Prohibited Items Detection in X-ray Images in YOLO Network

Abstract

In order to safeguard public spaces from security issues, such as terrorism, security mechanisms have long played a crucial role. With the increase of population and crowd density in public transportation hubs of big cities, rapid, automatic, and accurate detection of prohibited items in X-ray scanning images becomes increasingly significant. Therefore, a one-stage detection algorithm, namely an improved You Only Look Once (YOLO) algorithm, is proposed. Firstly, the datasets are put to the the third version of YOLO(YOLOv3) network for iterative training by using a loss function named Distance Intersection over Union (DIoU). Secondly, the Spatial Pyramid Pooling (SPP)[15] model is utilized in the YOLOv3 network, can help to obtain feature maps from images of any size. Finally, the training and test results are visualized through the Tensorboard toolkit for performance evaluation. The experiment is also trained in two datasets named COCO and PASCAL VOC. The experimental findings demonstrate that the approach employed in this paper has better Frame Per Second (FPS) than other one-stage object algorithms such as Single Shot Multibox Detector (SSD), Resnet50-SSD and YOLOv3. The mean Average Precision (mAP) improves 2% than the original YOLOv3 network. The SIXRay datasets, derived from real images acquired of security checks in several subway stations, is used for testing under real-world conditions. Overall, the new method has been proven highly effective and holding promising potentials for large-scale implementation.