Deep Learning for anomaly detection in vials

Details

Supervisors

Lee, John Aldo ; Matthieu Duvinage

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

Recent years have seen a shift in focus for scientific development with the surge of the COVID-19 pandemic, where speed became a requisite for vaccine development and production, especially for companies such as Glaxo-Smith Kline. During vaccine production, separation of faulty vials carries a danger of becoming a bottleneck, as current measures heavily depend on manual interruption. At this point, Deep Learning techniques come into play to streamline the process by classifying vials as clean or defective. Unfortunately, the methods used insofar either rely on manual labelling for both classes, or manual labelling due to low accuracy. The paper combines multiple anomaly detection and Deep Learning methods to combat the presented issues. Unsupervised Learning, through autoencoders with a U-Net architecture, reduces the labelling cost significantly. Furthermore, the use of specially curated image deformation and preprocessing techniques pave the way for much desirable accuracy and false-positive rates to be achieved.