Performance Testing of Commercially Available Tools for Logo and Text Detection and Identification for Safeguards Applications

Recent increases in performance of open-source computer vision models and the availability of enterprise or cloud-based computational power has made computer vision more tenable than ever before. Furthermore, the volume and diversity of open-source information to be reviewed by the International Atomic Energy Agency (IAEA) for potential safeguards relevance continues to grow – both providing an opportunity for deeper safeguards evaluation and creating an overwhelming burden on safeguards analyst resources. For the past five years, Lawrence Livermore National Laboratory and Sandia National Laboratories have been evaluating opportunities, technical trends, and capabilities related to the use of computer vision tools to support open-source information collection and analysis for IAEA safeguards. New tools to detect and identify logos and text within images could provide a new level of support to safeguards analysts seeking clues regarding a state’s nuclear activities beyond more traditional capabilities like image classification or object detection. In this paper, we will present findings from a recent evaluation of an open-source computer vision platform on logo and text identification, including evaluation on specially curated nuclear-relevant images containing text and logos.