Evaluating a dual convolutional neural network architecture for object-wise anomaly detection in cluttered x-ray security imagery.

X-ray baggage security screening is widely used to
\nmaintain aviation and transport secure. Of particular interestis the focus on automated security X-ray analysis for particular
\nclasses of object such as electronics, electrical items and liquids.
\nHowever, manual inspection of such items is challenging when
\ndealing with potentially anomalous items. Here we present a dual
\nconvolutional neural network (CNN) architecture for automatic
\nanomaly detection within complex security X-ray imagery. We
\nleverage recent advances in region-based (R-CNN), mask-based
\nCNN (Mask R-CNN) and detection architectures such as RetinaNet
\nto provide object localisation variants for specific object
\nclasses of interest. Subsequently, leveraging a range of established
\nCNN object and fine-grained category classification approaches
\nwe formulate within object anomaly detection as a two-class
\nproblem (anomalous or benign). Whilst the best performing
\nobject localisation method is able to perform with 97.9% mean
\naverage precision (mAP) over a six-class X-ray object detection
\nproblem, subsequent two-class anomaly/benign classification is
\nable to achieve 66% performance for within object anomaly
\ndetection. Overall, this performance illustrates both the challenge
\nand promise of object-wise anomaly detection within the context
\nof cluttered X-ray security imagery.