A NESTED AUTOENCODER APPROACH TO AUTOMATED DEFECT INSPECTION ON TEXTURED SURFACES


Öz M. A. N., Kaymakçı Ö. T., Mercimek M.

INTERNATIONAL JOURNAL OF APPLIED MATHEMATICS AND COMPUTER SCIENCE, cilt.31, sa.3, ss.515-523, 2021 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 31 Sayı: 3
  • Basım Tarihi: 2021
  • Doi Numarası: 10.34768/amcs-2021-0035
  • Dergi Adı: INTERNATIONAL JOURNAL OF APPLIED MATHEMATICS AND COMPUTER SCIENCE
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Compendex, Computer & Applied Sciences, INSPEC, zbMATH, Directory of Open Access Journals
  • Sayfa Sayıları: ss.515-523
  • Anahtar Kelimeler: autoencoders, defect detection, automatic visual inspection, deep learning, CLASSIFICATION
  • Yıldız Teknik Üniversitesi Adresli: Evet

Özet

In recent years, there has been a highly competitive pressure on industrial production. To keep ahead of the competition, emerging technologies must be developed and incorporated. Automated visual inspection systems, which improve the over-all mass production quantity and quality in lines, are crucial. The modifications of the inspection system involve excessive time and money costs. Therefore, these systems should be flexible in terms of fulfilling the changing requirements of high capacity production support. A coherent defect detection model as a primary application to be used ina real-time intelligent visual surface inspection system is proposed in this paper. The method utilizes a new approach consisting of nested au-toencoders trained with defect-free and defect injected samples to detect defects. Making use of two nested autoencoders, the proposed approach shows great performance in eliminating defects. The first autoencoder is used essentially for feature extraction and reconstructing the image from these features. The second one is employed to identify and fix defects in the feature code. Defects are detected by thresholding the difference between decoded feature code outputs of the first and the second autoencoder. The proposed model has a 96% detection rate and a relatively good segmentation performance while being able to inspect fabrics driven at high speeds.