Abstract: To address the issue of eyelid occlusion affecting recognition performance in iris recognition, a solution based on ViT(Vision Transformer) is proposed. Firstly, a FFM(Feature Fusion Module) is proposed to achieve feature extraction and fusion at different scales, solving the problem of information loss during feature extraction. Secondly, the local feature encoder is pre-trained by minimizing reconstruction loss to avoid forming triplets with heterogeneous irises sharing the same dominant features. This prior knowledge endows the model parameter adjustment with certain interpretability. An interactive encoding structure is constructed with ViT and residual blocks as the core, efficiently fusing information from different iris blocks to form comprehensive feature representation. Finally, the traditional triplet loss is improved by incorporating the threshold concept, providing a clearer learning direction for model training. Experimental results show that the proposed method can effectively eliminate the negative impact of occlusion on iris recognition and significantly improve recognition performance.

Key words: iris recognition, Vision Transformer, feature fusion, triplet loss