Evaluation of clustering algorithms for the analysis of thermal NDT inspections

Infrared thermography is a well-known technique for the Nondestructive Testing (NDT) of industrial components. Typically, the raw results of a thermal inspection are processed with an algorithm to enhance the defect detectability and then analyzed by an expert. A challenging point of this workflow is the final step, as the assessment made by the operator could be biased or subjective. To tackle this issue, clustering algorithms could be used to define, in an unsupervised manner, whether a region under inspection is defective or sound. In this work, a steel sample with flat bottom-hole defect is investigated in a Flash Thermography setup. The recorded thermal sequence is then analyzed with a clustering algorithm (k-means). The algorithm is applied varying different parameters and assessing, for each scenario, the performance of the clustering in terms of defect detection, quantified through specificity and sensitivity.