IFAG, Symate

The demonstrator provides an optical inspection solution working on the same or similar hardware and software environment. The demonstrator consists of several phases:

• (Pseudo-)anomaly detection with a pre-trained neural network (NN) for detecting deviations
• A pool of labelled images is generated for a supervised approach to build a classifier model by checking those deviations
• Deploying this model to unseen productive images and analysing the predictions which provide information about the target performance

Beyond state-of-the-art developments and impacts

The approach is innovative in making a two-step process by firstly using some unsupervised trained anomaly detection and secondly analysing (and with this step labelling) deviating data (e.g., images). With these data, a supervised training for generating a model is conducted, which can predict more information. This can be a blueprint on dealing with a lack of labelled data when starting an AI project.