Semantic segmentation for non-destructive testing with step-heating thermography for composite laminates

In this paper, semantic segmentation networks such as UNet and DeepLabV3+ are evaluated and compared against Random Forest and Support Vector Machines in the field of step-heating active infrared thermography for subsurface defect detection and localization. To collect information from an entire digital recording sequence into a particular image, post-processing methods such as PCT, PPT, Kurtosis, Skewness and TSR are used. Two datasets are created, one with 3-channel images using PCT, and one using all the above post-processing methods to condense the heating and cooling processes into 30-channel images. This evaluation study shows that DeepLabV3+ is able to detect most defects in specimens with a similar structure to training samples without false positives even for defects of different depth and area. UNet requires the use of 30-channel images to achieve results closer to DeepLabV3+. Random Forest and Support Vector Machines are unable to compete with the recent methods as they are unable to detect defects correctly.