Uncertainty Estimation for Non-destructive Detection of Material Defects with U-nets

Thermographic Imaging is a method for the detection of material defects inside the inspected specimen in a non-destructive manner. U-net was recently proposed to be used for thermographic imaging, and represents the new state-of-the-art in the field. In this paper, we propose two computationally very cheap methods to estimate the uncertainty of such u-net predictions. We demonstrate the efficacy of our method through the utilization as a loss proxy for rejection above a specific loss. For this, we use two models that differ in training data and compare with an ensemble based method.