Year
2017
Abstract
In a previous paper on this topic [1], we presented an open source software processing technique based on SURF (Speeded Up Robust Features) feature detection and FLANN (Fast Learning Artificial Neural Network) matching, reducing radiography images to a non-sensitive, irreversible set of feature vectors. Now we describe additional results and progress towards a realistic information barrier system that utilizes open source, embedded processor boards to run the software and introduces a separation of hardware into sensitive and non-sensitive processing steps. We also present additional image processing results that demonstrate the number of feature matches generated for varying degrees of similarity between the tested item and the reference set. A non-matching item, something that is substantively different from the reference, shows very few strong, qualifying feature matches as demonstrated by our previous work. Our new work provides matching results from inspected items that appear identical externally, but are missing internal components of significance – a key question for verification inspections. The degree to which very small changes are detectable is a subject of our ongoing research.