On using feature descriptors as visual words for object detection within X-ray baggage security screening

摘要

Here we explore the use of various feature point descriptors as visual word variants within a Bag-of-Visual-Words (BoVW) representation scheme for image classification based threat detection within baggage security X-ray imagery. Using a classical BoVW model with a range of feature point detectors and descriptors, supported by both Support Vector Machine (SVM) and Random Forest classification, we illustrate the current performance capability of approaches following this image classification paradigm over a large X-ray baggage imagery data set. An optimal statistical accuracy of 0.94 (true positive: 83%; false positive: 3.3%) is achieved using a FAST-SURF feature detector and descriptor combination for a firearms detection task. Our results indicate comparative levels of performance for BoVW based approaches for this task over extensive variations in feature detector, feature descriptor, vocabulary size and final classification approach. We further demonstrate a by-product of such approaches in using feature point density as a simple measure of image complexity available as an integral part of the overall classification pipeline. The performance achieved characterises the potential for BoVW based approaches for threat object detection within the future automation of X-ray security screening against other contemporary approaches in the field.