Detecting Overlapping Pharmaceutical Infusion Bags using Convolutional Neural Network

When producing pharmaceutical infusion bags, manual inspection of the production of the bags is very important. We set out to see if convolutional neural networks (CNNs) can be used to make one of these inspection steps easier. The task was to detect where infusion bags placed on a tray overlapped. The problem with detecting this overlap is that the infusion bags are made of transparent plastic and filled with transparent fluid, so it is hard to see the bag placements. Several different setups of CNNs were explored, as well as one image processing technique. All the neural networks tested were able to find the infusion bags. The U-Net CNN for the whole image seemed to be especially well suited this specific inspection problem, as this made it possible to make a direct input-to-output neural network with the least limitations. This paper describes the various methods and setups used, including their advantages and limitations. How to implement the selected setup in the actual inspection process and some further work we would like to do is also mentioned.