Author:
Khiarak Jalil Nourmohammadi,Oskouei Amin Golzari,Nasab Samaneh Salehi,Jaryani Farhang,Moafinejad Seyed Naeim,Pourmohamad Rana,Amini Yasin,Noshad Morteza
Abstract
AbstractIris segmentation and localization in unconstrained environments are challenging due to long distances, illumination variations, limited user cooperation, and moving subjects. Some existing methods in the literature have somehow mitigated some of the above-mentioned issues. In this paper, motivated by these weaknesses, we propose a framework that employs a deep neural network-based approach to iris segmentation and localization. The proposed framework is based on a U-Net architecture initialized with a pre-trained MobileNetV2 model. In addition, to better study the detectors in iris recognition scenarios, we have collected 1000 images. The provided dataset (KartalOl) is made publicly available for the research community. In the proposed framework, to have better generalization, we fine-tuned the MobileNetV2 model on the provided data for NIR-ISL 2021 from the CASIA-Iris-Asia, CASIA-Iris-M1, and CASIA-Iris-Africa and our dataset. Likewise, data augmentation techniques are applied on images. We chose the binarization threshold for the binary masks by iterating over the images in the provided dataset. The proposed framework is trained and tested in CASIA-Iris-Asia, CASIA-Iris-M1, and CASIA-Iris-Africa, along with the KartalOl dataset. The experimental results highlight that our method surpasses state-of-the-art methods on mobile-based benchmarks. The implementation source code of KartalOl is made publicly available at https://github.com/Jalilnkh/KartalOl-NIR-ISL2021031301.
Funder
European Union’s Horizon 2020 research and innovation program
Publisher
Springer Science and Business Media LLC
Subject
General Earth and Planetary Sciences,General Environmental Science