A Dangerous Goods Detection Approach Based on YOLOv3

In modern life, safety inspection plays an indispensable role at almost every corner of the society. Owing to the rapid growth of urban population and travel demand, people's requirement for security is growing. The measures of safety inspection changes from manual work to auxiliary machinery in the wake of scientific-technical progress. In the recent decades, as X-ray scanners have been widely used in safety inspection, real-time dangerous goods detection in X-ray images is more and more important. However, some dangerous goods are so small that it is difficult to detect them. In this paper, we present a detection method based on YOLOv3 which preprocess the data sets to improve detection accuracy of small objects. We divide the original images into several sub-images that are mainly in size of 416×416 by the contours of luggage. By training YOLOv3 with these sub-images, we reduce the errors of resizing input images. The experiment results show that our method perform better than YOLOv3 in small object detection.