This thesis presents a concept for improving the data annotation job for computer vision applications in manufacturing. The focus is on process optimization, cost reduction and resource conservation. The proposed concept is based on the masked autoencoder concept as a self-supervised learning approach. The goal is to provide a scalable, widely applicable solution that can be integrated in the machine learning operations lifecycle. Three specific manufacturing datasets are used for concept validation and the result evaluation is calculated by the loss function and image congruence. The thesis concludes that the proposed data annotation concept saves resources, improves model quality, and enables organizations to scale artificial intelligence, data, analytics, and model development. As a result, organizations benefit from efficiencies, cost reductions, more robust models, transparency, computer vision experience, and expanded deployment capabilities. Implementing a data annotation concept based on the self-supervised learning approach can significantly improve computer vision performance in the manufacturing industry.
Prof. Dr. Martin Zaefferer