EyePreserve: Identity-Preserving Iris Synthesis

TL;DR

EyePreserve tackles identity-preserving iris synthesis under pupil dilation by learning a non-linear deformation from data rather than relying on anatomical assumptions. It combines an autoencoder-based deformation model with triplet-based adversarial losses and realism constraints to generate ISO-compliant iris images while preserving identity across pupil changes. Empirical results show improvements over linear and biomechanical deformation baselines, especially for large dilation differences, and demonstrate the method's utility for dataset augmentation and forensic iris analysis. The work also provides fully reproducible resources including training data processing, mask estimation, and released code.

Abstract

Synthesis of same-identity biometric iris images, both for existing and non-existing identities while preserving the identity across a wide range of pupil sizes, is complex due to the intricate iris muscle constriction mechanism, requiring a precise model of iris non-linear texture deformations to be embedded into the synthesis pipeline. This paper presents the first method of fully data-driven, identity-preserving, pupil size-varying synthesis of iris images. This approach is capable of synthesizing images of irises with different pupil sizes representing non-existing identities, as well as non-linearly deforming the texture of iris images of existing subjects given the segmentation mask of the target iris image. Iris recognition experiments suggest that the proposed deformation model both preserves the identity when changing the pupil size, and offers better similarity between same-identity iris samples with significant differences in pupil size, compared to state-of-the-art linear and non-linear (bio-mechanical-based) iris deformation models. Two immediate applications of the proposed approach are: (a) synthesis of, or enhancement of the existing biometric datasets for iris recognition, mimicking those acquired with iris sensors, and (b) helping forensic human experts examine iris image pairs with significant differences in pupil dilation. Images considered in this work conform to selected ISO/IEC 29794-6 quality metrics to make them applicable in biometric systems. The source codes and model weights are offered with this paper.

Figures (12)

• Figure 1: EyePreserve accepts (a) a ISO/IEC 29794-6-compliant iris image (either synthetically generated of a non-existing identity, or authentic of an existing subject) and (b) a new/target shape of the iris expressed either as a mask or a new pupil-to-iris ratio. By combining (a) and (b) the EyePreserve synthesizes a new ISO/IEC 29794-6-compliant iris image with the iris texture deformed to match a given new iris shape. The proposed model preserves the identity and correctly models non-linear deformations of the iris muscle. This paper also offers the model synthesizing irises of non-existing subjects, which creates a complete pipeline for identity-preserving iris image synthesis and deformation.
• Figure 2: Illustration of the training mechanism with all loss function components explained in Sections \ref{['sssec:identity-preservation']} -- \ref{['sssec:triplet']}. Symbols: $I$ -- input image, $O$ -- output image, $T$ -- target sample, $N$ -- negative (impostor) sample, $P$ -- positive (genuine) sample, $D$ -- discriminator, $AE$ -- autoencoder
• Figure 3: Illustration of the target iris image mask estimation. Upper row: Getting the target iris mask for dilated pupils is easy, and cut a larger pupil circle in the original mask. Bottom row: For the constricted-pupil iris mask, we need to fill out the pupil region and then cut a smaller pupil circle. However, when the pupil is partially occluded by eyelids, there are residual regions that we remove by using an "inside-eyelid" mask.
• Figure 4: The similarity scores, offered by the HDBIF czajka2019domain (left) and DGR ren2020dynamic (right) methods as a function of the pupil-to-iris ratio difference ($\Delta$) between the probe and gallery samples. The dotted lines represent the similarity thresholds below which the image pair is considered a 'non-match' for a given matcher.
• Figure 5: Illustration of how state-of-the-art StyleGAN models, conditional (a) and unconditional (b), solve the identity-preserving iris texture modeling. For a conditional model, the pupil size was encoded as a condition. For unconditional model, the varying pupil-size images were generated by traveling along the linearly interpolated path in StyleGAN's latent space between latent codes representing input and target images (after projecting images into the StyleGAN latent space). HDBIF scores in green (match) or red (non-match) are calculated between the left images and the synthesized samples. We observe that identity is not preserved between two iris images with different pupil size by both StyleGAN models, while it is preserved by the EyePreserve.