FSPT: Foreground Separation Prompt Tuning for Few-Shot Fine-Grained X-ray Classification

With the rapid growth of global air traffic and the increasing complexity of security threats, aviation safety has become more critical than ever. As a key component of aerospace and electronic systems infrastructure, X-ray imaging systems is critical for ensuring operational safety in air transportation. Consequently, numerous researchs have focused on enhancing object detection performance in X-ray imagery. However, the few-shot fine-grained X-ray item classification (FSFG-XC) tasks has not received sufficient attention. Classifying prohibited items in X-ray screening poses considerable difficulties, mainly because of the high visual similarity among these items in X-ray images. Additionally, factors such as overlapping shapes, shadows, occlusion, and color fading complicate the distinction between items, especially from different viewing angles. In this work, we propose a Foreground Separation Prompt Tuning (FSPT) method to address these challenges. Our method consists of three key components, a Prompt-based Vision Transformer (PVT), an Attention-Based Foreground Separation Module (AFSM), and an Attention-guided Spatial Contrast Module (ASCM). PVT is designed to efficiently adapt tasks by updating only a small number of additional parameters while keeping the pretrained backbone frozen. Subsequently, AFSM mitigates background interference by integrating multi-layer attention mechanisms. ASCM enhances local fine-grained feature discrimination through spatial contrast learning. We also introduce our primary X-ray prohibited items dataset, FineXray, which consists of 3,955 images across 20 categories. This dataset aims to advance FSFG-XC studies in more challenging scenarios. Results from experiments conducted on the FineXray dataset along with widely used fine-grained benchmarks illustrate the superior efficacy of the proposed FSPT in comparison to current methodologies, offering a robust and efficient solution for enhancing intelligent threat recognition capabilities in aviation-focused electronic systems.