Automatic Threat Detection in Baggage Security Imagery using Deep Learning Models

Abstract

Automating object detection for surveillance purpose and threat detection is beneficial as it may compensate for the human error and will save time, which is of significant economic value. For the end-to-end classification process and feature extraction, the CNN approach requires large amounts of data. To overcome this limited availability of data, we have presented a transfer learning approach with various object detection models for single and multiple detections on two types of the dataset: Single-channelled (GDXray dataset) and Multichanneled(SIXray dataset). We have presented comparisons between the various models(Faster R-CNN with ResNet50, SSD with VGG16, YOLOv3 with ResNet50, and RetinaNet with ResNet50). The best results were achieved on Faster-RCNN(ResNet50) with 0.966 mAP for the four-class object detection problem(GDXray Dataset) and 0.845 mAP for the two-class object detection problem(SIXray Dataset).