End-to-End Anti-Attack Iris Location Based on Lightweight Network
Hongming Peng, Bingbing Li, Di He, Wang Junning · 2020
With the widespread application of iris recognition in high security fields such as public safety, banking, and mobile payment, robust iris location and pseudo iris attack detection have become hotspots in recent years. How to achieve high accuracy in iris location, prevent the pseudo iris attack and guarantee both the real-time implementation and robustness of the algorithm is an open problem. This paper proposes a lightweight anti-attack iris location algorithm named Lite Anti-attack Iris Location Network (LAILNet). This algorithm can filter out pseudo-iris attacks and achieve robust and highly accurate iris location. LAILNet's feature extraction network adopts dense connection mechanisms and depthwise separable convolutions, which can reuse the features and speed up the network training with the greatly reducing amount of network parameters and calculations. The multi-scale iris location network utilizes six kinds of scales to locate different sizes of irises, improving the accuracy of small-sized iris location. In order to better verify the effectiveness of the algorithm, the continuous infrared iris database IPITRT is constructed in this paper. Real iris data in this database contains extremely disturbing samples with such as far and near focal lengths, blur, changes in ambient lighting, reflected light spots, head movements and blinks. The IPITRT's pseudo iris dataset is obtained by the high-resolution printing and photographing of the real iris data, which has a high similarity with the real iris data. The results show that the average false positive rate of the LAILNet algorithm on the three databases (IPITRT, CASIA-Iris-V4.0 and CASIA-Iris-Fake) is only 0.145%, the average iris positioning accuracy is 99.87%, the model size is 1.3MB, and single frame processing time is only 1.6ms. The LAILNet algorithm can effectively filter out false iris attacks while maintaining high accuracy and high robustness in the iris location. From the perspectives of model size and processing time, the RLAILNet algorithm is lightweight and achieves real-time performance.