Performance Evaluation of Image Super-Resolution for Cavity Detection in Irradiated Materials

Radiation-induced swelling in structural materials is a significant challenge for the safe operation of nuclear power plants. One of the promising solutions to accelerate the understanding of swelling is the development of machine learning tools for accelerated characterization of irradiated microstructures, focusing on cavities which contribute to a materials swelling response. In this paper, we examine an object detection model for cavities, YOLOv8, for cavity detection using the datasets from the Canadian Nuclear Laboratory (CNL) and Nuclear Oriented Materials & Examination (NOME). Specifically, we explored the use of Image Super-Resolution (ISR) techniques to enhance the detection performance either in the underfocused or overfocused condition for improved small cavity detection. The overall F1-score increase of 6.8% via ISR, highlights the effectiveness of ISR techniques in enhancing the model’s performance in detecting irradiation-induced cavities across all modalities. Furthermore, this study presents a comparative evaluation of YOLOv8 and Faster R-CNN, discussing their detection trade-offs and suitability for different imaging conditions. In addition to evaluating overall detection accuracy, this work assesses the impact of ISR on both models’ precision and recall for cavity detection.