The ICB Competition on Cross-sensor Iris Recognition
The rich and unique texture present in the iris regions of human eyes makes iris recognition a very reliable way of identifying individuals. Iris recognition systems have been widely deployed in many mission critical applications such as national identity card, border control, banking, etc. The large market of iris recognition has attracted a number of companies to develop various iris sensors. Therefore it is necessary to match heterogeneous iris images captured by different types of iris sensors with an increasing demand of interoperable identity management systems. The significant differences among multiple types of iris sensors such as optical lens and illumination wavelength determine the cross-sensor variations of iris texture patterns. Therefore it is a challenging problem to develop robust iris image preprocessing, feature extraction and matching methods across different iris sensors. In order to track the state-of-the-art of cross-sensor iris recognition, the ICB Competition on Cross-sensor Iris Recognition (or CSIR2015 shortly) is organized by CASIA (Chinese Academy of Sciences’ Institute of Automation).
CSIR2015 is open to both academia and industry. There are two iris image databases used in CSIR2015 for training and testing purposes respectively. Both the training database CSIR-Train and the testing database CSIR-Test were collected using two iris cameras, i.e. IKEMB-220 produced by IrisKing and EyeGuard AD100 produced by IrisGuard. Each volunteer contributed 20 iris images of each eye for each camera so each iris class has 40 samples. The CSIR-Train is released to public domain for algorithm development. It contains 8,000 iris images of 200 eyes from 100 subjects (4,000 IKEMB-220 images and 4,000 EyeGuard AD100 images). The CSIR-Test is not available to competition participants. It contains 24,000 iris images of 600 eyes from 300 subjects captured by two cameras (12,000 IKEMB-220 images and 12,000 EyeGuard AD100 images). The main sources of intra-class variations in CSIR-Train and CSIR-Test include cross-sensor differences, non-linear deformations, eyeglasses and specular reflections, defocus and so on. Each image in the two databases is an 8-bit gray-level BMP file with a resolution of 640*480.
Fig. 1 Example iris images in the training database. (a) IKEMB-220. (b) EyeGuard AD100.
Fig. 2 Example iris images in the testing database. (a) IKEMB-220. (b) EyeGuard AD100.
All participants should submit two executables "irisenroll_AlgorithmName.exe" and "irismatch_AlgorithmName.exe" in the form of Win64 console applications for third-party performance evaluation. Moreover, a document including a brief introduction of the submitted algorithm is needed for competition summary report. The file "irisenroll_AlgorithmName.exe" is used to generate a feature template from an iris image. The file "irismatch_AlgorithmName.exe" is used to match two iris feature templates. The syntax of the two executables through command-line is as follows.
>irisenroll_AlgorithmName imagefile templatefile outputfile
> irismatch_AlgorithmName templatefile1 templatefile2 outputfile
All possible cross-sensor intra-class comparisons are implemented to evaluate the false non-match rate (FNMR) providing a total of 240,000 cross-sensor intra-class match results. One sample is selected from each iris class for each sensor to evaluate the false match rate (FMR) so there are totally 359,400 inter-class match results of cross-sensor iris recognition. If an intra- or inter-class comparison cannot be successfully implemented due to failure enrollment or failure match, a random variable ranging from 0 to 1 will be assigned as the matching score. Popular performance metrics of iris recognition such as FNMR, FMR, EER and ROC will be reported and the metric F4 (FNMR@ FMR=0.0001) will be used to rank the performance of submitted algorithms.
Each participant can maximally submit three algorithms. We only accept qualified iris recognition algorithms which meet the following requirements due to limited competition resources:
• The equal error rate (EER) must be less than 5% on the training database.
• The average processing time for feature encoding must be less than 3 seconds and the average matching time must be less than 0.1 second on a normal personal computer.
A public platform, the Biometrics Ideal Test (BIT; http://biometrics.idealtest.org) is used to organize the competition. The competition registration process is as follows.
1. Register an account in BIT (http://biometrics.idealtest.org ).
2. Download the training database after your account is approved and activated.
3. Develop an iris recognition algorithm and you can use CSIR-Train or other iris image databases as the training data.
4. Submit your algorithm to firstname.lastname@example.org. You can also send us a web link to download large size files. To help participants quickly master the programming protocols of algorithm evaluation, the source codes of the example files "enroll.c" and "match.c" are provided for download as below. Download C language source code here
5. Any problem of your algorithm during testing will be reported to you so please respond immediately to the organizers of CSIR2015.
6. The CSIR2015 Competition Committee will report the performance of all submitted algorithms and the competition results.
7. Authors of the winner algorithm will be invited to write a paper on their algorithm to be published in the ICB2015 proceedings.
Any questions on CSIR2015 can be addressed by emailing email@example.com.