Cooperative Verification Using Radiography Behind an Information Barrier

Sandia National Laboratories has been conducting research into utilizing radiography, combined with automated image processing algorithms, to create a novel method of non-invasive verification. In many treaty verification scenarios, inspectors must verify the authenticity or identity of items that contain sensitive features. While radiography is a powerful inspection tool, it also reveals a great deal of detail about an item that may not be allowed by a verification agreement. Automation of the image processing task enables use of an information barrier, giving inspectors confirmation that an inspected item matches a previous measurement or agreed template while protecting sensitive information about the item. Our technique utilizes feature matching in radiographic images of complex items. The SURF (Speeded Up Robust Features) method is used to extract features from the images. FLANN (Fast Learning Artificial Neural Network) is used in the matching process. The feature list becomes the template. The SURF features are somewhat rotation, scale, and translation invariant, which means the reference and target images need not be taken from the exact same position for the source and film, making data collection easier. A significant discovery is that we can discard the position information of the features and still perform the matching adequately. With no position information, geometry cannot be recovered; we believe it is impossible to reconstruct the image in this case, creating an irreversible transform that creates non-sensitive feature lists, or templates. This method is analogous to using a paper shredder to prevent reconstruction of an original while still being able to match features from the individual shredded pieces. Results of these image processing techniques on radiography simulations are promising, showing high correlation between features from identical items, even at slightly different measurement angles. Items not matching the original have significantly lower correlation with the feature set, enabling an automated decision process. We provide examples and results from complex electro-mechanical systems to demonstrate the effectiveness of this technique in the automatic verification of such items, and a path forward to the creation of a complete verification system with an information barrier.