Spatial and temporal deep learning for defect detection with lock-in thermography

Lock-in thermography is a common non-destructive testing method for the investigation of subsurface defects in thermal conductive materials. The advantages of this method are that it is contact free and large areas can be investigated. However, in comparison to other NDT techniques (e.g., X-ray computed tomography), the resolution is lower. In order to achieve cost savings, automating the evaluation of lock-in thermography data is of great interest. To this end, various data evaluation methods have recently been introduced which enable a higher degree of automation and use more of the information in the dataset. This study focuses on using deep learning in lock-in thermography. The image stack from a lock-in thermography measurement data contains temporal and spatial information. The temporal information is normally compressed into one image, which contains either the phase or amplitude information. Since these images represent a compression of the data, information is lost. In this study, adhesive-bonded metal sheets with defects inside them were used as specimens. The ground truth of the data is determined via X-ray computed tomography. This represents a challenging segmentation task since smaller defects are not captured in the thermography measurements, and due to artifacts in the thermography images. We investigate whether different deep learning approaches using temporal and spatial information (the phase image) within the data work for lock-in thermography data. It was shown that segmenting the phase images with a U-net performs best. The results of all approaches are evaluated using probability of detection curves and conventional segmentation metrics.