IrisFormer: A Dedicated Transformer Framework for Iris Recognition

Abstract

While Vision Transformer (ViT)-based methods have significantly improved the performance of various vision tasks in natural scenes, progress in iris recognition remains limited. In addition, the human iris contains unique characters that are distinct from natural scenes. To remedy this, this paper investigates a dedicated Transformer framework, termed IrisFormer, for iris recognition and attempts to improve the accuracy by combining the contextual modeling ability of ViT and iris-specific optimization to learn robust, fine-grained, and discriminative features. Specifically, to achieve rotation invariance in iris recognition, we employ relative position encoding instead of regular absolute position encoding for each iris image token, and a horizontal pixel-shifting strategy is utilized during training for data augmentation. Then, to enhance the model's robustness against local distortions such as occlusions and reflections, we randomly mask some tokens during training to force the model to learn representative identity features from only part of the image. Finally, considering that fine-grained features are more discriminative in iris recognition, we retain the entire token sequence for patch-wise feature matching instead of using the standard single classification token. Experiments on three popular datasets demonstrate that the proposed framework achieves competitive performance under both intra- and inter-dataset testing protocols.