Transfer Learning Neural Networks for Nuclear Forensic Image Morphology Using Image Splitting Techniques

Manual morphological analysis of actinide particles from scanning electron microscope (SEM) imagery is a critical component of nuclear forensics but is prone to significant inter‐analyst variability. To address this challenge, this work develops and evaluates an automated classification method using deep learning. We introduce a methodology based on partitioning 1906 SEM images, representing 13 classes of uranium compounds, into smaller patches for analysis. Three convolutional neural network (CNN) architectures of increasing complexity were compared: a custom baseline CNN, a simple transfer learning model using ResNet50v1, and a complex model featuring hierarchical feature extraction and a spatial attention mechanism built upon ResNet50v2. The final image classification was determined using a patch‐based, class‐balanced voting system. The complex model achieved a superior balanced accuracy of 94% on the test set, significantly outperforming the simple transfer learning model (89%) and the baseline model (77%). These results demonstrate that a sophisticated transfer learning architecture can serve as a robust, objective tool to increase the accuracy and consistency of actinide particle classification, thereby enhancing nuclear forensic capabilities.