Comparison between traditional vision and AI-based inspection systems for crimping quality checking in aseptic pharmaceutical manufacturing

Modern manufacturing processes are characterized by the request of amazingly high production ratios without impairing product quality. Until a few years ago it was sufficient to improve electro-mechanical systems, while adding just a few basic sensors. This is not enough anymore: computer vision has improved quality standards of existent tasks, but has also made possible to perform a multiplicity of new activities. In the past decades, vision systems were synonyms of "bulky" and "expensive", requiring customized architectures and long development time. Lighting conditions, geometry of the machine and environment control were key factors in obtaining an effective and robust vision system, deeply impacting in machine architectural choices. Neural networks and deep learning have helped to overcome these constraints. This thesis is focused on monitoring the quality of vials crimping after they have been filled with pharmaceutical liquid. The crimping process is carried out by rotating heads which grasp the bottle and accompany it towards a specifically designed profile that shapes the aluminum sheet around the bottle cap. Reliability of the precedent process is a key factor in preserving chemical and biological characteristics of the pharmaceutical liquid along time, with strict requirements to be met by law. This work presents a comparison between traditional computer vision techniques and modern deep learning ones, investigating also different sensitivity levels achieved using specific industrial architectures.