A research team led by Professor Wang Hongqiang from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences proposed a Global-Local Alignment Attention (GLAA) model based on an Asymmetric Siamese Transformer (AST), which markedly enhances the performance of Visible-X-ray cross-modality package re-identification tasks. 

This study was published in IEEE Transactions on Information Forensics and Security.

Visible-X-ray cross-modality package re-identification is a core technology in security inspection. The challenge lies in the significant pixel-level differences between the two modal images, making it difficult for traditional methods to extract robust cross-modality invariant features. 

In this study, the researchers incorporated an asymmetric design concept into the Siamese Transformer architecture by proposing a Cross-Modality Asymmetric Siamese Transformer (CAST) structure. Embedding LayerNorm layers and modality-aware encoding in one branch effectively enhances the model's ability to extract cross-modality invariant features. 

They also designed a Global-Local Cross-modality Alignment Attention module. By modeling the interaction between global and local features, it enhances fine-grained feature representation while addressing the spatial misalignment issues in cross-modality images.

Experimental results show that the key metrics of this model on a dedicated cross-modality package re-identification dataset show significant improvement over the current state-of-the-art methods, providing reliable technical support for the intelligentization of security inspection.

This work is the first to introduce the Transformer architecture into the cross-modality package re-identification task, breaking through the limitations of existing methods that rely on symmetric convolutional networks, according to the researchers.

Illustration of the VX-ReID task and main idea (Image by WANG Hongqiang)