DF-Iris: Discriminative Texture Discovery for Near-Eye Iris Recognition

Abstract

Iris recognition in unconstrained near-eye scenarios faces significant challenges due to off-axis gaze and occlusions, where traditional geometry-dependent pipelines relying on explicit segmentation become fragile. To address this limitation, we propose DF-Iris, an end-to-end framework that reformulates recognition from a geometric alignment task to a unified process of discriminative texture discovery. Inspired by fine-grained visual categorization, our approach treats identity localization as an implicit, learnable subtask, enabling the network to autonomously discover identity-relevant micro-textures directly from raw images. The framework incorporates a parallel differential convolution backbone for orientation-specific modeling to capture high-frequency annular patterns and a hierarchical attention mechanism for implicit localization to suppress spatially entangled periocular noise by maximizing inter-identity separability. Extensive evaluations on three public benchmarks confirm state-of-the-art performance with a compact model footprint. Specifically, the method achieves equal error rates of 0.72%, 0.73%, and 5.83% on CASIA-IrisV4-Thousand, CASIA-IrisV4-Lamp, and OpenEDS, respectively, demonstrating that discriminative evidence mining provides a robust and efficient alternative to explicit geometric segmentation.