Multiple similar radioactive source detection using deep learning

Loss of regulatory control over radioactive sources pose significant threats to public health and environmental safety. This paper proposes a deep-learning-based method for detecting multiple similar radioactive sources. We construct multiple single-source datasets and one multi-source dataset through Geant4. The single-source datasets vary in radiation-pixel-map resolution and in detection-window settings, while the multi-source dataset contains up to three unknown sources. We convert 1-D count-rate data into 2-D image features to generate radiation pixel maps, and we define detection windows to reduce the search space. The YOLOv8n is then employed to detect and locate radioactive sources within these images. Through comparative experiments, we determine the optimal settings for the radiation pixel map and detection windows, achieving an accuracy of over 95 % in detecting multiple un- known radioactive sources. The results demonstrate that the trained deep learning model is able to accurately and effectively detect radioactive sources.