizens, for example). However, extracting and interpreting these latent fingerprints poses a significant challenge for the imaging sciences community because they are often obscured or incomplete.

The current “golden standard” of print extraction is based on two-dimensional (2D) imaging, and archeologists typically employ acquisition methods to perform the necessary measurements and capture the prints. They investigate the print under a microscope, zoom in on particular areas of interest, and take a photo using a standard 2D camera. Portable cameras are therefore useful archeological tools, as they are user-friendly, lightweight, relatively cheap, and do not require specific operator training. Unfortunately, they are limited in their success, as material aging often reduces the prints’ visibility in traditional photographs and irregularly shaped prints cause perspective deformations in the captured data from a single camera.

Three-dimensional (3D) imaging techniques provide an attractive alternative to more traditional approaches. The main features of interest in 3D image-based studies are the fingerprint “points,” such as ending points, isolation points, and bifurcation points. Identifying and detecting these points is useful for the detailed review of prints in ancient pottery. During a minisymposium presentation at the 2020 SIAM Conference on Imaging Science, which took place virtually last week, Dzemila Sero of Centrum Wiskunde & Informatica (CWI) overviewed 3D imaging techniques that decode ancient fingerprints and enrich the resulting cultural heritage.

Several methods currently exist to acquire 3D prints. One such method is the use of structured light projection—which processes a known pattern and projects it onto a scene—to generate the necessary lines and produce a 3D shape. Multi-vision systems are another alternative. These set-ups place the object of interest in the field of view of multiple cameras, which shoot simultaneously. Each camera is programmed with unique settings to capture different light effects and other specific conditions. Multi-camera systems retrieve the desired image by overlapping the multiple views; matching the points between images eventually defines the 3D surface and yields a 3D surface model of the fingerprint. “This is a cool example of how a camera system like this works on the small details,” Sero said.

Each acquisition system has advantages and disadvantages. For example, complex systems are hard to transport and difficult to assemble in places where small details are necessary. They also cannot “see” through objects or detect fingerprints in hollow regions of clay artifacts, and lack the physical ability to capture prints in these hollow regions and other hard-to-reach places.

The FleX-ray Lab at CWI offers a potential solution. The lab possesses a state-of-the-art computed tomography (CT) scanner that can look inside 3D objects during the scanning process. This scanner consists of an X-ray source, a sample stage on which the object is placed, and a flat-panel detector, all of which yield a set of projections that reconstruct the object in question. For her study, Sero focused on a collection of statues from the Rijksmuseum in Amsterdam that are made of fired clay and survived centuries in a nearly unchanged state. Because they are clay-based, many fingerprints are present on the statues’ surfaces. “It’s really interesting for us to study these fingerprints to answer some questions related to the identity of the maker,” Sero said, excited by the prospect of potentially tracing a signature and determining whether the same craftsperson was responsible for multiple artifacts.

Researchers who employ scanning technology must scan each artifact in one take, with the machine’s parameters set correctly and the supports already in place. Many artifacts are not convenient shapes, and the supports must be made of strong materials that will keep the pieces stable during the scanning process to avoid blurred images. Due to the required level of precision, scientists use phantoms and FDK reconstructions to train CT machines and fine-tune their parameters beforehand; Sero followed these protocols before scanning the Rijksmuseum statues. Because every article is different, one cannot exactly apply one statue’s parameter settings to another.

In a forensic scenario, experts lift fingerprints from objects with film, molds, or dust to acquire a 2D version and extract the offending print. One can only assess a probability of identification if the features (or points) from two prints match; otherwise, a non-match leads to a non-identity. However, most ancient artifacts—including the Rijksmuseum statues—are too fragile to survive such invasive approaches. The latent fingerprints on these types of relics are almost always fermented, partial, and/or distorted, and thus rarely present in their entirety.

In the biometrics field, 2D feature extraction methods depend heavily on enhancement algorithms—such as orientation and frequency maps—which yield thin, simplified versions of the initial fingerprints. Scientists examine these simplified prints to calculate the points of interest. Another possible approach is the use of 2D feature extraction methods for 3D models via 3D-to-2D unwrapping algorithms. Researchers utilize parametric or nonparametric methods, as well as parametric shapes, to identify potential matches among fingerprints based on the fitted points.

Because the unwrapping process sometimes introduces distortions, Sero employs curvature maps to generate the lines. Recent advances also work to define 3D features using end-to-end neural networks. “These are really novel because there aren’t that many papers dealing with 3D fingerprint detection,” Sero said. All methods to mitigate unwrapping distortion rely on a 3D point cloud.

The final step in the 3D imaging process is feature mapping. Identifying matches is not as simple as one might think; in addition to being incomplete, fingerprints are often oriented in different ways.

Ultimately, Sero attempts to detect fingerprints on CT-reconstructed statues from the Rijksmuesum to define the number of corresponding features across the partial fingerprints, identify the best feature for the purposes of matching, and draw conclusions about the creators. Many steps throughout the process begin with strong data acquisition, which requires proper training. “Having good data influences every single step afterwards,” Sero said.

Because most 2D and 3D methods for fingerprint feature extraction are tailored to a single object, Sero concluded by acknowledging that expanding the pool of objects and resulting data-driven shape analysis might be a logical future step for this type of image processing.

Lina Sorg is the managing editor of SIAM News.