Application of Deep Learning Techniques in Uranium Microparticle Fission Track Detection

Uranium microparticle isotopic ratios are closely linked to uranium enrichment activities and distinctly differ from natural uranium isotopes. Through isotopic analysis of uranium microparticles, important information regarding material origins, production processes, and products can be obtained. This analysis aids in determining the presence of undeclared uranium enrichment activities within nuclear facilities. Therefore, the identification and analysis of uranium microparticles are of significant research value. Currently, the microparticle analysis of uranium microparticles predominantly relies on manual methods, which are heavily dependent on the operator’s experience and are laborious and time-consuming. The aim of this paper was to achieve automatic recognition and detection of uranium microparticle fission tracks, thereby significantly reducing manual labor costs and partially replacing traditional manual operations. This provides technical support for the subsequent intelligent detection of uranium microparticle tracks. Polycarbonate material was utilized to fabricate a double-layer structure for the formation of uranium microparticle fission track samples. These samples were irradiated under a thermal neutron fluence of 1.2×1014 cm−2 to create fission tracks. After cooling, the samples were transferred to a microscope to obtain the original images of uranium microparticle fission tracks. The YOLOv5 algorithm, a deep learning approach, was employed as the foundational network model. To address the issue of long-distance dependencies in convolutional operations, a window multi-head attention mechanism (swin transformer) was integrated to design the uranium microparticle detection network. This network was then subjected to experimental comparison. The results indicate that the YOLOv5-ST model achieves an average mean accuracy of 89.2%, a detection precision of 90.1%, and a recall rate of 89.6% for the detection of uranium microparticle fission tracks. Compared to the YOLOv5 model, the average mean accuracy, precision, and recall rate improve by 1.9%, 4.5%, and 10.1%, respectively. The performance of four loss functions—CIOU, DIOU, SIOU, and GIOU was compared. It is found that the integrated network alters the performance of the loss functions, with the DIOU loss function being most compatible with the YOLOv5-ST network. In conclusion, the YOLOv5-ST network model demonstrates superior detection effects for uranium microparticle track identification, particularly when paired with the DIOU loss function. Future research will focus on further enhancing the model’s detection accuracy and investigating detection time to increase its practical application value. Subsequent research on intelligent detection of uranium microparticle tracks can be expanded based on this technology.