LPIXray: A Large-scale Logistics Prohibited Item X-ray Dataset for the Application of Deep Learning in Security Inspection
Chengquan He, Tong Mu, Weiping Ren, Bohua Zhao · 2023
Detecting prohibited items in X-ray images is very important for logistics public safety. For object detection, deep learning with the support of a significant volume of data has enormous advantages and has gradually become a mainstream research method. However, there are relatively few public datasets on prohibited item X-ray images, and the existing datasets have the problem of fewer prohibited items in quantity and variety. To address the above challenges, this work contributes LPIXray dataset: a Logistics Prohibited Item X-ray dataset. The LPIXray contains 60950 X-ray images of prohibited items, 84785 instances, and is carefully divided into 6 classes with 18 subclasses. These X-ray images cover almost all common prohibited item categories, and each category has a wealth of data annotation. The LPIXray dataset is 6 or even 20 times larger than previous datasets, far exceeding previous data in terms of annotation quantity and category definition. Aiming at the problem that some X-ray images of prohibited item are difficult to collect and the samples are unbalanced, a data Augmentation approach based on the General Adversarial Network (GAN) and image fusion is designed. Finally, we conduct experimental verification on the LPIXray dataset based on You Only Look Once (YOLO) model.