Screening of Baggage X-ray Images Using Convolutional Neural Networks

For security screening of X-ray baggage, we address the problem as image classification task by the application of trained deep convolutional neural networks (CNN). Large quantities of training data are typically needed when using a deep multi-layer CNN technique to build a complete framework that obtains features and performs screening. To solve this problem, we use a transfer learning methodology that allows a pre-trained CNNs to be particularly tuned later for achieving the classification of baggage. In our study, for the classical threat and non-threat image classification, we experimented the classification task by a newly designed lighter CNN network without pre-training and compared the classification performance of pre-trained neural networks with SVM classifier using the features extracted from various layers of CNNs. Pre-trained networks achieve 99% classification accuracy and precision and exceeds the performance of CNN network without prior training. Further, the classification task by a newly designed lighter CNN network without pre-training achieves 96% accuracy, 95% precision, 5% false positive rate with SVM classifier.