TY - JOUR
T1 - Unsupervised fabric defect detection with local spectra refinement (LSR)
AU - Shakir, Sahar
AU - Topal, Cihan
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature.
PY - 2024/1
Y1 - 2024/1
N2 - The inspection of fabric defects is of great importance, as undetected and uncorrected defects entail poor production quality and expensive compensation. Due to the variety of defect types and sizes, it is a very tedious task to perform inspection manually. There are numerous automated systems in the literature; however, most of them require a training scheme where clean and defective fabric samples are manually fed to the system. Because of the diversity of fabric patterns and defect classes, supervised systems reduce convenience and ease of use in real practice. In this study, we propose an unsupervised, robust fabric defect detection method using spectral domain analysis. The proposed algorithm has a very simple flow and can run without any prior training scheme. First, the algorithm splits the input textile image into smaller patches and computes a generic spectral representation of the fabric pattern. Then, the method detects defective regions by measuring dissimilarities between the spectral representation and all local patches of the input fabric. We also introduce a textile fabric dataset, i.e., Ten Fabrics Dataset, which consists of ten different types of fabrics with 27 of the most common textile defects. According to the extensive set of experiments on two different datasets, the proposed method outperforms the state-of-the-art by achieving up to 94% accuracy.
AB - The inspection of fabric defects is of great importance, as undetected and uncorrected defects entail poor production quality and expensive compensation. Due to the variety of defect types and sizes, it is a very tedious task to perform inspection manually. There are numerous automated systems in the literature; however, most of them require a training scheme where clean and defective fabric samples are manually fed to the system. Because of the diversity of fabric patterns and defect classes, supervised systems reduce convenience and ease of use in real practice. In this study, we propose an unsupervised, robust fabric defect detection method using spectral domain analysis. The proposed algorithm has a very simple flow and can run without any prior training scheme. First, the algorithm splits the input textile image into smaller patches and computes a generic spectral representation of the fabric pattern. Then, the method detects defective regions by measuring dissimilarities between the spectral representation and all local patches of the input fabric. We also introduce a textile fabric dataset, i.e., Ten Fabrics Dataset, which consists of ten different types of fabrics with 27 of the most common textile defects. According to the extensive set of experiments on two different datasets, the proposed method outperforms the state-of-the-art by achieving up to 94% accuracy.
KW - Defect detection
KW - Spectral analysis
KW - Textile fabric inspection
KW - Unsupervised learning
UR - http://www.scopus.com/inward/record.url?scp=85175872284&partnerID=8YFLogxK
U2 - 10.1007/s00521-023-09080-0
DO - 10.1007/s00521-023-09080-0
M3 - Article
AN - SCOPUS:85175872284
SN - 0941-0643
VL - 36
SP - 1091
EP - 1103
JO - Neural Computing and Applications
JF - Neural Computing and Applications
IS - 3
ER -