Abstract
Security X-ray object detection faces significant challenges from "endogenous domain shift"—performance degradation caused by subtle variations in scanner hardware and environments. This paper adapts the ALDI++ (Align and Distill) framework to this domain. By employing a teacher-student self-distillation approach combined with modern transformer-based backbones (ViTDet), the proposed ALDI-ray method achieves new state-of-the-art results on the multi-domain EDS dataset, significantly outperforming previous specialized unsupervised domain adaptation (UDA) methods.
Introduction
Detecting prohibited items like knives or power banks in cluttered luggage is a critical security task. While deep learning has advanced general object detection, security X-ray imagery suffers from endogenous domain shift. These are discrepancies intrinsic to the data acquisition process—such as sensor characteristics or scanning geometry—that are often imperceptible to humans but detrimental to neural network performance. Since manual annotation for every new scanner deployment is impractical, Unsupervised Domain Adaptation (UDA) is essential for transferring knowledge from a labeled source domain to an unlabeled target domain.
Methodology
ALDI-ray leverages the ALDI++ framework, prioritizing robust self-distillation over traditional hard pseudo-labeling.
The architecture consists of several key components:
- Teacher-Student Model: A Student model learns from labeled source data, while a Teacher model—maintained as an Exponential Moving Average (EMA) of the student—generates targets for unlabeled target data.
- Multi-task Soft Distillation: Rather than filtering predictions with a hard confidence threshold, the student learns from the teacher’s raw "soft" outputs for both RPN localization and RoI classification, reducing confirmation bias.
- Robust Burn-in: A pre-training phase with heavy data augmentation ensures the Teacher provides reliable guidance before adaptation begins.
- Backbone Scaling: The framework evaluates the transition from traditional CNNs (VGG-16) to Vision Transformer-based detectors (ViTDet), which better capture global context in cluttered scenes.
Experimental Setup
The method was evaluated on the EDS dataset, a benchmark for endogenous shift in X-ray security. It contains three sub-datasets (EDS1, EDS2, EDS3) across 10 categories, including Drill/Bottle (DB), Knife (KN), and Powerbank (PB). Experiments covered all six possible transfer directions between these domains (e.g., $D_{1 \to 2}$).
Results and Discussion
ALDI-ray establishes a new state-of-the-art for X-ray UDA. With a VGG-16 backbone, it achieved an average mAP of 56.4%, surpassing the previous best method, PSN (54.77%).
The most significant gains came from scaling the backbone. As shown in the results, upgrading to ViTDet led to a massive performance leap. For instance, in the $D_{2 \to 3}$ transfer task, ViTDet reached 79.43% mAP, a substantial improvement over the 61.14% achieved by VGG-16. This suggests that transformers are particularly well-suited for the "visual noise" inherent in security X-rays. Furthermore, the model showed marked improvements in difficult, thin-profile categories like Knife (KN) and Glassbottle (GB).
Conclusion
ALDI-ray demonstrates that combining soft self-distillation with powerful transformer backbones effectively bridges the endogenous domain gap in security X-ray detection. While some categories like Laptops remain challenging due to high inter-domain variance, the framework provides a robust foundation for future work in automated threat detection. Future research will explore specialized modules to account for the unique physics of X-ray occlusion and density summation.
Summary updated