Inspection of commerce entering the country is currently a burdensome process, assisted greatly by non-intrusive inspection (NII) technologies such as X-Ray imaging. There are a variety of singular ML techniques that assist in the adjudication process, many are well known, common and widely available with proper data. However, the challenge is to develop a comprehensive system exhibiting certain properties that leverages many techniques simultaneously and combines results to improve overall performance. Here we describe the concepts, components, data and implementation of a system we use for NII adjudication that is well suited to adapting to future needs. We demonstrate several components of this system the benefits of it. Specifically, we show the combination of results from common methods achieves superior performance in concert. First, ensembeling has proven successful in combining results from a variety of object detection models. We show the impact of using our novel graph clique-based method to increase performance over independent models. Additionally, because of performant and reliable object detection results, we demonstrate how constructing an image level contextualization of a scene allows analysis at the metadata level. Lastly, we leverage variational autoencoder (VAE) based similarity models to describe anomalies. We deploy these algorithms in a hierarchical fashion, developing optical signatures for whole NII scans as well as their components (e.g, entire vehicle, trunk, wheels). Here, we quantify the impacts of the inclusion of these methods in image anomaly detection task and demonstrate them in the extensible framework we have developed for NII applications. This paper will serve as an overview of our approach and show results from both combinatorial techniques.
|