Blog - How iris recognition works

Iris recognition is among the most stable biometrics. Daugman’s core idea: normalize the iris annulus to a fixed map and compare binary texture codes.

Segmentation and normalization

First locate pupil and limbus circles. Map a ring point $(r, \theta)$ onto a rectangle (rubber sheet):

$$I(x(\rho,\theta), y(\rho,\theta)) \mapsto J(\rho, \theta)$$

where $\rho \in [0,1]$ runs from pupil to limbus and $\theta \in [0, 2\pi)$.

Iris code

Texture is filtered with Gabor / log-Gabor kernels. Phase responses become a binary code $C \in \{0,1\}^{n}$ plus an occlusion mask $M$ (eyelids, specularities).

Similarity is a masked Hamming distance:

$$HD(C_a, C_b) = \frac{\|(C_a \oplus C_b) \land M_a \land M_b\|_1}{\|M_a \land M_b\|_1}$$

Lower $HD$ means a better match. Decision: $HD \le \tau$.

Pipeline

Why NIR?

Near-infrared light stabilizes iris texture across eye colors and reduces glare. Visible RGB is too pigment-dependent.

Takeaway

Iris recognition wins on template repeatability: solid segmentation + occlusion masks + a calibrated Hamming threshold $\tau$.