Newsletter October 2022

Our newsletter from October 2022

Carton content inspection using deep learning

The end customer in this project was a leading international manufacturer of technology-oriented products for the eye. Together with our customer KOCH Pac-Systeme GmbH, we integrated an optical inspection system with deep learning algorithms.
At the end of a secondary packaging line, the individual blisters grouped in a folding box were to be checked for the presence of the required number of blisters and a leaflet. The variants to be checked were / 4 blisters without leaflet / 6 blisters without leaflet / 6 blisters with leaflet.

The challenge

As the blisters have room for maneuver within the outer packaging, there are strong fluctuations in position, which cause a high degree of variability in appearance. High position and size tolerances must therefore be taken into account. These are unfavorable conditions for a rule-based test approach, which experience has shown to lead to increased pseudo errors.

With this in mind, the deep learning approach was given preference over the rule-based approach for this solution.

Implementation

Three separate test programs were used on the camera to differentiate between the three variants. Each of the three inspection programs accesses a previously trained VIDI model.
When teaching the “blister patterns” into the DeepLearning model, the expectations in the model expand from time to time in terms of what can be expected and is good. When searching for the trained “blister pattern”, these are then also recognized in the event of strong stock fluctuations.
The two features “blister” and “package insert” were trained in the model. The number of both features and the positions of the features found are determined using additional functions and displayed in the camera image using graphical regions. The determined number of both features is compared with a fixed specification entered in the inspection program. The result of this comparison is then used for graphical OK / NOK signaling and for deciding the error image memory. The number of characteristics actually found is used as the result of the test for the control system.
A WEB view has been integrated for visualization in the HMI of the system.

This shows the current camera image with a surrounding colored frame to indicate the OK/NOK status of the respective test as well as the features found in graphical areas. Next to the camera image is the test program name, the display of the camera’s operating status, the menu for language switching and the statistics on the evaluated tests, including the reset button.

Result

A satisfied customer!

If you have specific questions, please feel free to contact us(thomas.hahn@i-mation.de)!