Logo des Repositoriums

Auflistung nach Autor:in "Bundscherer, Maximilian"

1 - 2 von 2
  • Konferenzbeitrag
    Machine Learning in Glass Bottle Printing Quality Control: A Collaboration with a Medium-Sized Industrial Partner
    (INFORMATIK 2024, 2024) Bundscherer, Maximilian; Schmitt, Thomas H.; Bocklet, Tobias
    In cooperation with a medium-sized industrial partner, we developed and evaluated two ML-based approaches for quality control in glass bottle printing. Our first approach utilized various filters to suppress reflections, image quality metrics for image comparison, and supervised classification models, resulting in an accuracy of 84%. We used the ORB algorithm for image alignment and to estimate print rotations, which may indicate manufacturing anomalies. In our second approach, we fine-tuned pre-trained CNN models, which resulted in an accuracy of 87%. Utilizing Grad-CAM, an Explainable AI method, we localized and visualized frequently defective bottle print regions without explicitly training our models for this use case. These insights can be used to optimize the actual manufacturing process beyond classification. This paper also describes our general approach and the challenges we encountered in practice with data collection during ongoing production, unsupervised preselection, and labeling.
  • Konferenzbeitrag
    Training a Computer Vision Model for Commercial Bakeries with Primarily Synthetic Images
    (INFORMATIK 2024, 2024) Schmitt, Thomas H.; Bundscherer, Maximilian; Bocklet, Tobias
    In the food industry, reprocessing returned products is a vital step to increase resource efficiency. [SBB23] presented an AI application that automates the tracking of returned bread buns. We extend their work by creating an expanded dataset comprising 2432 images and a wider range of baked goods. To increase model robustness, we use generative models pix2pix and CycleGAN to create synthetic images. We train state-of-the-art object detection models YOLOv9 and YOLOv8 on our detection task. Our overall best-performing model achieved an average precision AP 0.5 of 90.3% on our test set.