Noise Influence on the Fuzzy-Linguistic Partitioning of Iris Code Space

This paper analyses the set of iris codes stored or used in an iris recognition system as an f-granular space. The f-granulation is given by identifying in the iris code space the extensions of the fuzzy concepts wolves, goats, lambs and sheep (previously introduced by Doddington as ‘animals’ of the biometric menagerie) - which together form a partitioning of the iris code space. The main question here is how objective (stable / stationary) this partitioning is when the iris segments are subject to noisy acquisition. In order to prove that the f-granulation of iris code space with respect to the fuzzy concepts that define the biometric menagerie is unstable in noisy conditions (is sensitive to noise), three types of noise (localvar, motion blur, salt and pepper) have been alternatively added to the iris segments extracted from University of Bath Iris Image Database. The results of 180 exhaustive (all-to-all) iris recognition tests are presented and commented here.

💡 Research Summary

The paper investigates the stability of a fuzzy‑linguistic partitioning of iris‑code space when the underlying iris images are corrupted by various types of noise. The authors model the set of binary iris codes as an “f‑granular” space, where the granulation is defined by the extensions of four fuzzy concepts originally introduced by Doddington—wolves, goats, lambs, and sheep—collectively known as the biometric menagerie. Each iris code is assigned a fuzzy membership to one of these four categories based on its Hamming distance to other codes and a set of predefined thresholds. In a noise‑free scenario, this partitioning yields a relatively stable distribution of codes among the four animal groups, reflecting the intrinsic variability of the biometric data.

To assess how robust this partitioning is under realistic acquisition conditions, the authors introduce three distinct noise models to the iris segments extracted from the University of Bath Iris Image Database: (1) local variance noise, which simulates spatially varying illumination and contrast; (2) motion‑blur noise, which mimics camera shake or subject movement; and (3) salt‑and‑pepper noise, which represents impulsive pixel corruption. For each noise type, 60 exhaustive all‑to‑all matching experiments are performed, resulting in a total of 180 test runs. The same Hamming‑distance matcher and fuzzy‑membership calculation are applied to the noisy codes, allowing a direct comparison of the resulting f‑granulation with the baseline (noise‑free) case.

The experimental results reveal that the fuzzy partition is highly sensitive to noise. Motion blur produces the most pronounced disruption: the distribution of codes shifts dramatically, with many genuine matches (originally classified as “goats”) drifting into the “lamb” region, while impostor matches (originally “wolves”) become less distinct and often migrate toward the “sheep” region. Local variance noise yields moderate changes; it perturbs the membership values of a subset of bits, causing a noticeable but less severe reshuffling of codes among the four categories. Salt‑and‑pepper noise, despite its high‑frequency nature, leads to relatively modest alterations because it affects only isolated bits, leaving the overall structure of the fuzzy partition largely intact.

These findings challenge the assumption that the biometric menagerie provides a static, objective categorization of users. In practical deployments, where illumination changes, eye movement, and sensor imperfections are inevitable, the fuzzy animal‑group assignments can fluctuate significantly, potentially undermining security policies that rely on fixed “hard” or “soft” user classifications. The authors argue that any system leveraging the menagerie concept must incorporate dynamic re‑evaluation mechanisms, such as adaptive threshold tuning or online learning of fuzzy membership functions, to maintain reliable performance under noisy conditions.

In the discussion, the paper proposes several avenues for future work. First, more sophisticated noise models—including combined noise scenarios and realistic sensor‑specific artifacts—should be explored to better approximate field conditions. Second, the authors suggest integrating machine‑learning techniques (e.g., fuzzy‑c‑means clustering with adaptive weighting or deep neural networks that learn noise‑robust feature representations) to automatically adjust the fuzzy granulation in response to observed data drift. Third, they recommend developing preprocessing pipelines (advanced denoising, motion‑compensation, contrast normalization) that can mitigate the impact of noise before code generation, thereby stabilizing the fuzzy partition.

Overall, the study demonstrates that the f‑granular partitioning of iris‑code space, while conceptually appealing, is not inherently robust to common image degradations. The sensitivity to noise underscores the need for adaptive, noise‑aware designs in biometric systems that aim to exploit the biometric menagerie for user profiling, risk assessment, or access control. By highlighting these vulnerabilities, the paper provides a valuable foundation for subsequent research aimed at building more resilient, fuzzy‑logic‑based biometric frameworks.