Adaptive Deep Iris Feature Extractor at Arbitrary Resolutions

TL;DR

The paper tackles iris recognition under arbitrary image resolutions by introducing a resolution-adaptive feature extractor built from a gating module, three resolution-expert modules (HR, MR, LR), and a shared output module. Through distillation, the LR experts align their features with the HR expert to preserve identity information while enhancing robustness to down-sampling and blurring, all with minimal runtime overhead. The approach outperforms traditional HR-trained and restoration-based methods across multiple datasets and generalizes to unseen degradations, enabling practical iris recognition in less constrained environments. The framework is adaptable to existing backbone models and offers a favorable balance between recognition accuracy at LR and computational efficiency for real-world deployment.

Abstract

This paper proposes a deep feature extractor for iris recognition at arbitrary resolutions. Resolution degradation reduces the recognition performance of deep learning models trained by high-resolution images. Using various-resolution images for training can improve the model's robustness while sacrificing recognition performance for high-resolution images. To achieve higher recognition performance at various resolutions, we propose a method of resolution-adaptive feature extraction with automatically switching networks. Our framework includes resolution expert modules specialized for different resolution degradations, including down-sampling and out-of-focus blurring. The framework automatically switches them depending on the degradation condition of an input image. Lower-resolution experts are trained by knowledge-distillation from the high-resolution expert in such a manner that both experts can extract common identity features. We applied our framework to three conventional neural network models. The experimental results show that our method enhances the recognition performance at low-resolution in the conventional methods and also maintains their performance at high-resolution.

Figures (10)

• Figure 1: Proposed iris recognition method using automatically switching networks for iris images with arbitrary resolutions.
• Figure 2: Architecture difference between conventional baseline and proposed models. The proposed feature extractor has resolution-expert modules (REMs) on the input side and shared module on the output side. Each expert in REMs is trained for different ranges of resolution. Our gating module selects a module in REMs for feature extraction from a resolution condition of an input image.
• Figure 3: Best-performing module labels for each degradation condition (upper) and examples of degraded iris images (lower).
• Figure 4: Training scheme for proposed lower-resolution modules. First, HR module and shared modules are trained for iris recognition, Afterward, lower-resolution module is trained with four losses while parameters of HR and shared modules are fixed.
• Figure 5: Results (EER) of ablation study using CASIA-Iris-Thousand for split points between REMs and shared module. Bold and underlined mean the best and second, respectively.