Banknotes deteriorate over their lifetime, becoming dirty, creased, faded or torn from being handled by many people and machines. Central banks inspect paper currency based on "fitness" levels, withdrawing damaged banknotes from circulation and replacing them with newly printed ones. This process ensures currency stays structurally sound and machine-readable by ATMs.
Images of an unfit Euro 5 banknote with 3D information and surface description. Three images from left to right: Color image, Depth variation in millimeters, and Curvature of depth image. On the curvature image, the color blue corresponds to ridge-like structures, the color red to valley-like structures, and green tones are uncreased or slightly creased regions. Image Credit: Chromasens GmbH
Automated 2D optical sorting machines have long been used to evaluate the fitness of banknotes for recirculation by comparing each bill against strict criteria. While 2D systems can identify non-dimensional faults, such as stains, fading or tears, they lack the precision to detect dimensional classes of defects, specifically creasing, also known as wrinkling, crumbling or random folds. Shading effects from creases are frequently indistinguishable by 2D sensors from printed patterns and stains.
Detecting Creases With 3D Cameras
Creasing is one of the main reasons why banknotes become unfit for recirculation, as it may obscure serial numbers for machines to read. For reliable discrimination between creasing and other types of defects, 3D imaging of the surface is required to reduce false fit or false unfit determinations.
To tackle this challenge, the Digital Safety & Security Department at the AIT Austrian Institute of Technology GmbH (Vienna, Austria) has developed a 3D approach for automatic banknote fitness inspection. Due to the industrial setting of cash recycling workflows and the need for line-scan stereo vision, the team selected a Chromasens 3DPIXA line-scan camera combined with highly efficient algorithms for inline assessment. The Chromasens 3DPIXA camera is capable of lateral pixel resolutions of 15 μm (10 μm x 10 μm pixel size), a line rate of 20 kHz and a nominal height resolution of 3 μm. With its unique mix of line-scan technology and 3D stereo calculation on graphics cards, the 3DPIXA simultaneously delivers 3D data and 2D color images.
Testing consisted of a data set of 30 examples of Euro 5 banknotes which were labeled by national bank experts. Both sides of the banknotes were acquired by the camera. The system separated texture-like elements on the surface from the 3D properties of deformed banknotes by analyzing concepts borrowed from differential geometry, particularly extracted local curvature features. Analysis showed wrinkle structures and elevations in different colors. Blue corresponded to ridge-like structures and the color red to valley-like structures, with green tones signifying uncreased or slightly creased regions.
AIT's experimental acquisition setup identified the requirements with respect to 3D camera resolution and the discriminative power of algorithms. Further testing of different types of paper currency should validate the approach with the introduction of machine learning to obtain decision rules and thresholds for banknote fitness determination.