Biometrics recognition using deep learning: a survey

Shervin Minaee, Amirali Abdolrashidi, Hang Su, Mohammed Bennamoun, David ZhangArtif. Intell. Rev 2023 · DOI

TL;DR

This paper provides a comprehensive survey of deep learning techniques applied to various biometric recognition modalities. It highlights the significant advancements deep learning has brought to biometrics, covering key architectures, datasets, and challenges, and points towards future research directions.

What Problem Does This Paper Address?

Biometric recognition systems, used for authentication and identification, have traditionally relied on hand-crafted features. However, these methods often struggle with the variability and complexity of real-world data, leading to limitations in accuracy and robustness. The paper addresses the challenge of improving biometric recognition performance by surveying how deep learning models, with their ability to automatically learn hierarchical representations, have revolutionized this field. It aims to provide a structured overview for researchers entering this domain.

Methods

The paper surveys various deep learning architectures that form the backbone of modern biometric recognition systems. These include:

  • Convolutional Neural Networks (CNNs): Widely used for image-based biometrics like face, fingerprint, and iris recognition, CNNs excel at capturing spatial hierarchies and local patterns.
  • Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM): Employed for sequential data like voice and gait recognition, where temporal dependencies are crucial.
  • Auto-Encoders: Used for unsupervised learning of data representations, often for dimensionality reduction or feature extraction.
  • Generative Adversarial Networks (GANs): Explored for generating synthetic biometric data to augment datasets and for tasks like fake biometric detection.

The survey also discusses the importance of transfer learning, where models pre-trained on large datasets (like ImageNet) are fine-tuned for specific biometric tasks, and the development of novel loss functions (e.g., large-margin softmax, triplet loss) designed to enhance feature discriminability.

Datasets

The paper lists numerous datasets used across different biometric modalities. Some prominent examples include:

  • Face Recognition: Labeled Faces in the Wild (LFW), VGGFace2, CASIA-WebFace, MegaFace, IJB-C.
  • Fingerprint Recognition: FVC Fingerprint Database, PolyU High-resolution Fingerprint Database, CASIA Fingerprint Dataset.
  • Iris Recognition: CASIA-Iris-1000 Database, UBIRIS Dataset, IIT Delhi Iris Dataset.
  • Palmprint Recognition: PolyU Multispectral Palmprint Dataset, CASIA Palmprint Database, IIT Delhi Touchless Palmprint Database.
  • Ear Recognition: IIT Ear Database, AWE Ear Dataset, Multi-PIE Ear Dataset.
  • Voice Recognition: NIST SRE, SITW (Speakers in the Wild), VoxCeleb datasets.
  • Signature Recognition: ICDAR 2009 SVC, SVC 2004, Offline GPDS-960 Corpus.
  • Gait Recognition: CASIA Gait Database (multiple subsets), Osaka Treadmill Dataset, Osaka University Large Population (OULP) Dataset.

The paper does not introduce new datasets but provides a comprehensive overview of commonly used ones for each modality.

Key Insights and Ideas

  • Deep Learning's Dominance: Deep learning models have consistently outperformed traditional hand-crafted feature methods across nearly all biometric modalities, often achieving near-human or even super-human performance on benchmark datasets.
  • The Power of End-to-End Learning: Deep learning enables end-to-end training, allowing models to jointly learn feature extraction and classification, simplifying the pipeline and improving overall performance.
  • Loss Functions as Key Enablers: The development of specialized loss functions (e.g., large-margin, angular, triplet losses) has been crucial for learning highly discriminative features, especially for face recognition.
  • Transfer Learning is Essential: Due to the often limited size of specific biometric datasets, transfer learning from models pre-trained on large, general-purpose datasets (like ImageNet) is a widely adopted and effective strategy.
  • Challenges Remain: Despite significant progress, challenges like handling intra-class variations (pose, illumination, aging), large-scale datasets with millions of identities, and robustness to occlusions persist.

Key Results

The paper presents extensive tables summarizing the performance of various deep learning models on different biometric recognition tasks. For instance:

  • Face Verification: Deep learning models like FaceNet and ArcFace achieve accuracies exceeding 99.5% on the LFW dataset, significantly surpassing classical methods.
  • Face Identification: On the MegaFace challenge, models report Rank-1 identification accuracies above 95% even with millions of distractors.
  • Other Modalities: Similar trends of high accuracy are reported for fingerprint, iris, palmprint, ear, voice, signature, and gait recognition, with deep learning models consistently achieving state-of-the-art results on their respective benchmarks.

Limitations and Future Work

  • Dataset Scale and Diversity: Current datasets, while large, often do not fully represent real-world scenarios with millions of identities and extreme intra-class variations (e.g., aging, severe occlusions). Future work should focus on creating larger, more challenging datasets.
  • Interpretability of Deep Models: Understanding why deep learning models work so well and how they are susceptible to adversarial attacks remains an open question. Research into interpretable deep models is needed.
  • Few-Shot and Self-Supervised Learning: Developing models that can learn effectively from very few samples (few-shot learning) or without explicit labels (self-supervised learning) is crucial for biometrics, especially for rare identities or new modalities.
  • Biometric Fusion: Combining multiple biometric modalities (e.g., face + voice) is essential for robust and reliable systems, especially for challenging tasks like distinguishing identical twins. Developing effective fusion strategies at various levels (feature, score, decision) is a key area for future research.
  • Real-time Performance and Memory Efficiency: Many deep learning models are computationally intensive. Future work needs to focus on developing efficient models that can operate in real-time on resource-constrained devices.
  • Security and Privacy: Addressing presentation attacks, template security, and privacy concerns related to biometric data is paramount. Research into visual cryptography and robust anti-spoofing methods is essential.

Summary updated

BibTeX
@article{minaee2023biometrics,
  author = {Shervin Minaee and Amirali Abdolrashidi and Hang Su and Mohammed Bennamoun and David Zhang},
  title = {{Biometrics recognition using deep learning: a survey}},
  journal = {Artif. Intell. Rev.},
  year = {2023},
  doi = {10.1007/s10462-022-10237-x},
}
Figure