| Title |
Improving Industrial Metal Surface Defect Detection through the Integration of Fringe based Phase Data |
| Authors |
김기범(Gibeom Kim) ; 김혜진(Hyejin Kim) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.11.66 |
| Keywords |
Deep learning; Machine learning; Unsupervised anomaly detection; Fringe pattern |
| Abstract |
Phasemap data acquired from fringe patterns have been employed to detect minute defects on highly reflective surfaces such as metals in a non-contact manner. However, previous studies have generally relied on either the wrapped phase map or the unwrapped phase map alone, or have adopted supervised approaches that require defect data. In this study, we propose a method that identifies defects after training only on normal patterns via an anomaly detection technique, thus operating without defect labels, and that robustly detects defects ranging from small and shallow to large and deep by simultaneously using both wrapped and unwrapped phasemap data to compensate for the weaknesses of each. We applied the representative anomaly detection models PaDiM and PatchCore to fringe based phase data and detected surface defects by fusing the anomaly scores obtained from the two types of phase maps. Performance evaluation using defect data generated by simulation confirmed that the proposed method achieves high detection accuracy irrespective of the defect full width at half maximum (FWHM) and depth. In our experiments, we observed the following: (i) the wrapped phasemap detects defects more readily when both depth and FWHM are large; (ii) the unwrapped phasemap is more effective when depth is large and FWHM is small; and (iii) combining the two mitigates phase-wrapping ambiguity and patch level misses, improving robustness across the full range. |