ProjectNov 2024 – Feb 2025
Iris Recognition: Traditional Biometric Pipeline vs Deep Learning
The classic Daugman pipeline (via the OpenIris library) against ResNet-50 embeddings trained with triplet loss, on the 20,000-image CASIA-Iris-Thousand dataset. The traditional pipeline wins clearly: about 0% EER and 99.7% rank-1 identification.
The iris is one of the most reliable biometric traits: its texture is unique, even between identical twins, and stays stable for life. This project put two ways of recognising it side by side, under the same data and the same evaluation protocol.
Two pipelines
Traditional (OpenIris). The standard biometric chain: segmentation with a dual-decoder network that separates eye geometry from noise like eyelashes and reflections, normalisation with Daugman’s rubber sheet model, Gabor filters to turn the texture into a binary iris code, and Hamming distance for matching.
Deep learning (ResNet-50). The network produces a 2048-dimensional embedding per image, trained with a triplet loss (margin 0.4) and compared with cosine similarity. We trained two versions: one on full eye images and one on the normalised iris strips.
Data and protocol
CASIA-Iris-Thousand has 20,000 infrared eye images; OpenIris segmented 18,725 of them (about 93%). We evaluated both verification (is this the claimed person?) and identification (who is this, or is it someone unknown?), both all-against-all and probe-against-gallery.
Results
Identification, probe vs gallery:
| Method | EER | Rank-1 identification | AUC |
|---|---|---|---|
| OpenIris | ≈ 0% | 99.8% | – |
| ResNet-50, full eye | 6% | 93.4% | 0.97 |
| ResNet-50, normalised iris | 10% | 90.2% | 0.94 |
In verification the gap is smaller (about 0% against 1% EER), but the traditional pipeline still comes out ahead, and its decision threshold (0.375) matched the one recommended by the library’s authors. On a clean, well-controlled dataset, decades of hand-designed iris coding are hard to beat with a general-purpose CNN; the deep models remain competitive and need no hand-tuned feature design, which would matter more on messier data.