Topics

Nuclear-Material Classification and Unknown-Material Detection via Electron Microscope Image Analysis for Nuclear Forensics
−A Novel Nuclear-Material Identification Technique Using Deep Metric Learning Models−

 

Fig.1 Discrimination of nuclear material electron microscope images using a CNN model trained by deep metric learning approach

Fig. 1 Discrimination of nuclear material electron microscope images
using a CNN model trained by deep metric learning approach

Electron microscope images of the nuclear material segmented into small squares are projected into a multidimensional space using a convolutional neural network model. The unknown-material images will be detected based on the distance from the representative points of the known-material images. In this example, the images of UOC-3, which is assumed to be a hypothetical unknown material, are distributed far from the representative points of the known-material images (UOC-1: 0(→), UOC-2: 1(→)).

 As part of the development of nuclear forensic analysis technology to determine the origins of nuclear and radioactive materials recovered from a crime scene, we aim to develop a technique to identify nuclear materials with high accuracy by analyzing morphological characteristics that can be observed using electron microscopes using convolutional neural network (CNN) models.
 Although these CNN models realize highly accurate image classification, they find difficulty in accurately detecting unknown materials that have not been used for model training. This can be a serious concern in nuclear forensic analysis, which may involve samples originating from unknown materials. We have developed a novel technique that simultaneously helps in classifying nuclear materials and identifying unknown materials through electron microscope image analysis by employing an approach called “deep metric learning,” which is used in facial recognition, in CNN model training (Fig. 1a).
 By employing natural uranium ore concentrate (UOC) reference materials, CNN model optimized via deep metric learning (deep metric learning model) can perform UOC classification with high accuracy; these models can also detect unknown materials with higher accuracy compared to the models trained using the conventional training approach (Fig.1b).
 This technique will enable nuclear-material identification in nuclear forensic analysis.

 
Click hire to see details

  • ● Difference between the conventional CNN model and deep metric learning model

     A CNN model converts an input image into a vector (feature vector) that extracts local features, and the image data is analyzed as a coordinate point in a multidimensional space called a feature space. For image classifications, the feature vector is converted into a final model prediction called Logits using an arbitrary function, and the probability of belonging to each image class will be calculated by inputting Logits into the Softmax function (Fig. 2a).
     The deep metric learning applied in this study is an optimization approach that improves the image discriminability of the CNN model by changing its “Head” part, which converts feature vectors into the Logits, to a special structure. The deep metric learning model is able to convert images into feature vectors in such a way that the distance between different class images is increased, and the distance between same class images is decreased. This makes it possible to clearly distinguish between images of known and unknown materials (Fig. 2b shows an example of the feature vector conversion using a conventional CNN model and a deep metric learning model). Since the deep metric learning model can also calculate the class probabilities, it can also be used for image classifications in the same way as a conventional CNN model.

    •  

  • Fig.2 Differences between the conventional CNN model and the deep metric learning model

    Fig. 2 Differences between the conventional CNN model and the deep metric learning model

    •  
     

    謝辞

    This study was conducted as part of a project funded by MEXT under the "Subsidy for Nuclear Security Enhancement Promotion Project."

     

    Author (Researcher) Information

     
    Name | Yoshiki Kimura
    Technology Development Promotion Office, Integrated Support Center for Nuclear Nonproliferation and Nuclear Security
     

    Reference

    Application of Deep Metric Learning Model to Microscope Image Analysis for the Determination of UOC Samples in Nuclear Forensics Analysis, Journal of Radioanalytical and Nuclear Chemistry, 3rd International Conference on Radioanalytical and Nuclear Chemistry (RANC 2023), Budapest, Hungary, vol.333, issue 7, 2024, p.3541‐3551.

    Paper URL: https://doi.org/10.1007/s10967-023-09300-w

    ISCN Newsletter No.0331 > 3. Technology and Research (in Japanese)

    February 14, 2025

     Development of Science & Technology for Nuclear Nonproliferation 

    To R&D Navigator top ▼