Exploring Source-Free X-Ray-Prohibited Item Detection With an Assistant–Teacher Network

Wanying Jia, Tong Jia, Bowen Ma, Cunhe Huang, Yixin Tian, Hao Wang, Dongyue Chen

Abstract: Security scanners are important for maintaining long-term social stability and safety. However, different security scanners exist in endogenous domain shifts. Traditional unsupervised methods require access to the source domain for adaptation during training. However, in practical deployment, data privacy restricts access to data between different security scanners. To explore this issue, an assistant-teacher network (ATN) is proposed to address the endogenous domain shift of X-ray-prohibited item detection in the case of source-free domain adaption. In particular, to suppress the problem of false positives and loss of important information in the source-free domain adaption, we propose a common cognitive mechanism (CCM). To further prevent the model from optimizing in the wrong direction under false-positive labels, an external ATN loop is designed. In addition, an image de-stylization mechanism (IDM) is proposed to alleviate the endogenous domain shift, allowing the network to learn the essential features of the X-ray image. By utilizing internal and external loops, the ATN can adaptively apply to multiple security scanners in a source-free domain adaptation (SFDA). The comprehensive experimental results on X-ray images and natural images indicate that our method is effective. Especially in X-ray-prohibited item detection, our method has achieved state-of-the-art results compared to recent methods. Our code will be available at https://github.com/JiaWanying/ATN.git

External IDs:doi:10.1109/tim.2025.3544749