Deep Learning for Topmost Roller Chain Detection Using Data Augmentation

Positioning topmost object from a pile of metal sheets using image processing is a typical topic in the field of industry. Relative new methodologies mainly focus on improving robustness under complex state based on deep learning. While deep learning and CNN have demonstrated capacity in instance segmentation, the preparation of dataset is time consuming. To accelerate the preparation, this paper proposes a data augmentation technique that can simulate imaging features of roller chains and generates a large dataset. Experiments of roller chains show no significant difference between the dataset by manually annotating and the dataset by data augmentation. And the methodology reduces manual annotation by 99.9% and needs few labors.