T-Center: A Novel Discriminative Feature Extraction Approach for Iris Recognition

Yifeng Chen, Cheng Wu, Yiming WangAAAI 2019 · DOI

Abstract

Iris recognition systems face challenges in scalability due to the difficulty of establishing reliable classification thresholds as database sizes increase. This paper, presented at AAAI-19, proposes T-Center (Tight Center) loss, an enhancement of the standard center loss that incorporates L2-normalization to optimize for angular margins. By training convolutional neural networks (CNNs) with a joint Softmax and T-Center loss, the proposed method aligns the training objective with the cosine similarity metric used in deployment. Experimental results on benchmark datasets demonstrate a significant reduction in intra-class variance and improved separability between positive and negative pairs, leading to more robust iris recognition.

Introduction

While iris recognition is a highly effective biometric, it encounters significant hurdles as databases scale. A primary issue is the difficulty in setting a universal "classification threshold"—the value used to determine if two iris images belong to the same individual.

The root of this problem is the lack of sufficient discriminative power in standard deep learning models. While Softmax loss is effective for inter-class separation, it does not inherently minimize intra-class variance. Furthermore, standard Center Loss utilizes Euclidean distance, creating a mismatch with the cosine similarity (an angular metric) typically used to evaluate iris features. This paper addresses these limitations by introducing a loss function that optimizes directly for angular compactness.

Methodology: T-Center Loss

The authors propose T-Center Loss ($L_{tc}$), which refines the traditional center loss approach through three key mechanisms:

  1. L2-Normalization: T-Center loss applies L2-normalization to both feature vectors ($x_i$) and class centers ($c_{y_i}$) prior to distance calculation. This forces optimization onto a hypersphere, aligning the training process with the angular nature of cosine similarity.
  2. Joint Supervision: To maintain global separability while enhancing local compactness, the model is trained using a weighted combination of Softmax and T-Center loss: $$L = L_{softmax} + \frac{\lambda}{2} L_{T-Center}$$
  3. Dynamic Center Updating: Class centers are updated by averaging features within each mini-batch, moderated by a scalar learning rate ($\alpha$) to ensure stability against noise.

Experimental Evaluation

The approach was validated using two prominent benchmark datasets: CASIA-IrisV4 and IITD2.0. The researchers compared three distinct configurations:

  • Model A: Softmax loss only.
  • Model B: Softmax loss combined with standard Center Loss.
  • Model C: Softmax loss combined with the proposed T-Center Loss.

Results and Discussion

The implementation of T-Center loss yielded substantial improvements in feature discriminability.

Variance and Separability

Model C achieved a 43.8% reduction in intra-class variance compared to Model A, and a 30.2% reduction compared to Model B. This increased compactness is critical for establishing reliable thresholds in large-scale systems.

Distribution of cosine similarity scores for positive and negative pairs across Model A (left), Model B (middle), and Model C (right). Model C shows the clearest separation.

As shown in the similarity distributions, T-Center loss (Model C) creates a significantly wider gap between the scores of positive pairs (same identity) and negative pairs (different identities). This "gap" is essential for minimizing both False Acceptance Rates (FAR) and False Rejection Rates (FRR).

Feature Space Visualization

Visualization of the feature space using PCA confirms that T-Center loss produces tighter and more distinct clusters.

3D visualization of iris feature clusters, demonstrating the high intra-class compactness achieved by T-Center loss.

By penalizing the angular distance between samples and their respective class centers, the model learns to ignore non-discriminative variations, such as lighting changes or minor misalignments, focusing instead on the unique patterns of the iris.

Conclusion

T-Center loss provides a robust solution for enhancing the discriminative power of CNNs in iris recognition. By aligning the training objective with the evaluation metric through L2-normalization, the method significantly reduces intra-class variance. While this study focuses on iris recognition, the principles of T-Center loss are likely generalizable to other biometric tasks, such as face or gait recognition, where angular similarity is paramount. Future work may explore the computational efficiency of this approach and its robustness against severe occlusions.

Summary updated

BibTeX
@inproceedings{chen2019t,
  author = {Yifeng Chen and Cheng Wu and Yiming Wang},
  title = {{T-Center: A Novel Discriminative Feature Extraction Approach for Iris Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33019929},
}
Figure