An Asymmetric Siamese Transformer With Global-Local Alignment Attention for Visible-X-Ray Cross-Modality Package Re-Identification

Visible-X-ray Cross-Modality Package Re-Identification (VX-ReID) is a critical task in security inspection, aiming to match visible-light images with X-ray images. The significant modality gap between these image types poses substantial challenges in extracting robust and fine-grained modality-invariant features. To effectively address these challenges, this paper introduces a novel cross-modality feature extraction framework, the Asymmetric Siamese Transformer with Global-Local Alignment Attention (AST-GLAA). The network comprises two key components: Cross-modality Asymmetric Siamese Transformer Structure (CAST-S) and Global-Local Cross-Modality Alignment Attention (GL-CMA). CAST-S leverages an asymmetric design in one branch of the Siamese Transformer network by introducing a LayerNorm layer and incorporating modality embeddings to enhance the robustness of modality-invariant features. Meanwhile, GL-CMA facilitates the interaction between global and local features, significantly improving the representation of fine-grained features while effectively addressing spatial misalignment issues in cross-modality images. Experimental results demonstrate that the proposed method achieves state-of-the-art (SOTA) performance on the VX-ReID task, highlighting its effectiveness and potential in addressing the challenges of cross-modality package re-identification.