For a wholesaler of coins we developed a machine vision solution to determine the position and rotation of various collector coins for subsequent special printing. The position must be detected with an accuracy of 5/100 mm, otherwise the coins would be printed incorrectly and thus become unusable.

The coins are presented to the machine vision system, consisting of four high-resolution 20MP GigE monochrome cameras with corresponding lenses, on a tray. The system uses a matching process to recognise the coins and determine their position in space. A combination of LED backlights and white diffuse LED line lights is used to perfectly illuminate the scene. The processing is done on a compact image station with 4 PoE GigE interfaces and a monitor. The system is installed in a closed box into which different coin trays are inserted via a drawer.

On the software side, the system consists of two separate components developed by phil-vision, which are based on MVTec's Halcon image processing library:

coinTeach
Application where a model of the inserted coin is trained and stored.

coinFinder
Application with the help of which the coins on the tray are found based on the previously generated model and where their position and rotation are determined.

In a closed structure, four cameras look from above at a defined image field. A black tray with several fixed markers, made by the customer, is pushed into the system via a drawer. The system is currently designed for two different tray sizes. The large tray consists of four areas (one area per camera) with one predefined area used for teaching. Here, the individual coin to be taught is placed and the model is generated. The entire tray area is then used to find the coins with their respective position and rotation.

If the tray and the coin to be trained are correctly placed under the system, a matching model of the coin is taught using the coinTeach application. Instead of the camera image, a previously generated binary print image can optionally be used to create the model.

After training, the coinFinder is started and the previously saved model of the coin to be found is loaded. Using the search function, the coins present on the tray are now found and their x- and y-position as well as the rotation are determined in comparison to the trained model. The results for each coin are displayed on the user interface and saved as a CSV file. The information from the CSV file is then used in the printer to print the coins accurately.