Frequency-modulated thermography and clustering analysis for defect detection in acrylic glass

The frequency-modulated thermal wave imaging (FMTWI) technique has shown its potential as a substitute for traditional lock-in thermography in non-destructive testing (NDT) as it can scan the entire material thickness in a single cycle and a relatively short amount of time. However, using FMTWI for the inspection of defects still requires tremendous human effort. In this work, multi-dimensional cluster analysis was proposed to post-process phase images within a frequency band altogether, without prior knowledge of the defect location and depth. Three clustering methods, k-means, fuzzy c-means and Gaussian mixture model (GMM), were applied to segment the defect area from a series of phase images and reconstruct the defect dimension and their performances were compared. It was found that for the shallow buried defects with 6 mm diameter, all three methods could estimate the dimension with less than 10% error, whereas for other detectable defects, GMM could maintain less than 25% error while the other two algorithms reached an error as large as 244.78%. The results demonstrated that the proposed method could automatically segment the defect area from a series of phase images and GMM would be more suitable for processing FMTWI experimental data than the other two methods investigated.