Metrologically interpretable feature extraction for industrial machine vision using generative deep learning

Deep Learning (DL) is leveraged in a growing number of industrial applications. One strength is the data-driven ability to extract characteristic features from complex inputs in form of a latent vector without the need for closed formulation or derivation from a priori known quantities. This work proposes a framework based on generative DL methods to interpret these latent vectors as metrological quantities. The approach is explored in the machine vision domain by implementing a model utilising style-based adversarial latent autoencoders, principal component analysis, and logistic regression. It is successfully evaluated on an industrial image set of aluminium die casting surfaces.