Towards to Human Intention: A few-shot open-set object detection for X-ray hazard inspection

X-ray-prohibited item inspection is an important measure to maintain public safety and has been widely applied in industries such as transportation and logistics. Without bells and whistles, it is a classic few-shot object detection task due to the long-tailed distribution of data in the real world, i.e., rare items occur at low frequencies. Taking into account that most previous few-shot methods are designed to detect known categories and are unable to perceive unknown objects, this is not conducive to the screening of novel types of prohibited items because the predefined categories in specific datasets cannot cover all possible classes in the natural world. In this work, we propose a Few-shot Open-set X-ray Object Detection method (FOXOD) composed of four novel components, namely the Overlapping Object Separator (OOS), Unknown Interest Advisor (UIA), Knowledge Augmentation (KA), and Discriminant Classifier (DC). Our method aims to achieve perceptual alertness for all known prohibited items and potential unknown objects based on a few data. To alleviate the problem of high overlap between objects, OOS utilizes a hierarchical attention mechanism to sample local knowledge and decouple the features of overlapping individual targets. In addition, UIA is carefully designed to assist RPN in perceiving potential unknown proposals without supervision, while KA mitigates catastrophic forgetting of known classes by the model. Moreover, DC prevents overfitting and further identifies unknown positive objects from a large number of background proposals. Extensive experiments on SIXray and VOC datasets demonstrate that FOXOD outperforms other baselines in both X-ray and common scenarios. Code is available at: https://github.com/liumaozhen-lmz/-FOXOD.