【摘要】 虹膜纹理容易被纹理隐形眼镜隐藏甚至伪造，进而对虹膜识别系统的安全性构成了威胁。针对真实虹膜与纹理隐形眼镜伪造虹膜光学特性和纹理特征差异较小的问题，提出了一种循环注意力隐形眼镜虹膜防伪检测方法 RAINet。利用循环注意力机制对能区分真伪虹膜的关键区域进行无监督定位，并通过多层级特征融合提升防伪检测精度，构建了端到端防伪检测网络，无需图像预处理即可直接进行真伪特征检测。采用MobileNetV2作为特征分类网络，在保持检测精度的同时，减少了网络的参数量和计算量。在包含真实虹膜样本和隐形眼镜虹膜样本的两个公开数据库（IIITD CLI和ND系列）上进行了实验验证。结果表明，RAINet的检测精度优于其他防伪检测网络，在同传感器、跨传感器和跨数据库实验条件下的平均正确分类率分别可达到99.93%、97.31%和97.86%。更多还原

【Abstract】 Iris textures are easily hidden or even forged by textured contact lenses, which further threatens the security of the iris recognition system. Considering the tiny differences in the optical properties and texture features of authentic irises and irises forged by textured contact lenses, this paper proposes an anti-spoofing detection method for contact lens irises based on recurrent attention, namely recurrent attention iris net(RAINet). Specifically, the recurrent attention mechanism is employed to locate the key regions that can be used to distinguish authentic irises from forged ones in an unsupervised manner, and multi-level feature fusion is applied to improve the anti-spoofing detection accuracy. An end-to-end antispoofing detection network is built for the direct detection of authentic and forged features without image pre-processing.MobileNetV2 is used as the feature classification network to reduce the number of parameters and amount of computation of the network in addition to maintaining the detection accuracy. Experimental verification is performed on two public databases(IIITD CLI and ND series) containing both authentic iris samples and contact lens iris samples. The results show that the proposed RAINet outperforms other anti-spoofing detection networks in detection accuracy. Its average correct classification rates under intra-sensor, inter-sensor, and inter-database experimental conditions reach 99. 93%,97. 31%, and 97. 86%, respectively.更多还原