yuanzhuohao / irisparsenet

Joint Iris Segmentation and Localization Using Deep Multi-task Learning Framework

IrisParseNet

Joint Iris Segmentation and Localization Using Deep Multi-task Learning Framework

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

As shown in the figure above, we present a novel deep learning based joint iris segmentation and localization method, named as IrisParseNet. Unlike the former many FCN based iris segmentation methods, this method not only realizes accurate iris sgementation, but also solves the most import problem in the iris preprocessing step, namely iris localization. Besides, this method achieves state-of-the-art performances on various benchmarks. To promote the research on iris preprocessing (especially iris segmentation and iris localization), we have made our ground-truth annotations, annotation codes and evaluation protocols freely available to the community.

Citation

If you use this code or ground-truth annotations for your research, please cite our papers.

@article{casiairis2019,
  title={Joint Iris Segmentation and Localization Using Deep Multi-task Learning Framework},
  author={Caiyong Wang and Yuhao Zhu and Yunfan Liu and Ran He and Zhenan Sun},
  journal={arXiv preprint abs/1901.11195},
  year={2019}
}

Prerequisites

• Halcon 10.0/13.0 or above for labeling iris inner/outer boundary
• matlab R2016a (other versions are ok) for evaluating the performance of the proposed method

Getting Started

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 code/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.

Annotation ground truths

We do not have control of the links to the original databases except for CASIA iris database. Please do a Google or similar search to try to find the database website if the links below do not work.

We provide the following annotation results:

directory name	README
center	pupil center image
center_rough	dilated pupil center image with a circular structuring element of radius 3
inner	iris inner boundary image
Inner_rough	dilated iris inner boundary image with a circular structuring element of radius 3
outer	iris outer boundary image
Outer_rough	dilated iris outer boundary image with a circular structuring element of radius 3
pupil_edge_mask	filled iris inner boundary image
pupil_edge_mask_binary	filled iris inner boundary binary image
iris_edge_mask	filled iris outer boundary image
iris_edge_mask_binary	filled iris outer boundary binary image
circle_params	the parameter of labeled iris inner/outer circular boundary
ellipse_params	the parameter of labeled iris inner/outer elliptical boundary
edge	iris boundary image where iris inner/outer boundary is labeled as 1/2, respectively
label_center	the x-y coordinate and exist status (1: existing 2: not existing, meanwhile x,y=-1) parameter of pupil center position
JPEGImages	original input image
SegmentationClass	segmentation ground truth
SegmentationClass_1D	binary segmentation ground truth

Detailed database description are listed as follows:

• CASIA-Iris-Distance (CASIA)

The CASIA database contains 2576 images from 142 subjects with resolution of 2352×1728 pixels captured under near infrared illumination (NIR). Here, we use a subset introduced in paper[2] (totalling 300 iris images for training, 100 iris images for testing), for which iris segmentation and localization groundtruths is available. Besides, all images are resized to 640×480 pixels.

Link to the original database and ground truth: [Google Drive] [Baidu Drive]

• UBIRIS.v2 (UBIRIS)

The UBIRIS database contains 11102 images from 261 subjects which are acquired in visible light illumination (VIS). Here, we use a subset introduced in NICE.I competition (totalling 500 iris images for training, 445 iris images for testing), for which iris segmentation and localization groundtruths is available. Besides, all images are resized to 400×300 pixels.

Link to the original database: click here

Ground truth segmentation files: click here

Ground truth localization files: [Google Drive] [Baidu Drive]

• MICHE-I (MICHE)

The MICHE images were captured by three mobile devices including iPhone5 (abbreviated IP5, 1262 images), Samsung Galaxy S4 (abbreviated GS4, 1297 images), and Samsung Galaxy Tab2 (abbreviated GT2, 632 images) in uncontrolled conditions with visible light illumination (VIS). Here, we use a subset introduced in paper[3] (totalling 140 iris images for training, 429 iris images for testing), for which iris segmentation and localization groundtruths is available. Besides, the width of all iris images is resized to 400 and height is resized to maintain the same proportions as the original image. Finally, the size of resized image is approximately 400x400.

Link to the original database: click here

Ground truth segmentation files: Please contact [email protected] or [email protected] (the author of paper[3]).

Ground truth localization files: [Google Drive] [Baidu Drive]

Note that because the boundary of some images is very difficult to label, we deleted these images. Hence the dataset finally used in the experiment is summaried as follows:

dataset	CASIA	UBIRIS	MICHE
No. of training set	296	498	140
No. of testing set	99	445	429

In order to access the annotation files, it must be sent an email to Dr. Wang([email protected]) asking for the password of the zip file.

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 code/evaluation.

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.

Questions

Please contact [email protected]