Few-shot X-ray Prohibited-item Detection Based on Multi-scale Feature Fusion and Sample Balancing

Few-Shot Learning (FSL) aims to tackle machine learning tasks with limited data and has garnered widespread attention. Previous research has demonstrated that few-shot object detection methods based on model fine-tuning yield strong performance. However, the fine-tuning-based few-shot object detection approach also has its drawbacks: (1) limited learning ability on complex datasets, (2) high variability of a few samples during the fine-tuning phase often resulting in unstable performance, and (3) poor generalization of the model due to under-utilization of the visual features of a few samples. In this paper, we focus on X-ray security inspection of prohibited items as our research context. Firstly, we propose a sample balancing module that differs from random sampling by dividing the sampling interval into IOU values for uniform sampling, which helps uncover highly valuable challenging samples. Secondly, we introduce a multi-scale feature fusion module that enhances multi-level features by utilizing the same depth of fused balanced semantic information. This approach addresses the high variability of a small number of samples, leading to more accurate localization. Our method performs well across multiple shot settings on the PASCAL VOC, COCO, and X-ray prohibited items datasets, such as UNICOMP, with an increase of 8.8% to 27.9% in UNICOMP, outperforming most state-of-the-art methods.