Few-Shot X-Ray Prohibited Item Detection Based on Self-Attention Prototype Learning

In Prototype network for few-shot learning, prototypes are commonly calculated by averaging samples. However, this approach does not account for differences in sample quality. It also ignores the relationships between different metatasks. These limitations weaken the prototypes' performance. To address these two issues, this paper proposes a self-attention-based prototype learning network for few-shot X-ray prohibited item detection. The approach employs a self-attention mechanism to allocate weights to support samples. This enables the model to learn better prototype representation and achieve more accurate detection. The model is evaluated on two X-ray datasets and compared with state-of-the-art methods. Our method achieves improvements of 3.4% and 1.5% in novel class accuracy respectively. The experimental results validate the effectiveness of the proposed modules and demonstrate that the proposed method enhances the accuracy of prohibited item detection under few-shot conditions.