Developing Training Data for Sandpaper Defect Detection with a Labeling Platform
Hsu, Elsa (2023)
Hsu, Elsa
2023
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2023060220739
https://urn.fi/URN:NBN:fi:amk-2023060220739
Tiivistelmä
This paper investigates a number of factors that influence the maximal utilization of training data in a detection effort of sandpaper defects: 1) the selection of defects and compilation of labeling instructions, 2) the data-centric practices of preparing and labeling training data, and 3) the utilization of an external labeling workforce. The exploration of these factors intends to cast light on their significance and influence in the overall detection effort. Furthermore, we present precision, recall, and F-1 scores based on labeling results from three rounds, indicating notable progress in optimizing the quality of the labeling output through systematizing the labeling instructions and standardizing the labeling process. The results demonstrate the gains achieved by establishing informative instructions and optimizing the process, leading to improved accuracy and reliability of the labeled data. The findings and reflections also suggest potential improvements in future instructions and reflect on the untapped opportunities for facilitating the process.