PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Periocular Biometrics Research

Fernando Alonso-Fernandez, Josef Bigun

arXiv:1810.03360v1cs.CV

TL;DR

Face and iris recognition can face non-ideal acquisition conditions, motivating interest in periocular biometrics for unconstrained use. This survey synthesizes databases, detection and segmentation, recognition features, and related tasks; it reports advantages over iris in difficult settings and utility for soft-biometrics and altered appearance.

  • Problem

    The periocular region is of interest as an independent or complementary biometric when face or iris systems encounter non-ideal conditions and relaxed acquisition is desirable.

    Source: §1 · p. 1

  • Method

    The paper surveys periocular research across databases, detection and segmentation, global and local recognition features, discriminative regions, modality comparisons, and related tasks.

    Source: §1 · p. 1, §8 · p. 13

  • Results

    Under difficult acquisition conditions periocular recognition is reported superior to iris, while periocular features are comparable to whole-face features for gender and ethnicity classification and can aid or outperform face matchers after plastic surgery or gender transformation.

    Source: §6 · p. 10, §8 · p. 13

  • Takeaways & Limitations

    The survey presents periocular biometrics as a useful independent or complementary modality across unconstrained recognition and tasks beyond identity matching.

    Source: §1 · p. 1, §8 · p. 13

    Use of databases acquired with personal devices such as smartphones or tablets remains limited in the reviewed research.

    Source: §8 · p. 13

Abstract

from arXiv · show

Periocular refers to the facial region in the vicinity of the eye, including eyelids, lashes and eyebrows. While face and irises have been extensively studied, the periocular region has emerged as a promising trait for unconstrained biometrics, following demands for increased robustness of face or iris systems. With a surprisingly high discrimination ability, this region can be easily obtained with existing setups for face and iris, and the requirement of user cooperation can be relaxed, thus facilitating the interaction with biometric systems. It is also available over a wide range of distances even when the iris texture cannot be reliably obtained (low resolution) or under partial face occlusion (close distances). Here, we review the state of the art in periocular biometrics research. A number of aspects are described, including: i) existing databases, ii) algorithms for periocular detection and/or segmentation, iii) features employed for recognition, iv) identification of the most discriminative regions of the periocular area, v) comparison with iris and face modalities, vi) soft-biometrics (gender/ethnicity classification), and vii) impact of gender transformation and plastic surgery on the recognition accuracy. This work is expected to provide an insight of the most relevant issues in periocular biometrics, giving a comprehensive coverage of the existing literature and current state of the art.

1. Introduction

Periocular biometrics can support unconstrained recognition because it permits relaxed acquisition and avoids iris segmentation. This survey reviews the research landscape, from databases and detection to recognition features and related tasks.

1. Introduction

  • Relaxed acquisition makes the periocular region suitable for unconstrained or uncooperative scenarios and avoids difficult iris segmentation.It can also complement face and iris recognition under non-ideal conditions.Source: §1 · p. 1
  • The survey covers periocular databases, detection and segmentation algorithms, recognition features, modality comparisons, soft-biometrics, and effects of gender transformation and plastic surgery.It includes face and iris databases as well as newer databases designed specifically for periocular research.Source: §1 · pp. 1–2
  • Recognition features are organized as global descriptors of an entire ROI or local descriptors around selected key points.Source: §1 · p. 1
  • Holistic processing of a predefined eye ROI can include non-identity information such as hair or glasses, and some features may vary in discriminative value across the region.The survey reviews work addressing these issues and comparisons or fusion with face and iris modalities.Source: §1 · p. 1

2. Databases

Periocular studies draw on face, iris, and dedicated periocular databases, whose acquisition conditions and annotations vary. Reported database accuracies are only approximate because studies may use different subsets or protocols.

2. Databases

  • Reported best accuracies are approximate because studies may use different database subsets or protocols; facial databases generally show better accuracy, while newer periocular databases are only some steps behind.Facial databases are also the most used, so later work builds on previous research.Source: §2 · p. 2

2.1. Facial Databases

  • Facial databases span varied capture conditions, including portal video, controlled and uncontrolled images, aging, pose, expression, illumination, and occlusion.Examples include MBGC portal sequences with NIR and visible-light video and Plastic Surgery images with one pre- and one post-surgery image per person.Source: §2.1 · p. 2

2.2. Iris Databases

  • Iris databases used in periocular research include close-up NIR collections and UBIRIS v2 visible-light images captured at 3–8 meters, including a real-world second session.The real-world session includes natural light, reflections, contrast changes, defocus, occlusions, blur, and off-angle images.Source: §2.2 · pp. 2–3

2. Databases

  • Table 1 covers publicly available databases and indicates reported best performance and ground-truth availability.Source: Table 1 · p. 3

2.3. Periocular Databases

  • Dedicated periocular databases include UBIPr images captured at 4–8 meters and FOCS portal video with small irises and substantial image-quality variation.IMP includes NIR, visible-light, and night-vision data; CSIP captures four smartphones across ten camera and flash setups.Source: §2.3 · p. 3

3. Detection and segmentation of the periocular region

Detection and segmentation research seeks to locate periocular regions without relying on full-face or iris detection. Approaches include direct eye localization, eyebrow and component segmentation, and methods whose success can vary under challenging conditions.

3. Detection and segmentation of the periocular region

  • On the difficult FOCS portal database, correlation-filter eye detection achieved 95% success, while evaluated iris segmentation algorithms achieved 51%–90% accuracy.The eye-position result does not by itself ensure successful iris segmentation.Source: §3 · p. 3
  • Full-face detection can make periocular extraction depend on accurate face localization, while iris segmentation may also be unreliable under challenging conditions.Eye detection can instead be a preprocessing step for iris segmentation.Source: §3 · p. 3
  • Direct eye-localization methods include face-part detectors, hyperspectral band combination, Gabor features, symmetry filters, and eyebrow shape models with graph-cut segmentation.Symmetry filters need no training and can use a few one-dimensional convolutions; the eyebrow model segments pixels after locating the eyebrow region.Source: §3 · pp. 3–4
  • A pixel-level method uses seven specialized classifiers to label iris, sclera, eyelashes, eyebrows, hair, skin, and glasses; sclera-only methods use color spaces guided by prior iris-boundary detection.Source: §3 · p. 4

4. Recognition using periocular features

This section reviews periocular recognition approaches using global and local descriptors, including common feature choices and preprocessing for image variation and recognition conditions. It also covers cross-spectral recognition and Light Field Camera research, including spectrum-aware matching, multi-focus capture, and recognition features.

4.1.1. Textural-based features

  • LBP is a widely used texture matcher, with uniform, circular, three-patch, hierarchical, and four-patch variants; HOG and SIFT are also used.Standard LBP compares each pixel with its 3 × 3 neighborhood and pools the resulting codes into patch histograms; subspace methods and applying LBP to other features are also common.Source: §4 · p. 4, §4.1.1 · pp. 6–7

4.3. Literature review of periocular recognition works

  • Human observers and computer experts performed similarly when comparing periocular images, and humans performed better with VW than NIR illumination.Source: §4 · p. 6

4.1. Global features

  • Global features summarize the ROI in fixed-length vectors by concatenating patch descriptors, making vector-distance matching time efficient.Source: §4.1 · p. 6

4.1.2. Shape-based features

  • Shape-based descriptors encode eyelid curvature and contours or eyebrow geometry and shape context; eyebrow shape has also produced accurate standalone recognition.Source: §4.1.2 · p. 7, §4.3 · p. 9

4.1.3. Color-based features

  • Local Color Histograms performed best in RG space with 4 × 4 histograms, yielding 16 features per patch.Source: §4.1.3 · p. 7

4.2. Local features

  • Local methods match variable numbers of keypoint descriptors, increasing computation, while grid-sampled descriptors provide fixed-size vectors without keypoint detection.SIFT and LBP are among the most popular methods; SIFT encodes orientation histograms in a 128-dimensional descriptor, while BRISK and ORB are presented as faster alternatives to SIFT or SURF for keypoint extraction.Source: §4.2 · pp. 7–8

4.3. Literature review of periocular recognition works

  • Studies also addressed aging and pose or illumination variation, evaluated periocular recognition across databases and modalities, and explored fusion with iris recognition.Juefei-Xu et al. (2011) used pose correction and illumination and periocular-region normalization to address age-related degradation; Woodard et al. (2010b) reported the best accuracy up to that date on FRGC VW images using RG color histograms.Source: §4.3 · pp. 8–9
  • UBIRIS v2 results advanced from PIGP to PILP, while UBIPr performance improved with varied features and eye-corner-based ROI definition.Bakshi et al. (2014) reported PIGP's Rank-1 performance surpassed prior UBIRIS v2 studies, and Bakshi et al. (2015) reported the best Rank-1 periocular performance to date on that database. Padole and Proenca (2012) reduced UBIPr EER from ∼30% to ∼20% using the midpoint of the eye corners; Uzair et al. (2015) reported the best published Rank-1 performance to date across four databases.Source: §4.3 · p. 9
  • Sharma et al. (2014) presented the IMP database with VW, NIR, and night images and used neural networks to learn spectral variability, testing LBP and HOG variants.Cross-spectral matching was the recognition task.Source: §4.3 · p. 9
  • Jillela and Ross (2014) also studied cross-spectral recognition using a proprietary database of NIR and VW images.Source: §4.3 · p. 9
  • Raghavendra et al. (2013) and Raja et al. (2014) used a Light Field Camera database, reporting superior performance over a conventional camera and proposing SRC and BSIF features.The camera captures multiple images at different focus in one capture, addressing focus quality; the database was not made available. Both studies also tested iris–periocular fusion.Source: §4.3 · pp. 9–10

5. Best regions for periocular recognition

Periocular recognition depends on which parts of the eye region are used and on illumination: useful cues differ between NIR and visible-wavelength images. Selective inclusion can improve identification while excluding irrelevant components.

5. Best regions for periocular recognition

  • Holistic regions of interest can include hair or glasses that may bias recognition, and features need not be equally discriminative across the periocular area.Source: §5 · p. 10
  • In NIR, regions around the iris are most useful, while visible-wavelength images favor surrounding skin texture and often the eyebrow region.Human judgments and automatic-algorithm studies broadly agree on these illumination-dependent patterns.Source: §5 · p. 10
  • Including eyebrows improved Rank-1 identification by 8–19%, while occluding iris and sclera reduced Rank-1 accuracy by up to 41%.The effects varied by machine expert and were observed in visible-wavelength FRGC images.Source: §5 · p. 10
  • Cheek information added no significant discrimination in one visible-wavelength study, while segmenting seven components enabled avoiding hair and glasses and reduced EER from 12.8% to 9.5%.The improvement used fused LBP, HOG, and SIFT features on UBIRIS v2 images.Source: §5 · p. 10

6. Comparison and fusion with other modalities

The survey compares periocular recognition and fusion with iris, sclera, and face modalities. Across reported difficult acquisition conditions, periocular recognition often compares favorably with iris or face, and fusion can improve performance.

6. Comparison and fusion with other modalities

  • Under difficult conditions, periocular recognition is reported as superior to iris in many settings and more robust than face under blur, downsampling, or partial occlusion.Iris fusion can still improve performance, including with visible-wavelength data.Source: §6 · p. 10

6.1. Iris Modality

  • Across iris studies, periocular–iris fusion improved performance in difficult NIR and visible-wavelength data, with modality comparisons varying by illumination.A CSIP cross-sensor periocular result was 15.5% EER, while fusion achieved 14.5% EER versus 34.4% for iris Gabor wavelets; another study reported an average 20% EER improvement.Source: §6.1 · pp. 10–12

6.2. Sclera Modality

  • Combining periocular and sclera features significantly improved EER for identity verification on UBIRIS v1.Source: §6.2 · p. 12

6.3. Face Modality

  • Periocular recognition outperformed face under extreme blur or downsampling and was more robust to partial face occlusion in reported studies.For plastic-surgery matching, face–periocular fusion reached Rank-1 recognition of 87.4%, the highest reported with that database at the time.Source: §6.3 · p. 12

7. Soft-biometrics, gender transformation and plastic surgery analysis

The survey describes periocular features as useful for soft-biometrics classification and reviews recognition studies involving gender transformation and plastic surgery. Reported results show strong classification accuracy and better periocular performance than other face regions in the cited transformation study.

7. Soft-biometrics, gender transformation and plastic surgery analysis

  • Periocular features have been used for gender and ethnicity classification, with reported accuracies comparable to those obtained using the entire face.Reported examples include 85% gender accuracy on low-resolution Flickr images, 89% or higher for gender and ethnicity, and 90% for gender on FERET.Source: §7 · p. 12
  • In a study of 38 subjects undergoing hormone replacement treatment, periocular accuracy greatly outperformed other face components and the full face.The database contained over 1.2 million face images spanning several months to three years; face matchers began to fail after only a few months of treatment.Source: §7 · p. 12
  • For matching images before and after plastic surgery, a fusion approach combining face and periocular information outperformed earlier studies using only full-face matchers.Source: §7 · p. 12

8. Conclusions and future work

The survey presents periocular biometrics as a discriminative option for unconstrained recognition, with less constrained acquisition than face or iris modalities. It points to at-distance, on-the-move, and heterogeneous-data recognition as directions for further work.

8. Conclusions and future work

  • The survey describes periocular biometrics as a discriminative option for unconstrained recognition that can tolerate expression variability and occlusion.It also notes its capacity to match partial faces.Source: §8 · pp. 12–13
  • Table 5 surveys periocular-feature studies of soft-biometrics, gender transformation, and plastic-surgery effects.Soft-biometrics classifies broad traits such as gender or ethnicity rather than uniquely identifying a person.Source: Table 5 · p. 13
  • The survey anticipates more research on ocular recognition at a distance and on the move, given periocular acquisition’s lower constraints.It notes limited use of smartphone, tablet, and surveillance-camera databases, with recognition accuracy still behind other setups.Source: §8 · p. 13
  • Detection and segmentation can define the periocular region without first segmenting the iris or detecting the full face.Eye corners are suggested as ROI reference points because they are less sensitive to gaze variation and remain visible in closed eyes.Source: §8 · p. 13
  • Periocular features show face-comparable accuracy for gender and ethnicity classification and may aid or outperform face matchers after plastic surgery or gender transformation.The survey also identifies cross-modality, cross-spectral, cross-sensor, and stand-off matching as emerging areas.Source: §8 · p. 13
  • Heterogeneous periocular data from different sensors and spectral ranges is identified as an important direction for future work.The survey connects this direction to settings where multiple sensors and vendor solutions must coexist.Source: §8 · p. 14
Loading 1810.03360v1…