Prohibited Items Detection in X-ray Images Based on Task Decoupling YOLOv5

X-ray security inspection is a critical security measure in airports, train stations and other areas with dense populations. However, due to the intricate nature of X-ray imaging and the intense occlusion between objects, the result of general object detection algorithms is not satisfactory. By exploiting the efficient YOLOv5 algorithm, we propose an Attention Task Decoupling Head (ATDH)to decouple the features used for classification and regression tasks. ATDH consists of a channel attention adjustment module (CAAM) and a spatial attention adjustment module (SAAM). These two lightweight modules make task-specific adjustments to the input features of YOLOv5's shared prediction heads from channel dimensions and spatial dimensions, respectively. The unreasonable situation of using the same feature to predict classification tasks and regression tasks with different information preferences is avoided. In addition, we also have implemented SimOTA dynamic sample assignment approach to flexibly adapt to the requirements of different training stages and different object instances for dividing positive and negative samples. Experiments on datasets including OPIXray, SIXray, CLCXray, and HIXray show that our approach has a significant performance improvement over the YOLOv5 benchmark.