Improving feature-based object recognition for X-ray baggage security screening using primed visualwords

Abstract

We present a novel Bag-of-Words (BoW) representation scheme for image classification tasks, where the separation of features distinctive of different classes is enforced via class-specific featureclustering. We investigate the implementation of this approach for the detection of firearms in baggage security X-ray imagery. We implement our novel BoW model using the Speeded-Up Robust Features (SURF) detector and descriptor within a Support Vector Machine (SVM) classifier framework. Experimentation on a large, diverse data set yields a significant improvement in classification performance over previous works with an optimal true positive rate of 99.07% at a false positive rate of 4.31%. Our results indicate that class-specific clustering primes the feature space and ultimately simplifies the classification process. We further demonstrate the importance of using diverse, representative data and efficient training and testing procedures. The excellent performance of the classifier is a strong indication of the potential advantages of this technique in threat object detection in security screening settings.

Object classification in 3D baggage security computed tomography imagery using visual codebooks

“…In the classification of baggage containing handguns they reported that the method does not work well in isolation but results can be improved using the extra information available from the colour image (indicating material type). In an extension of [29], Turcsany et al [30] present a BoW model using the Speeded-Up Robust Features (SURF) [31] and a Support Vector Machine (SVM) [32] classifier for automated object recognition within 2D X-ray baggage imagery. Correct classification rates in excess of 99% and a false positive rate of approximately 4% are demonstrated on a diverse dataset.…”

On using Feature Descriptors as Visual Words for Object Detection within X-ray Baggage Security Screening

“…This poses an interesting challenge for the use of automatic object recognition approaches akin to the prior work of [1,2,3,4]. In addition, an associated ability to automatically assess the underlying complexity of a given X-ray baggage image facilitates the potential of "auto-clearing" low complexity baggage items (e.g.…”

“…The work of Baştan et al [1] considers the concept of Bag-of-Visual-Words (BoVW) within X-ray baggage imagery using Support Vector Machine (SVM) classification with SIFT feature descriptors [7] achieving performance of 0.7, 0.29, 0.57 recall, precision and average precision, respectively. Turcsany et al [4] followed a similar approach, extending the work of [1], using BoVW with SURF feature descriptors [8] and SVM classification together with a modified version of vocabulary generation to yield 99.07% true positive, and 4.31% false positive on firearms detection over 2000 examples. A BoVW approach with SIFT feature descriptors, augmented with SPIN X-ray intensity features [9], and SVM classification is also used in [3] for the classification of single and dual view X-ray images with best average precisions achieved for gun and laptop objects of 94.6% and 98.2%.…”

“…With a notable common focus across a range of prior work in this domain on the BoVW feature representation approach [1,4,5,3,12,11,2], the objective of this study is thus to investigate the relative performance of a number of state-of-the-art feature point detection and description approaches for this task. With contrasting trends towards the use of end-to-end convolutional neural network (CNN) for the complete feature detection to classification pipeline (e.g.…”

“…[14,15]), here we aim to provide a definitive benchmark of over a range of hand-tuned features of varying complexity (namely: FREAK [16], DAISY [17], BRISK [18], ORB [19], KAZE [20], AKAZE [21]) against the mainstay of prior work in the field (i.e. SIFT [7] / SURF [8] with [1,4,3,2]). To these ends, we present an overview of a classical BoVW architecture (Section 2) into which we present both the by-product availability of complexity analysis ( Figure 2) and our comparative performance evaluation (Section 3).…”