Homsh's Groundbreaking Breakthrough: ViT+ArcFace
Iris recognition accuracy reaches the world's top level
With an Equal Error Rate (EER) of only 0.29% and ROC AUC approaching the theoretical limit —
We have redefined the boundaries of iris recognition with Vision Transformer
▲ Vision Transformer redefines the underlying paradigm of iris feature extraction
I. This Time, It's Not Just Progress — It's a Paradigm Shift
If you ask an engineer who has worked in iris recognition for two decades: "What is the hardest problem you have ever tackled?"
He will probably pause for a moment, then say: "The Rubber Sheet."
Since John Daugman proposed the IrisCode algorithm in 1993, the "Rubber Sheet unwrapping" process has been like an incantation etched into the DNA of iris recognition systems worldwide. Unwrapping the circular iris into a rectangular image, then extracting textures using Gabor filters... this workflow has been used for three decades, and no one questioned it.
Until we decided to throw it away.
II. Why Did the Rubber Sheet Stop Working?
Vision Transformer (ViT for short) is one of the most dazzling technological breakthroughs in the deep learning field over the past three years. It slices an image into a number of 16×16 "patches", uses the self-attention mechanism of language models to understand the global structure of the image, and outperforms the convolutional neural networks (CNNs) that dominated for years in multiple top-tier visual tasks.
When we first tried to apply ViT to iris recognition, the initial results were disappointing: the Equal Error Rate (EER) was as high as 4.65%, far below expectations.
The team quickly identified the root cause: the Rubber Sheet "flattens" the 64 × 512 pixel annular iris into a rectangle, which is then scaled to the 224 × 224 input required by ViT — a 3.5x vertical stretch and 2.3x horizontal compression. The natural radial/circumferential texture structure of the iris was severely distorted, making it impossible for ViT's patch attention mechanism to perceive the semantics within.
In other words: we had been feeding the smartest model in the wrong way.
The solution sounds simple, yet it required the courage to break convention — abandon the Rubber Sheet and switch to ROI circular cropping: with the center of the iris as the origin, crop a square area (2.5x the radius) to preserve the natural spatial symmetry of the iris, then directly resize it to 224×224 and feed it into ViT. In this way, each 16×16 patch can perceive the authentic, undistorted iris texture.
III. Key Metrics: EER = 0.29%, ROC AUC = 0.9999
Changing this single preprocessing step brought about a world of difference:
| Solution |
EER |
Remarks |
| Round 1: ViT + Rubber Sheet |
4.65% |
Traditional workflow |
| Round 2: CNN + Rubber Sheet |
2.80% |
Backbone replacement with limited improvement |
| Round 3: ViT + ROI Cropping |
~0.12%* |
Critical breakthrough |
| Final Version: ViT-S/16 + ROI + Regularization |
0.29% |
Production-grade solution |
*Round 3 results are not subject to rigorous statistical verification and contain optimistic bias.
The final released system adopts ViT-S/16 (22.1M parameters) + ArcFace angular margin loss, trained on a fusion of 8 public datasets (a total of 4,480 identities / 67,704 images). After rigorous statistical verification, the results are as follows:
● EER = 0.29% (Equal Error Rate)
● 95% Confidence Interval: [0.21%, 0.40%] (200 Bootstrap resampling rounds)
● ROC AUC = 0.9999 (nearly perfect score)
● Mean genuine pair similarity: 0.8742 (high consistency for the same individual)
● Mean impostor pair similarity: 0.0450 (complete feature separation for different individuals)
● At FRR=1%, FAR = 0.00% (zero false recognition at high-security operating points)
▲ ROC Curve (AUC=0.9999) and Genuine/Impostor Score Distribution — Two Peaks Completely Separated
IV. Training Data: Not Just Large, But Diverse
This study fused 8 public datasets, including the two most challenging scenarios in the industry:
Twin Data (CASIA-Iris-Twins)
Iris data from 200 pairs of twins — even with nearly identical genes, the iris textures are completely different. This is the "ultimate test" to verify the discriminative power of the algorithm.
Visible Light Unconstrained Scenarios (UBIRIS.v2)
518 identities with over 11,000 images, captured under natural lighting with motion blur, out-of-focus distortion, and illumination variations — this is the dataset closest to real-world deployment scenarios.
Training was completed on an Apple Silicon M2 Ultra (Mac Studio) in approximately 12.3 hours (90 training epochs), with a peak inference latency of only ~35ms (including ROI cropping and feature extraction).
V. Horizontal Comparison with Top Industry Work
| Method |
Backbone |
Preprocessing |
EER |
| Daugman IrisCode |
Gabor |
Rubber Sheet |
~0.10% (Controlled Environment) |
| UniqueNet (2016) |
Siamese CNN |
Rubber Sheet |
0.18% |
| IrisFormer (2023) |
ViT-B/16 |
Rubber Sheet |
0.22% |
| PolyIRIS (2021) |
Multi-scale CNN |
Rubber Sheet |
(Single Dataset) |
| Homsh ViT+ArcFace (This Release) |
ViT-S/16 |
ROI Cropping |
0.29% (8 Datasets) |
▲ From 4.65% to 0.29% EER: The Technological Evolution Path of Four Rounds of Iteration
VI. Next Steps
1.Cross-Dataset Independent Evaluation
Blind testing on the IIT Delhi dataset not involved in training to verify real-world generalization ability.
2.Liveness Detection Integration
Combine multi-frame flash response or texture analysis to defend against photo playback attacks and build a complete anti-spoofing system.
3.Medium and Long-Range Iris Recognition
Introduce medium-range (3m) data to extend to scenarios with larger capture distances — the next blue ocean for commercial implementation.
4.Lightweighting and Edge-Side Deployment
Distill the ViT-S/16 model to <5M parameters to adapt to resource-constrained edge devices (NPU/FPGA).
Conclusion: A Thirty-Year Convention Deserves Re-examination
Daugman's Rubber Sheet was the optimal solution of its era. But the essence of technology is this: when better tools emerge, the old paradigm should step aside.
Vision Transformer has changed the underlying logic of image recognition. Through four rounds of experiments and four months of exploration, we have found the correct way for ViT to truly unlock its potential in iris recognition — not to make ViT adapt to the old workflow, but to design a new preprocessing paradigm tailored for ViT.
An EER of 0.29% is just a number, but also a declaration:
Iris recognition has entered the Transformer era, and Homsh is at the starting line.
About Homsh
WuHan Homsh Technology Co., Ltd. (HOMSH), founded in 2011, is one of the few high-tech enterprises in the world that holds independent intellectual property rights for core iris recognition algorithms and chips. Its core Phaselirs™ algorithm and Qianxin Series FPGA/ASIC intelligent chips for iris recognition have been widely used in financial collection, customs clearance, government certificate issuance, military security and other fields.
