Link to paper | PDF

Authors: R. Augustauskas, L. Zabulis, A. Lipnickas, S. Jokubauskas

Published in: 2023 IEEE 12th International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS)

Abstract: Computer vision-empowered quality is an important feature in modern manufacturing. These systems can perform defect inspections faster and more accurately than humans. There has been a rise in adopting deep-learning-based solutions for visual inspection tasks. However, new data is still necessary for each specific problem, and precise dataset annotation can be time-consuming considering precision-demanding approaches such as segmentation or region detection. This investigation proposes a method inspired by class activation mapping-based (CAM) to enhance deep neural network-based image classification algorithms by outputting an attention map where the defective area is indicated. With additional lightweight operations, it helps to roughly localize defects place in the image without using pixel-wise annotations. The proposed method is tested on glass bottles (Oliena), printed circuit boards (PCB defects), and industrial machine tool component surface defects (BSData) datasets.