Towards Complete and Accurate Iris Segmentation Using Deep Multi-Task Attention Network for Non-Cooperative Iris Recognition

Created by Caiyong Wang @ Institute of Automation, Chinese Academy of Sciences (CASIA)

We propose a high-efficiency deep learning based iris segmentation approach, named IrisParseNet. The proposed approach first applies a multi-task attention network to simultaneously predict the iris mask, pupil mask and iris outer boundary. Then, based on the predicted pupil mask and iris outer boundary, parameterized inner and outer iris boundaries are achieved by a simple yet effective post-processing method. Overall, the proposed approach is a complete iris segmentation solution, i.e., iris mask and parameterized inner and outer iris boundaries are jointly achieved, which facilitates the subsequent iris normalization as well as iris feature extraction and matching. Hence the proposed approach can be used for iris recognition as a general drop-in replacement. To help reproduce our method, we have made the models, manual annotations and evaluation protocol codes freely available to the community.

In the above figure, (a) original iris images from CASIA.v4-distance (top two), MICHE-I (middle three), and UBIRIS.v2 (bottom three) iris databases, (b) ground truth iris mask (blue), and inner (green) and outer (red) boundaries of the iris, (c) segmentation results of IrisParseNet (false positive error pixel (green), false negative error pixel (red), and true positive pixel (blue)), (d) iris outer boundary predicted by IrisParseNet, (e) pupil mask predicted by IrisParseNet, and (f) localization results of IrisParseNet after post-processing (inner boundary (green) and outer boundary (red)) are shown from left to right.

Citation

If you use this model or corresponding codes/datas for your research, please cite our papers.

@article{wang2020iris,
  title={Towards Complete and Accurate Iris Segmentation Using Deep Multi-Task Attention Network for Non-Cooperative Iris Recognition}, 
  author={Wang, Caiyong and Muhammad, Jawad and Wang, Yunlong and He, Zhaofeng and Sun, Zhenan},
  journal={IEEE Transactions on Information Forensics and Security}, 
  year={2020},
  volume={15},
  pages={2944-2959},
  publisher={IEEE}
}

Anyone is permitted to use, distribute and change this program for any non-commercial usage. However each of such usage and/or publication must include above citation of this paper.

Prerequisites

linux
Python 2.7 ( python3 is not supported!)
CPU or NVIDIA GPU + CUDA CuDNN
Caffe
matlab R2016a
Halcon 10.0/13.0 or above

Getting Started

Installing Caffe

We have provided the complete Caffe codes. Just install it following the official guide. You can also refer to our another extended Caffe Version.

Caffe [Google Drive] [Baidu Drive](zuhv)

Model training and testing

The complete codes for training and testing the model are placed at Codes/IrisParseNet, and the post-processing executable program is placed at Codes/Post-processing.

We have released the trained models on CASIA.v4-distance (casia for short), MICHE-I (miche for short), and UBIRIS.v2(nice for short).

CASIA.v4-distance model [OneDrive] [Baidu Drive](yavp)
MICHE-I model [OneDrive] [Baidu Drive](zpi4)
UBIRIS.v2 model [OneDrive] [Baidu Drive](376p)
VGG_ILSVRC_16_layers.caffemodel [OneDrive] [Baidu Drive](7ncn)

Evaluation protocols

The iris segmentation and localization evaluation codes are provided. During realizing the evaluation protocols, we've referenced a lot of open source codes. Here, we thank them, especially USIT Iris Toolkit v2, TVM-iris segmentation, GlaS Challenge Contest.

Please read our paper for detailes. The evaluation codes can be found in evaluation folder.

Annotation codes

we use the interactive development environment (HDevelop) provided by the machine vision software, i.e. MVTec Halcon. Before labeling, you need to install Halcon software. Halcon is a paid software, but it allows to try out for free, please refer to the page: https://www.mvtec.com/products/halcon/now/.

Our halcon based annotation codes can be found in annotation. The code can help us to label iris inner/outer bounadry and output a variety of kinds of annatation results as much as possible.

Data

We have provided all training and testing datasets with ground truths to help reproduce our method. Since we do not have permission to release the original iris images for MICHE-I and UBIRIS.v2 databases, hence if you want to use the ground truths of these two databases, you can email the owners of both databases to request permission and let us know if given permission. We will provided the password of ground truth files.

Original iris database:

MICHE-I Email: biplab@unisa.it.
UBIRIS.v2 Email: hugomcp@di.ubi.pt.

Ground truth:

CASIA.v4-distance [OneDrive] [Baidu Drive](kdd6)
MICHE-I [OneDrive] [Baidu Drive](84ya)
UBIRIS.v2 [OneDrive] [Baidu Drive](iudc)

Reference

[1] Zhao, Zijing, and Ajay Kumar. "An Accurate Iris Segmentation Framework Under Relaxed Imaging Constraints Using Total Variation Model." international conference on computer vision (2015): 3828-3836.

[2] Liu, Nianfeng, et al. "Accurate iris segmentation in non-cooperative environments using fully convolutional networks." Biometrics (ICB), 2016 International Conference on. IEEE, 2016.

[3] Hu, Yang, Konstantinos Sirlantzis, and Gareth Howells. "Improving colour iris segmentation using a model selection technique." Pattern Recognition Letters 57 (2015): 24-32.

[4] Proença H, Alexandre L A. The NICE. I: noisy iris challenge evaluation-part I[C]//2007 First IEEE International Conference on Biometrics: Theory, Applications, and Systems. IEEE, 2007: 1-4.

[5] De Marsico M, Nappi M, Riccio D, et al. Mobile iris challenge evaluation (MICHE)-I, biometric iris dataset and protocols[J]. Pattern Recognition Letters, 2015, 57: 17-23.

Disclaimer

This package is only provided on "as it is" basis and does not include any warranty of any kind.

Questions

Please contact caiyong.wang@cripac.ia.ac.cn

polyuYangZhongqi/IrisParseNet