Martin Braunsperger
Data-Centric Methods for Efficient Deep Assembly Verification: A Case Study in Electronic Lock Manufacturing

Information

Abstract

This thesis investigates data-efficient deep learning methods for visual assembly ver- ification in a highly customizable electronic lock production line. We examine three data-centric strategies. First, ROI-based methods leverage the fixed geometry of the workpiece carriers by reformulating object detection as a set of classification or regression tasks over predefined regions. Second, pseudo bounding boxes are created by combining image-level labels with known part positions to automatically generate object detection annotations without manual labelling. Third, synthetic training data is produced by compositing cropped part images with background scenes, thereby increasing dataset diversity and reducing the need for extensive manual data collection. These methods are evaluated on a real-world dataset collected during regular production to assess their effectiveness in reducing manual annotation effort. We provide a compre- hensive comparison of data-centric approaches, highlighting their respective strengths and limitations. The results demonstrate that leveraging the structured nature of the assembly environment enables accurate model performance with substantially reduced annotation requirements.

Additional Files and Images

Additional images and videos

teaser: An example from the dataset presented in the thesis showing a Double-sided hybrid Euro cylinder with electronic thumbturn and keyway teaser: An example from the dataset presented in the thesis showing a Double-sided hybrid Euro cylinder with electronic thumbturn and keyway

Additional files

thesis
thesis

Weblinks

No further information available.

BibTeX

@bachelorsthesis{Braunsperger-2025_bsc,
  title =      "Data-Centric Methods for Efficient Deep Assembly
               Verification: A Case Study in Electronic Lock Manufacturing",
  author =     "Martin Braunsperger",
  year =       "2025",
  abstract =   "This thesis investigates data-efficient deep learning
               methods for visual assembly ver- ification in a highly
               customizable electronic lock production line. We examine
               three data-centric strategies. First, ROI-based methods
               leverage the fixed geometry of the workpiece carriers by
               reformulating object detection as a set of classification or
               regression tasks over predefined regions. Second, pseudo
               bounding boxes are created by combining image-level labels
               with known part positions to automatically generate object
               detection annotations without manual labelling. Third,
               synthetic training data is produced by compositing cropped
               part images with background scenes, thereby increasing
               dataset diversity and reducing the need for extensive manual
               data collection. These methods are evaluated on a real-world
               dataset collected during regular production to assess their
               effectiveness in reducing manual annotation effort. We
               provide a compre- hensive comparison of data-centric
               approaches, highlighting their respective strengths and
               limitations. The results demonstrate that leveraging the
               structured nature of the assembly environment enables
               accurate model performance with substantially reduced
               annotation requirements.",
  month =      jun,
  address =    "Favoritenstrasse 9-11/E193-02, A-1040 Vienna, Austria",
  school =     "Research Unit of Computer Graphics, Institute of Visual
               Computing and Human-Centered Technology, Faculty of
               Informatics, TU Wien ",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2025/Braunsperger-2025_bsc/",
}