Metrologically interpretable machine learned features using latent spaces of Generative Deep Learning Models

• Applicant: Professor Dr.-Ing. Robert Schmitt
• Subject Area: Production Systems, Operations Management, Quality Management and Factory Planning
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 507911127

The extraction of relevant features is a key step in the process of building industrial machine vision systems that can measure quality characteristics of a process or product. However, traditional feature extraction methods require an a priori formulation of the features, that describe a quality characteristic, in a closed analytical form. However, these formulations are often not available or adequate for given the complexity and variance of data from industrial quality assurance applications. While Deep Learning can learn these features inherently, it requires massive amounts of annotated data and operates discriminative, hence, assigning nominal properties only. Moreover, machine learned features are opaque to any user, which hampers the ability to take remedial actions based on the decision of the Deep Learning model. The overall goal of “MIMICRI” is the development and establishment of a metrologically interpretable feature extraction method based on Generative Deep Learning to measure properties in industrial quality assurance applications, that cannot be assessed through traditional feature extraction methods. Within the cooperation with the Universidade Federal de Santa Catarina, WZL’s focus will be to develop the new feature extraction method in the domain of industrial Machine Vision. Generative Deep Learning methods learn the underlying distribution of a given dataset and have attracted a lot of attention due to their astonishing success in high-quality image synthesis and image editing. They can be trained in an unsupervised manner, requiring no costly annotations, and extract characteristic features of the given dataset in form of a structured and regularized latent space, that has shown a high degree of disentanglement. The disentangled representations can thus be interpreted and the association to quality characteristics established. For a metrological interpretation a quantitative assessment is necessary. For this purpose, distance measure in the latent space will be derived based on Riemannian manifolds to define scales in the latent space that enable the measurement of quality characteristics by means of reference points.

• DFG Programme: Research Grants
• International Connection: Brazil
• Partner Organisation: Coordenação de Aperfeiçoamento de Pessoal de Nível Superior; (CAPES); Setor bancario Norte
• Cooperation Partner: Professor Dr.-Ing. Armando Albertazzi Concalves Junior