Abstract
Iris restoration from complexly degraded iris images, aiming to improve iris recognition performance, is a challenging problem. Due to the complex degradation, directly training a convolutional neural network (CNN) without prior cannot yield satisfactory results. In this work, we propose a generative iris prior embedded Transformer model (Gformer), in which we build a hierarchical encoder-decoder network employing Transformer block and generative iris prior. First, we tame Transformer blocks to model long-range dependencies in target images. Second, we pretrain an iris generative adversarial network (GAN) to obtain the rich iris prior, and incorporate it into the iris restoration process with our iris feature modulator. Our experiments demonstrate that the proposed Gformer outperforms state-of-the-art methods. Besides, iris recognition performance has been significantly improved after applying Gformer.
Introduction
Iris recognition is a highly secure biometric, but its real-world performance is often hampered by poor image quality. Factors like long-distance capture (low resolution), out-of-focus subjects (defocus blur), and subject movement (motion blur) create "complexly degraded" images. Gformer is a new hierarchical transformer-based model designed to restore these low-quality iris images by embedding a pretrained generative "iris prior." By combining the long-range modeling capabilities of Transformers with the rich structural knowledge of a StyleGANv2 generator, the system significantly improves both visual image quality and the accuracy of downstream iris recognition systems.
The Challenge of Iris Restoration
Standard Convolutional Neural Networks (CNNs) often struggle with iris restoration for two primary reasons:
- Local Receptive Fields: CNNs are excellent at capturing local patterns but struggle to model long-range dependencies across an image, which is necessary for aligning global iris structures.
- Lack of Prior Knowledge: When an image is severely degraded, there isn't enough information left to reconstruct fine iris microstructures (like filaments and crypts) without a "template" of what a healthy iris should look like.
Methodology: The Gformer Architecture
The authors introduce Gformer, a U-shaped encoder-decoder architecture that bridges the gap between Transformer flexibility and GAN-based generative power.
The Encoder (Transformer-based)
To avoid the quadratic computational explosion of standard Transformers on high-resolution images, Gformer uses Depth-Wise Self-Attention (DSA). This mechanism calculates attention across the channel dimension rather than the spatial dimension, keeping the complexity linear and practical for restoration. This allows the model to capture global context essential for aligning iris structures without massive memory overhead.
The Decoder (Generative Prior)
The decoder is built using the backbone of a pretrained StyleGANv2 generator. This "generative iris prior" provides the model with a deep understanding of iris anatomy, allowing it to "fill in the blanks" for missing textures. Using a pretrained GAN ensures the restored microstructures look like real iris features rather than random noise.
Iris Feature Modulator
This component acts as the "glue" between the encoder and decoder. It uses Channel-Split Spatial Feature Transform (CS-SFT) to fuse features. By only modulating half of the feature channels, the model preserves the identity information from the input while injecting high-quality textures from the generative prior. An MLP extracts a latent code ($F_{latent}$) from the encoder to ensure the semantic identity of the subject is preserved throughout the decoding process.
Experimental Setup
The researchers used two major datasets from the Institute of Automation, Chinese Academy of Sciences (CASIA):
- CASIA-Iris-Lamp: 822 images from 411 subjects.
- CASIA-Iris-Thousand: 20,000 images from 1,000 subjects.
They combined and processed these into a dataset of 28,089 iris images across 2,738 categories. To simulate real-world conditions, they applied a synthetic degradation model involving Gaussian blur, motion blur, and downsampling.
Key Results
Gformer outperformed several state-of-the-art (SOTA) methods, including EDSR, RCAN, and DeblurGANv2, particularly in recognition-focused metrics:
- Recognition Accuracy: On the True Acceptance Rate (TAR) at a False Acceptance Rate (FAR) of 0.001, Gformer achieved 0.8074, a massive jump from the degraded input's 0.1992 and significantly higher than the next best SOTA (HiFaceGAN at 0.7549).
- Error Rates: It achieved the lowest Equal Error Rate (EER) of 0.0455.
- Visual Quality: Qualitatively, Gformer produced much sharper and more realistic iris textures (crypts and furrows) compared to the "blurry" or "over-smoothed" results of pure CNN approaches.
Conclusion and Future Work
While Gformer sets a new benchmark for iris restoration, it faces some limitations. With an inference time of 146ms, it is significantly slower than simpler models like SRGAN (10ms), which might limit its use in real-time mobile authentication. Additionally, the model was trained on synthetic degradations; moving to real-world sensor noise and lighting conditions remains a challenge. Future work will focus on optimizing the balance between complexity and performance to make this architecture more suitable for edge devices.
Summary updated