Analysing information content in images of paper straws using image processing, machine learning and deep learning

Abstract: Continuous production control is important for the production using Tetra Pak equipment, as problems that are found early in the chain of production can minimize the need for waste greatly. In production of paper straws, a possible production control could be to photograph straws during production, analyse these and predict how good the straw is, using the image. If the predicted performance of the straws are steadily decreasing, production can be stopped and control parameters tuned in order to get back to peak performance. The task described above is built on the assumption that there is information about how good a paper straw is, in a picture. This thesis has the intention of showing whether images alone can be used for production control. Methods such as image processing and the classical machine learning algorithm Random forest regressor will be compared to, and used together with, a deep learning option, Convolutional Autoencoders. The results show that an approximate maximum of 10% of the variations in open- ing force can be explained by the information in the photographs. Best performing was the autoencoder, quite closely followed by image processing at just under 10% of variations explained. The suggested improvements are many, but can be divided into two categories: further improvements of the models and information extraction, and improvements to increase predictability overall. In the first category, the inclusions of more sam- ples, trying out other models beside Random forest regressor, and examining other network architectures for the autoencoder, are some of the advised options. In the second category, parameters that are deemed important but that cannot be seen in figures are recommended to be added to the data set. These parameters may include details on the glue used to seal the paper straw, or some sort of measure on how sharp the knifes that cut the straw are, to mention two examples.