Iris Recognition based on Convolutional Neural Networks and Multi-Index Hashing

Abstract

Efficient iris recognition at scale requires indexing structures that can search large galleries without a linear scan. This work evaluates 7 feature extractors, with 5 pre-trained CNNs (VGG16, ResNet50, ConvNeXt-Tiny, MobileNetV2, InceptionV3) and 2 classical descriptors (SIFT, ORB), combined with median binarization and Multi-Index Hashing on the CASIA-Iris-Lamp dataset, containing 16212 images and 411 subjects. The experiments compare extractors under a fixed 32-bit segment parameterization, analyze the trade-off between hit rate and penetration rate as a function of the number of tables, and evaluate thermometer binarization as an alternative to median. CNN extractors achieve rank-1 identification rates near 99%, while SIFT and ORB reach only 10.60% and 51.45% despite similar hit rates, showing that hit rate alone does not reflect recognition quality. VGG16 offers the best trade-off, combining a rank-1 rate of 99.45% with a penetration rate of 18.02%. Thermometer binarization improves rank-1 but produces a penetration rate of 100% across all tested configurations, eliminating the indexing advantage of Multi-Index Hashing.