Neural network based automated defect detection using induction thermography for surface cracks of forged parts

A fully convolutional neural network was set up for the detection of crack-type defects and for the defect shape prediction of thermography datasets. The method uses a supervised neural network for sematic segmentation (U-Net). For these tasks, training datasets of forged parts were acquired through induction thermography. The approach provides a significant improvement over conventional methods of thermal signal and image processing used in active thermography. Furthermore, the results may lead to new procedures for a quantitative evaluation of flaws and defects in non-destructive testing using infrared thermography.