Iris Recognition Using an Enhanced Pre-trained Backbone Based on Anti-Aliased CNNs

To harness the power of pre-trained image classification models on the extensive ImageNet dataset, early layers of pre-trained backbones have been used to capture relevant information about iris texture in the field of Iris Recognition (IR). This addresses the lack of extensive iris datasets that are required for adjusting millions of parameters in deep networks for IR. However, using intermediate layers as feature extractors introduces problems due to image alterations and aliasing effects due to down-sampling. This paper proposes an IR method that employs low-level convolutional filters from a pre-trained backbone, incorporating low-pass filters to mitigate aliasing effects resulting from both down-sampling and horizontal shifting, caused by minor head rotations during image acquisition. Furthermore, the backbone is enhanced by adapting circular padding to reduce the impact of horizontal discontinuities in the Rubber Sheet representation. The extracted features are then encoded using a novel adaptive threshold method, increasing the variability of binary iris codes, which are matched by using a single Hamming Distance without a bitshifting process. The proposed model exhibits significantly improved performance on publicly available CASIA and IITD datasets, along with datasets containing examples of substantial pupil dilation. Finally, we tested the robustness of our proposed method against head rotations by using datasets that were modified by horizontal shifting to simulate small head rotations.