Logo des Repositoriums

Auflistung nach Autor:in "Binder, Frank"

1 - 2 von 2
  • Konferenzbeitrag
    Classifying figures and illustrations in electronics datasheets: A comparative evaluation of recent computer vision models on a custom collection of 4000 technical documents
    (INFORMATIK 2023 - Designing Futures: Zukünfte gestalten, 2023) Perakis, Lymperis; Balling, Julian; Binder, Frank; Heyer, Gerhard; Kreupl, Franz
    We report findings from a comparative evaluation of several recent object detection models applied to a domain-specific use case in technical document analysis and graphics recognition. More specifically, we apply models from the EfficientDet and YOLO model families to detect and classify figures in electronics datasheets according to a custom classification scheme. We identify YOLOv7-D6 as the most accurate model in our study and show that it can successfully solve this task. We highlight an iterative approach to figure annotation in document page images for creating a comprehensive and balanced custom dataset for our use case. In our experiments, the object detection models show impressive performance levels on par with state-of-the-art results from the literature and related studies.
  • Textdokument
    Digitizing Drilling Logs - Challenges of typewritten forms
    (INFORMATIK 2021, 2021) Bürgl, Kim; Reinhardt, Lea; Binder, Frank; Müller, Lydia; Niekler, Andreas
    In this work, we show prospects of how mining and geological documentation in the form of drilling reports can be digitized and further processed. Processing these typed and handwritten forms poses challenges for document management in renaturation projects. We highlight the structural problems of drilling reports and present three approaches for recognizing and processing the information documented in them. We use optical character recognition and document layout analysis techniques to approach the problem. Layout analysis was performed using a heuristic approach and a neural network for layout recognition. In detail, we show the approaches Form Processing (A), Table detection by line counting (B) and processing with Mask-R-CNN (C). A case study is used to show initial results and challenges. B and C are more robust than A to small changes in the form. C can recognize columns better with more training data than B in cases where table boundaries are not respected. B and C also allow other language models to be used for OCR and can thus also recognize handwriting with appropriate training data.