Enabling Instance Segmentation: A Semi-Automatic Method for Thermal Event Annotation

Contemporary infrared imaging systems in thermonuclear fusion devices are preventing thermal overloads on plasma-facing components (PFCs) relying on the surface temperature. Automatic delineation and classification of thermal events would facilitate scene understanding, contributing to advanced machine protection, control, and physics exploration applications. However, the absence of image annotations, which require a significant amount of expert labor and are prone to inconsistencies, limits the use of deep learning computer vision methods in fusion devices. A semi-automatic annotation method based on deterministic infrared image processing is proposed to reduce annotation efforts while maintaining consistency. The method exploits discharge sequence properties to minimize expert involvement. It was evaluated on infrared images from the Wendelstein 7-X (W7-X) stellarator by comparing the generated annotation with manually prepared ground-truth annotations. The generated annotations have a high mean similarity to the manual annotations, measured with (), equal to <tex-math notation="LaTeX">0.825</tex-math> with a sample standard deviation of <tex-math notation="LaTeX">0.030</tex-math> . Furthermore, a customized metric temperature over limit weighted SDC (tlwSDC), which weighs pixel severity based on the surface temperature relative to the PFC temperature limit, is proposed, and this mean similarity is equal to <tex-math notation="LaTeX">0.904</tex-math> with a sample standard deviation of <tex-math notation="LaTeX">0.018</tex-math> . Encouraging results for an infrared image from the W Environment in Steady-state Tokamak (WEST) tokamak indicate that the method might be cross-device viable. The proposed semi-automatic method enabled the generation of an annotated image dataset and, consequently, the training of the first W7-X instance segmentation model.