IEEE Consumer Electronics Magazine - March 2018 - 61

Network 1

0.0883
0.0643
0.0558
0.0444

0.0034
0.0018
0.0018
0.0019

0.0572
0.0308
0.0315
0.0328
FDR

FPR
Network 2

FNR
Network 3

SPDNN

FIGURE 9. The comparisons for all of the networks. (Measurements closer to 0 indicate better performance.) FDR: false discovery rate; FPR: false positive rate.

CONCLUSION
From the training and testing results, we can see that our
methodology helped the merged network to achieve an optimal design and increased the convergence in the training stage
and in the test phase. In more extensive studies on heterogenous data sets, we have shown that this methodology can 1)
generalize the merged network beyond a trivial combination
of parent networks and 2) achieve a significant reduction in
size and computational scale for the merged network [3]. In
other words, the merged network can be significantly smaller,
often a size similar to the largest of the individual parent networks, and perform better on a broader selection of input data
than a simple combination of the original networks. In effect,
the merged network learns a more optimal solution that combines elements of each of the individual parent deep networks.
If you are interested in applying these techniques to your own
deep-learning problems, we encourage you to contact us and
explore the potential for collaboration.

ACKNOWLEDGMENTS
This research was funded under the Science Foundation Ireland (SFI) Strategic Partnership Program by SFI and FotoNation Ltd., project 13/SPP/I2868 on "Next Generation Imaging
for Smartphone and Embedded Platforms." We gratefully
acknowledge the support of NVIDIA Corp. with the donation
of a Titan X GPU used for this research. Portions of the
research in this article use the CASIA-IrisV4 collected by the
Chinese Academy of Sciences' Institute of Automation.

ABOUT THE AUTHORS
Shabab Bazrafkan (s.bazrafkan1@nuigalway.ie) earned his
B.Sc. degree from Urmia University, Iran, in electrical engineering in 2011 and his M.Sc. degree from the Shiraz University of Technology, Iran, in telecommunications engineering,
image-processing branch, in 2013. Currently, he is a Ph.D.
student at the Center for Cognitive, Connected, and Computational Imaging, College of Engineering and Informatics,
National University of Ireland, Galway, and also works with
FotoNation Ltd. His research interests are in the fields of
deep neural networks and neural network design.
Peter Corcoran (dr.peter.corcoran@ieee.org) has been a
university professor for 28 years and was vice dean of research

and graduate studies for seven years in the College of Engineering and Informatics, National University of Ireland, Galway. He is the co-inventor of more than 300 granted U.S.
patents and has been a member of the IEEE Consumer Electronics Society for more than 20 years. He is the emeritus
editor-in-chief and founding editor of IEEE Consumer Electronics Magazine, an industry consultant and expert witness,
and the cofounder of several startup companies, including
FotoNation Ltd. He is a Fellow of the IEEE.

REFERENCES
[1] V. Vinge, "Signs of the singularity," IEEE Spectr., vol. 45, no. 6, pp.
76-82, 2008.
[2] J. Lemley, S. Bazrafkan, and P. Corcoran, "Deep learning for consumer devices and services: Pushing the limits for machine learning,
artificial intelligence, and computer vision," IEEE Consum. Electron.
Mag., vol. 6, no. 2, pp. 48-56, 2017.
[3] S. Bazrafkan, H. Javidnia, J. Lemley, and P. Corcoran. (2017). Depth
from monocular images using a semi-parallel deep neural network
(SPDNN) hybrid architecture. arXiv. [Online]. Available: https://arxiv
.org/abs/0811.1674v2
[4] S. Thavalengal and P. Corcoran, "Iris recognition on consumer devices: Challenges and progress," in Proc. IEEE Int. Symp. Technology and
Society (ISTAS), 2015, pp. 1-4.
[5] S. Thavalengal, P. Bigioi, and P. Corcoran, "Iris authentication in
handheld devices: Considerations for constraint-free acquisition," IEEE
Trans. Consum. Electron., vol. 61, no. 2, pp. 245-253, May 2015.
[6] S. Thavalengal, I. Andorko, A. Drimbarean, P. Bigioi, and P. Corcoran, "Proof-of-concept and evaluation of a dual function visible/NIR camera for iris authentication in smartphones," IEEE Trans. Consum.
Electron., vol. 61, no. 2, pp. 137-143, May 2015.
[7] S. Thavalengal, P. Bigioi, and P. Corcoran, "Evaluation of combined
visible/NIR camera for iris authentication on smartphones," in Proc.
IEEE Conf. Computer Vision and Pattern Recognition Workshops
(CVPRW), 2015, pp. 42-49.
[8] E. Wambui, "Samsung drives multiple modes of mobile biometrics
but suffers recall setback," Biometric Technol. Today, vol. 2016, no. 10,
pp. 1-2, 2016.
[9] C. Rathgeb, A. Uhl, and P. Wild, Iris Biometrics: From Segmentation
to Template Security. New York: Springer, 2013.
[10] J. R. Matey, R. Broussard, and L. Kennell, "Iris image segmentation
and sub-optimal images," Image Vis. Comput., vol. 28, no. 2, pp. 215-222,
Feb. 2010.
[11] S. Thavalengal, P. Bigioi, and P. Corcoran, "Efficient segmentation
for multi-frame iris acquisition on smartphones," in Proc. IEEE Int.
Conf. Consumer Electronics (ICCE), 2016, pp. 202-203.
[12] Chinese Academy of Sciences, Institute of Automation. (2010). Biometrics ideal test. [Online]. Available: http://biometrics.idealtest.org/
[13] S. Bazrafkan and P. Corcoran, "Enhancing iris authentication on
handheld devices using deep learning derived segmentation techniques,"
in Proc. IEEE Int. Conf. Consumer Electronics (ICCE), 2018.
[14] S. Bazrafkan. (2017). DB augmentation. GitHub. [Online]. Available: https://github.com/C3Imaging/Deep-Learning-Techniques/tree/Iris_
SegNet/DBaugmentation

MARCH 2018

IEEE Consumer Electronics Magazine

https://arxiv.org/abs/0811.1674v2 https://arxiv.org/abs/0811.1674v2 http://biometrics.idealtest.org/ https://www.github.com/C3Imaging/Deep-Learning-Techniques/tree/Iris_SegNet/DBaugmentation https://www.github.com/C3Imaging/Deep-Learning-Techniques/tree/Iris_SegNet/DBaugmentation

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - March 2018