PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Iris Recognition: A Survey

Kien Nguyen, Hugo Proença, Fernando Alonso-Fernandez

arXiv:2210.05866v1cs.CVcs.AI

TL;DR

Iris recognition research has rapidly adopted deep learning, but its benefits vary across recognition stages and challenging operating conditions. This survey synthesizes more than 200 papers, reports, and repositories across segmentation, recognition, attacks, forensics, human-machine pairing, tools, and emerging challenges. It identifies progress alongside unresolved generalization, privacy, acquisition-distance, and segmentation difficulties.

  • Problem

    Deep learning’s uneven impact across iris-recognition stages and its robustness in challenging applications require a comprehensive synthesis of methods, resources, and open problems.

  • Method

    The survey reviews and critically compares deep-learning approaches for iris acquisition, segmentation, feature encoding, matching, recognition, presentation attacks, forensics, human-machine pairing, and related resources.

  • Results

    The survey finds advances across iris-recognition tasks, including deep feature extraction, human-machine pairing, and end-to-end methods, while documenting persistent challenges in cross-domain robustness and difficult data.

  • Takeaways & Limitations

    Deep learning broadens iris-recognition research beyond conventional pipelines toward forensic, attack-resistant, human-machine, unconstrained, and synthetic-data applications.

  • Takeaways & Limitations

    Cross-database, cross-sensor, and unseen-attack generalization remain difficult, while synthetic iris generation raises identity-leakage and identity-preservation concerns.

Abstract

from arXiv · show

In this survey, we provide a comprehensive review of more than 200 papers, technical reports, and GitHub repositories published over the last 10 years on the recent developments of deep learning techniques for iris recognition, covering broad topics on algorithm designs, open-source tools, open challenges, and emerging research. First, we conduct a comprehensive analysis of deep learning techniques developed for two main sub-tasks in iris biometrics: segmentation and recognition. Second, we focus on deep learning techniques for the robustness of iris recognition systems against presentation attacks and via human-machine pairing. Third, we delve deep into deep learning techniques for forensic application, especially in post-mortem iris recognition. Fourth, we review open-source resources and tools in deep learning techniques for iris recognition. Finally, we highlight the technical challenges, emerging research trends, and outlook for the future of deep learning in iris recognition.

1 INTRODUCTION

Iris recognition is motivated by the iris’s highly distinctive, stable texture and the established scalability of classical IrisCodes. The survey examines where deep learning improves the recognition pipeline and system security.

  • Iris biometric properties: The iris texture develops before birth, remains relatively stable across the lifetime, and can support forensic identification shortly after death.The passage notes disease- and aging-related biological changes as exceptions.
  • Iris biometric properties: High randotypic randomness, permanence, and entropy give iris texture strong discriminating power compared with face or fingerprint biometrics.
  • Classical recognition: IrisCodes remain dominant because XOR-based matching supports massive-scale search with low false-match probability and high speed.The approach underlies large deployments including Aadhaar, HART, and NEXUS.
  • Classical recognition: Daugman’s IrisCodes encode iris images into 2,048-bit binary sequences using Gabor filtering and Cartesian-to-polar normalization.Hamming distance compares the resulting codes after normalization to a canonical structure.
  • Survey scope: Deep learning has been adopted across iris-recognition stages, but its impact is uneven across preprocessing, segmentation, encoding, and matching.The survey assesses where deep learning improves performance and security and where more complex models have not produced comparable benefits.
  • Survey scope: The survey reviews deep learning for segmentation, recognition, presentation-attack detection, forensic applications, human-machine pairing, unconstrained analysis, and open-source resources.

2 DEEP LEARNING-BASED IRIS SEGMENTATION

Iris segmentation must parameterize both iris boundaries and distinguish usable from occluded regions before recognition. Deep learning provides diverse segmentation frameworks, including approaches for degraded, off-angle, post-mortem, and segmentation-less data.

  • Segmentation challenges: Segmentation is difficult because iris images may be motion-blurred, poorly focused, partially occluded, off-angle, or affected by severe visible-light reflections.
  • Method landscape: Deep-learning segmentation methods are surveyed in a chronological comparison organized by data type, datasets, and proposal characteristics.The survey omits aggregate performance summaries because methods use heterogeneous validation protocols and metrics.
  • Segmentation tasks: Iris segmentation comprises inner-boundary parameterization, outer-boundary parameterization, and discrimination of unoccluded versus occluded iris regions.These outputs support dimensionless polar representations used by later feature-extraction methods.
  • Method landscape: Recent approaches include multispectral preprocessing, dense fully convolutional networks, encoder-decoder models, interleaved residual U-Nets, and off-angle re-projection.
  • Method landscape: Other methods address interactive learning, post-mortem texture extraction, handheld degraded images, and hybrid pipelines combining edge information with deep networks.
  • Segmentation-less recognition: Segmentation-less recognition bypasses accurate boundary segmentation by learning agreeing patterns directly from non-segmented or roughly segmented iris data.This direction is motivated by segmentation’s difficulty and its role as a major source of recognition errors.

3 DEEP LEARNING-BASED IRIS RECOGNITION

Deep-learning iris recognition converts normalized iris data into feature representations and matches them through classifiers or learned similarity functions. The survey covers generic and iris-specific architectures, softmax and pairwise objectives, and domain-tailored designs.

  • Feature extraction: DL-based feature extraction maps dimensionless iris representations to compact feature vectors or embeddings used for recognition.
  • Feature extraction: Feature-extraction studies examine attention mechanisms, pretrained CNNs, iris-specific weights, multi-level representations, and feature fusion.
  • Matching strategies: Matching strategies comprise conventional classifiers, softmax-based losses, and pairwise-based losses.
  • Matching strategies: ImageNet-pretrained networks provide plug-and-play iris features without requiring large iris datasets, followed by classifiers such as SVM, RF, or sparse representation.
  • Network paradigms: Classification networks learn known-identity labels end to end with softmax losses, whereas similarity networks learn distances between same- and different-identity image pairs.
  • Network paradigms: Similarity networks support verification and identification over open identity sets by pulling genuine pairs together and pushing impostor pairs apart.
  • Network design: Similarity architectures commonly use shallower fully convolutional networks that preserve spatial correspondence for pixel-to-pixel matching.
  • Network design: Complex-valued networks are proposed as domain-tailored models because iris texture is stochastic and lacks the consistent shapes or semantic structure of generic objects.The design retains phase and corresponds more closely to complex-valued representations used by handcrafted iris features.

4 DEEP LEARNING-BASED IRIS PRESENTATION ATTACK DETECTION

Deep learning-based PAD addresses presentation attacks that can impersonate or conceal identities, with methods increasingly designed for varied attacks, sensors, databases, and imaging conditions. The survey reviews attention, lightweight, multimodal, synthetic-data, and cross-domain approaches while identifying generalization, limited data, demographic bias, and authentication-use questions as open concerns.

  • 4 DEEP LEARNING-BASED IRIS PRESENTATION ATTACK DETECTION: Presentation attacks use fake or altered iris samples to impersonate another identity or conceal the attacker’s own identity.Examples include printouts, screen replays, deceased-irises, and textured contact lenses.
  • 4.2 End-to-end Classification Networks: Lightweight and specialized architectures target mobile or heterogeneous settings while supporting cross-attack, cross-sensor, and cross-database evaluation.MobileNetv3-small uses overlapped micro-stripes and majority voting, while D-NetPAD and MVANet address varied presentation attacks and imaging environments.
  • 4.2 End-to-end Classification Networks: Attention-based methods improve PAD by weighting informative regions or combining spatial and channel dependencies, often outperforming baseline or prior state-of-the-art systems.WRN combines local region weighting with global classification; AG-PAD uses heatmaps to highlight annular iris texture; A-PBS fuses multi-scale spatial attention.
  • 4.2 End-to-end Classification Networks: Alternative inputs and modalities can strengthen PAD: edge-enhanced fusion, supervised class encoding, and OCT each showed benefits under particular evaluation conditions.OCT outperformed NIR and VW for intra-attack evaluation, whereas NIR generalized better to unseen attacks; cosmetic lenses remained harder to detect than artificial eyes.
  • 4.4 Adversarial Networks: Synthetic image generation improved hand-crafted and deep PAD methods, but robust generalization to unseen sensors, databases, imaging conditions, and attacks remains difficult.Existing databases are limited in size, demographic performance differences have been observed, and cosmetic lenses may be worn normally rather than to fool authentication.

5 DEEP LEARNING-BASED FORENSIC IRIS RECOGNITION

Forensic iris recognition extends iris biometrics to deceased subjects, but decomposition substantially changes iris appearance and creates segmentation challenges. Deep learning methods address these artifacts and can provide visual explanations for human examiners.

  • Post-mortem iris recognition targets deceased subjects, including comparisons across intervals from hours after death to several weeks.
  • Decomposition causes post-mortem images to differ significantly from live images, with drying-related highlights and visible muscle furrows complicating segmentation.
  • Trokielewicz et al. combined a SegNet segmenter with Siamese feature extraction trained solely on post-mortem iris samples.
  • Coarse masks enabled rubber-sheet normalization, while fine masks excluded decomposition-driven artifacts from encoding.
  • Kuehlkamp et al. used Class Activation Mapping to highlight salient regions, helping examiners inspect iris areas or verify algorithmic decisions.

6 HUMAN-MACHINE PAIRING TO IMPROVE DEEP LEARNING-BASED IRIS RECOGNITION

Human-machine pairing is presented as a way to combine complementary attention patterns in forensic iris comparison. Machine-generated visual cues may help examiners identify discriminatory regions that human observers miss.

  • Forensic iris decisions require confirmation by trained human experts, creating a role for deep learning in human-assisted analysis.
  • Human attention maps are learned from gaze tracking, whereas machine attention maps use Grad-CAM to identify regions contributing to model predictions.
  • Humans and machines focus on limited iris areas, but their salient regions differ in location, appearance, and density.
  • Machine visual cues may complement forensic expertise by highlighting discriminatory regions that human examiners might miss.
  • Machines can outperform humans on healthy, easy iris pairs, whereas humans can outperform machines on disease-affected pairs.

7 RECOGNITION IN LESS CONTROLLED ENVIRONMENTS: IRIS/PERIOCULAR ANALYSIS

Periocular recognition extends iris analysis to the surrounding eye region, offering a complementary biometric signal for poor-quality or less constrained imagery. Deep learning methods use transfer learning, attention, generative models, semantic information, and score fusion across iris and periocular data.

  • Periocular recognition uses information around the eye as a trade-off between face and iris biometrics, particularly in poor-quality environments.
  • Off-the-shelf CNN features can transfer effectively to periocular recognition despite being optimized for generic object classification.
  • CNN-based approaches address surveillance variability through adaptive regions of interest, mid-level feature selection, and channel-spatial attention.
  • Other methods reduce intra-class variability, compensate for surveillance image variation, or learn features jointly from identity and semantic data.
  • Cross-spectral approaches translate images between domains or jointly learn spectrum-specific features and cross-spectral correspondence.
  • Fusion methods combine iris and periocular information to adapt matching or exploit complementary evidence under less constrained conditions.

8 OPEN-SOURCE DEEP LEARNING-BASED IRIS RECOGNITION TOOLS

The survey reviews datasets and publicly released tools for deep learning-based iris segmentation, recognition, and presentation-attack detection. Existing resources are concentrated in near-infrared data and common attack types, while multiple open implementations cover segmentation and recognition pipelines.

  • The resource review summarizes datasets, open-source software, and tools for segmentation, recognition, and presentation-attack detection.
  • Near-infrared data dominate iris resources, although visible-wavelength images are widely used in segmentation and recognition studies.
  • PAD datasets classify instruments including printouts, textured lenses, replay displays, artificial eyeballs, synthetic irises, and post-mortem irises.
  • Presentation-attack datasets mainly use live textured lenses and paper printouts, while replay attacks remain residual in datasets and recent studies.
  • Released segmentation tools include U-Net, off-the-shelf CNN, multi-task attention, benchmarked CNN, and V-Net implementations across several datasets.
  • Open recognition packages include triplet-loss, pretrained ResNet, hybrid handcrafted-feature, and post-mortem-oriented segmentation and recognition systems.

9 EMERGING RESEARCH DIRECTIONS

Emerging directions target resource efficiency, interpretability, synthetic-data risks, long-range acquisition, privacy, and robustness to variable acquisition conditions. The survey also identifies weakly supervised or unsupervised segmentation as a major challenge.

  • 9.1 Resource-aware designs of iris recognition networks: Resource-aware designs trade model size and computation against accuracy through lightweight architectures, compression, and neural architecture search.Lightweight CNNs reduce parameters and FLOPs, while NAS can incorporate computation and memory into architecture design.
  • 9.2 Interpretability and human-machine pairing: Interpretability remains necessary because deep networks are opaque, motivating visualization methods that show how iris regions contribute to matching scores.Class Activation Maps and pixel-level decompositions are cited as approaches for making matching outputs more interpretable.
  • 9.3 Deep learning-based synthetic iris generation: Synthetic iris generation reduces collection burdens and privacy concerns but raises risks of identity leakage and difficulty preserving identities during controlled generation.The survey notes that identity leakage can expose training-set information, while identity preservation for existing subjects remains insufficiently addressed in iris synthesis.
  • 9.4 Deep learning-based iris super-resolution: Long-range iris recognition is constrained by the need for high-resolution imagery, with current acquisition typically requiring subjects to remain within 60 cm.Super-resolution methods based on stacked auto-encoders and CNNs are explored as responses to low-resolution biometric imagery.
  • 9.6 Segmentation challenges: Segmentation remains especially difficult in less constrained environments because acquisition conditions produce highly varying data features.Emerging work emphasizes interpretable models, minimal architectures, and weakly supervised or unsupervised frameworks.

10 CONCLUSIONS

Deep learning has expanded across the iris-recognition processing chain, motivating a comprehensive survey of methods and open problems. The paper concludes by emphasizing unresolved challenges such as unsupervised learning, black-box models, and online learning as directions for future research.

  • Deep learning methods now span preprocessing, segmentation, feature extraction, matching, and recognition in iris biometrics.
  • The survey reviews historical and state-of-the-art approaches across iris acquisition, segmentation, quality assessment, feature encoding, matching, and recognition.It also critically compares methods and identifies open problems for each phase.
  • Unsupervised learning, black-box models, and online learning remain prominent challenges and potential directions for future research.
Loading 2210.05866v1…