Classification and Detection of Prohibited Objects in X-Ray Baggage Security Images

Abstract

Airport security is strengthened by baggage screening to prevent potential threats and ensure safety. Researchers have been striving to create more precise models for baggage screening through the use of techniques such as fusion, denoising, and histogram equalization. While object detection in X-ray baggage images mostly relies on the Bag of Visual Words (BoVW) model, Convolutional Neural Networks (CNN) have also demonstrated potential. In recent studies, image projection and transfer learning in deep neural networks have been utilized to increase the accuracy of detecting potentially dangerous items. One study specifically improved the detection of obscured prohibited items by incorporating a De-occlusion Attention Module into a custom dataset of X-ray scanned images. In this research, the focus is on exploring the potential of computer vision for prohibited item detection and classification through image enhancement using a custom dataset resembling X-ray scanned images, as the main contribution. X-Ray machines are commonly used at high-security locations such as airports to scan baggage for hidden prohibited objects, like guns, knives, and razor blades. The research team added new layers to a pre-trained model using transfer learning, manually labeled the images, and enhanced the contrast. The YOLOv2, YOLOv3 and YOLOV5 models were employed to classify X-ray baggage images and detect obscured prohibited items, resulting in a mean average precision (mAP@0.50) of 59.44%, 98.61% and 84.1% respectively. These outcomes indicate the practical applicability of the suggested strategy for real-time baggage screening systems.