Int J Performability Eng ›› 2020, Vol. 16 ›› Issue (12): 1910-1920.doi: 10.23940/ijpe.20.12.p7.19101920

• Orginal Article • Previous Articles     Next Articles

Detection and Classification of Surface Defects of Magnetic Tile based on SE-U-Net

Xincheng Caoa, b, Wanshan Liua, b, Bin Yaoa, b, *, Qixin Lana, b, Weifang Sunc, *   

  1. aSchool of Aerospace Engineering, Xiamen University, Xiamen, 361005, China;
    bShenzhen Research Institute of Xiamen University, Shenzhen, 518000, China;
    cCollege of Mechanical and Electrical Engineering, Wenzhou University, Wenzhou, 325035, China
  • Submitted on ; Revised on ; Accepted on
  • Contact: * Corresponding author. E-mail address: (B. Yao); (W.F. Sun)
  • Supported by:
    This research is supported financially by the National Natural Science Foundation of China (No. 51605403), Aeronautical Science Foundation of China (No. 20183368004), and Fundamental Research Funds for the Central Universities (No. 20720190009).

Abstract: Defects such as blowholes and cracks are inevitable in the manufacturing of the magnet tile, and online full detection is a necessary process. Image-based surface defect detection is of great significance for improving product quality and production efficiency. This paper introduces a pixel-level surface defect detection method based on a deep full convolutional network, which realizes the detection and classification of defects at the same time. Combining the U-Net architecture and the squeeze-excitation module, an SE-U-Net that adaptively fuses shallow local information and deep semantic information is constructed. With a small amount of additional computation, U-Net's accuracy in detecting small defects from the large background is improved. Data augmentation via data transfer reduces the imbalance between image background and defects and improves the learning speed of the model. The proposed method was compared with SegNet and U-Net and achieved more accurate defect detection, and the average pixel accuracy reached 0.97, which demonstrates the superiority of the improved SE-U-Net for magnetic tile surface defects.

Key words: magnetic tile, multi-defect detection, deep learning, semantic segmentation, convolutional networks