Balanced X-ray Security Dataset based on SIXray and PIDray

The experiments in this study were conducted on a balanced dataset created from two X-ray security inspection datasets, SIXray and PIDray, which have overlapping prohibited item categories. The SIXray dataset is a large-scale dataset for prohibited item detection in X-ray security screening. It contains 8,929 X-ray images with manual annotations for prohibited items, primarily covering five object categories: knives, guns, wrenches, pliers, and scissors. The PIDray dataset was collected from various scenarios including airports, subway stations, and train stations, comprising a total of 47,677 images. This dataset was generated using three different security inspection devices, resulting in images of various sizes and color schemes, covering twelve common prohibited item categories: guns, knives, wrenches, pliers, scissors, hammers, handcuffs, batons, spray cans, power banks, lighters, and bullets. All prohibited item categories in the SIXray dataset are included within those of the PIDray dataset. Due to a severe imbalance in the number of labels corresponding to prohibited item categories between the two datasets, this paper first merged the SIXray and PIDray datasets. Then, data was randomly selected from the combined dataset to form a new balanced dataset. This new dataset consists of a total of 13,728 images. The distribution of the number of samples per class label in the balanced dataset is shown in Fig. 12. The resulting dataset was partitioned into training, validation, and test sets according to an 8:1:1 ratio.