This study presents an investigation into the uncertainty of images reconstructed by the total focusing method (TFM) using non-destructive evaluation (NDE) and phased array probes. Four neural network architectures based on the U-Net model are used to probabilistically segment TFM images and evaluate the uncertainty of the segmentation results. The models are trained on three simulated phased-array datasets, which contain various sources of uncertainty from the simulated defects or surrounding material. Physical limitations, such as the defect’s shadow zone, led to high uncertainty. Results demonstrate that probabilistic segmentation can be helpful in determining the source of uncertainty within segmented TFM images. The model performance is investigated based on several metrics, and the influence of defect size on model performance is shown. The probabilistic U-Net shows the highest F1-score overall test datasets. This study contributes to the advancement of NDE using TFM by providing insights into the uncertainty of the reconstructed images and proposing a solution for addressing it.     «

This study presents an investigation into the uncertainty of images reconstructed by the total focusing method (TFM) using non-destructive evaluation (NDE) and phased array probes. Four neural network architectures based on the U-Net model are used to probabilistically segment TFM images and evaluate the uncertainty of the segmentation results. The models are trained on three simulated phased-array datasets, which contain various sources of uncertainty from the simulated defects or surrounding m...     »