Real-Time Instance Segmentation of Metal Screw Defects Based on Deep Learning Approach

Authors

  • Wei-Yu Chen Department of Biomedical Engineering, National Yang Ming Chiao Tung University, Taipei, Taiwan
  • Yu-Reng Tsao Department of Biomedical Engineering, National Yang Ming Chiao Tung University, Taipei, Taiwan
  • Jin-Yi Lai Department of Biomedical Engineering, National Yang Ming Chiao Tung University, Taipei, Taiwan
  • Ching-Jung Hung Department of Biomedical Engineering, National Yang Ming Chiao Tung University, Taipei, Taiwan
  • Yu-Cheng Liu Department of Biomedical Engineering, National Yang Ming Chiao Tung University, Taipei, Taiwan
  • Cheng-Yang Liu Department of Biomedical Engineering, National Yang Ming Chiao Tung University, Taipei, Taiwan

DOI:

https://doi.org/10.2478/msr-2022-0014

Keywords:

defect inspection, metal screw, deep learning, convolutional neural network

Abstract

In general, manual methods are often used to inspect defects in the production of metal screws. As deep learning shines in the field of visual detection, this study employs the You Only Look At CoefficienTs (YOLACT) algorithm to detect the surface defects of the metal screw heads. The raw images with different defects are collected by an automated microscopic camera scanning system to build the training and validation datasets. The experimental results demonstrate that the trained YOLACT is sufficient to achieve a mean average accuracy of 92.8 % with low missing and false rates. The processing speed of the trained YOLACT reaches 30 frames per second. Compared with other segmentation methods, the proposed model provides excellent performance in both segmentation and detection accuracy. Our efficient deep learning-based system may support the advancement of non-contact defect assessment methods for quality control of the screw manufacture.

Downloads

Published

22.04.2022

How to Cite

Chen, W.-Y., Tsao, Y.-R., Lai, J.-Y., Hung, C.-J., Liu, Y.-C., & Liu, C.-Y. (2022). Real-Time Instance Segmentation of Metal Screw Defects Based on Deep Learning Approach. Measurement Science Review, 22(3), 107–111. https://doi.org/10.2478/msr-2022-0014