OcclusionNetPlusPlus: a multi-scale similarity network with adaptive occlusion detection for robust iris recognition

Abstract

A significant challenge in iris recognition systems is the presence of occlusions affecting the iris, face, and periocular regions. To address this issue, this study proposes an OcclusionNetPlusPlus framework which employs carefully designed bank of Gabor filters to capture iris texture patterns at different scales and orientations. We then inject 2D positional encodings into these filter responses to embed explicit (x,y) location information, enabling downstream modules to reason about where each feature came from. The innovation in our approach is the introduction of an occlusion detection mechanism that generates probability maps based on local variance analysis, effectively identifying occluded regions in the iris image. These probability maps are used to dynamically weight the extracted features, reducing the influence of unreliable regions during similarity computation. The framework incorporates a custom loss function that optimizes feature similarity while maintaining discriminative power across different iris patterns. Training and evaluation were conducted on publicly available iris recognition datasets, ensuring a diverse test bed for assessing performance across different occlusion scenarios. We evaluated OcclusionNetPlusPlus on CASIA-Iris-Thousand and IIT Delhi V1.0. In controlled tests, it achieves an EER of 0.51 %, an FRR of 0.54 % at FAR = 1 % (0.61 % at FAR = 0.1 %), and a d-prime of 7.04. Even under simulated unconstrained conditions—adding noise, blur, and random occlusions—EER stays around 2 %.