February 12, 2025

The use of biometric recognition is expanding in application fields such as payment and smartphone authentication. Nevertheless, for applications that require advanced security, it tends to be combined with other authentication schemes such as one-time passwords. In contrast, face and iris multimodal biometric recognition has the advantage of being able to verify identity by both face and right/left iris recognition at the same time. NEC’s new development is a technology that further simplifies this recognition. We spoke with our researchers about the details of this technology.

High-speed recognition of 10 million to 100 million users

―Tell us about the innovative points and advantages of the new face and iris multimodal biometric authentication.

Ito: At NEC, we developed a face and iris multimodal biometric recognition that combines NEC’s world number one ranked accurate face recognition and iris recognition.*1 This has been released as an actual service since March 2023. The greatest advantage of face and iris multimodal biometric recognition is its high accuracy. By combining the three elements, face recognition, right-eye iris recognition, and left-eye iris recognition, it delivers a super-high accuracy recognition that can cover as many as 10 billion users. In order to ensure higher accuracy, face recognition tends to be used in combination with other elements such as ID cards and PIN codes. In comparison, face and iris multimodal biometric authentication can smoothly achieve multi-element recognition by collecting three pieces of information from the face: overall face and right and left irises.

On the other hand, the issue with face and iris multimodal biometric recognition was large size in terms of equipment. Aside from needing a separate camera for each of face recognition and iris recognition, a tilt mechanism to move the lens upward and downward is necessary for pin-point, high-precision image capture around the iris. This upscales the size of the entire equipment, limiting where it can be installed and how it can be used. Therefore, application was only sought in limited implementations, such as that allows for placement of large systems while requiring exceptionally high accuracy, examples being immigration check and citizen ID operations.

The most notable point in this newly developed scheme is that it downsizes the equipment and expands the range of applications. It can also be connected to existing systems, including customers’ computers and tablet devices, to implement high-speed face and iris multimodal biometric authentication that covers 10 million to 100 million users. We are anticipating application in various settings without the limitation of installation location, for example, incorporating into ATMs and POS registers, and identity verification at outdoor gates requiring higher security.

Achieving compactness by using only one camera

―What led to the smaller size of the equipment?

Ito: We consolidated the cameras for face recognition and iris recognition into one. Iris recognition uses the iris parts taken out of the image used for face recognition.


Toizumi: In the field of iris recognition, it has been thought that the iris must be captured with a high resolution, from 160 to 200 pixels in 1 cm iris, for iris recognition. This belief expands system size and costs, driven by the need for specialized cameras and mechanisms specifically designed to capture the iris region. To avoid these problems, our team initially explored the possibility of iris recognition for low-resolution, less quality iris images.

Shoji: We started out with using AI to process low-quality iris image signals cut out from images for face recognition in an attempt to upgrade the quality. Improving image quality is a technology area that is widely studied outside of our research theme. While we were able to confirm a certain level of performance, we hit the problem of the time taken for image processing. As we were aiming to realize fast recognition processing,  this method was not practical―therefore, we searched for another way. We then set our eyes on a method called “knowledge distillation.”


Toizumi: In general, knowledge distillation is a deep learning technique used to create a faster lightweight model from the large-scale model that achieves higher performance but are typically slower. We applied this method to low-resolution iris recognition. 


Shoji: Upgrading the quality of the images themselves can take more than a couple seconds to process. However, what we are trying to develop is authentication―this does not need high-quality image signals, which slow down processing. We only need features with higher quality for recognition. Once we realized that, we attempt this approach to extract features with higher quality.


Toizumi: The adoption of this method enables high-speed processing, and it eliminates waiting time for users. Generally, increasing robustness for lower-quality images compromises accuracy for higher-quality images. On the other hand, the distillation method achieves the robustness for lower-quality images with maintaining its performance for higher-quality images. As a result, we achieved a scalable structure that can utilize the most accurate iris recognition engine. 


Shoji: We presented the algorithm of this technology at the IEEE International Joint Conference on Biometrics (IJCB), the world's largest international conference on biometric technologies, in the past September and had a great response.

Lightweight processing by using only the necessary and sufficient data

―Are there any other technical key points?

Toizumi: In addition to the low-resolution iris recognition engine, optimizing the entire camera system for multimodal recognition is also a key aspect of improvement. 


Sasaki: Once we got to a point where we briefly understood what level of resolution is needed on the engine side for authentication , we selected image sensors and lenses to meet those requirements and configured the camera system.  What was the most problematic was the data volume of images during communication. While we can now do with lower resolutions for iris recognition, we still need to capture the face with a larger image compared to that for face recognition. This increases the image data volume transferred from the camera.  Since transmitting all the data slows down the process, only the necessary and sufficient data for iris recognition is extracted and transmitted, further speeding up the process.  To be more specific, the system quickly and accurately finds features from images and sends only the necessary data to NEC’s detection engine. This process makes it possible to increase the frame rate of input data reaching the engine.  As a result, we built an architecture that can perform high-speed, high-accuracy multimodal authentication with just generally available parts.


Ito: We also added a camera control mechanism that optimizes image quality for recognition. Consequently, the recognition system functions robustly even under outside light. Optimizing image quality for recognition in addition to lightweight processing makes possible high-speed, high-accuracy authentication.

Breaking through the norm of iris recognition to complement the missing piece in biometric recognition

―Please tell us about the future prospects and goals for this technology.

Sasaki: While we managed to shape our research into technology, I personally feel that we still have some work to do before commercialization. For example, users may think that “irises may be difficult to capture” or “registration is difficult” or have other concerns. To eliminate such concerns, it is most effective to have people actually experience a demo version. I would like to continue working on developing prototypes and demos so that we can correctly and efficiently communicate the technology we made as a team.


Shoji: I want to further lift the restrictions in recognition so that the technology is available and accessible to more people. This time, we were able to break through the restriction of the “need for high-quality images for iris recognition,” but there are still more restrictions out there. I would like to continue research with the aim to creating a technology that enables high-accuracy recognition in a wider range of applications by easing these restrictions.


Toizumi: I am currently developing the core component of NEC's iris recognition engine. Our goal is to further enhance and expand the performance of our world-leading engine. Improving performance with low-quality images has been a significant achievement for us.
As I mentioned earlier, iris recognition has traditionally been believed to require high-quality images to work effectively. The textures of right and left irises are unique, and it does not change over life time. It offers great potential for accuracy in recognition. By removing the previous need for high-quality images, we can expand opportunities for enhanced performance and broader application possibilities. 


Ito: That is exactly the case. Compared to other biometric recognition technologies, iris recognition has been somewhat behind in deployment. Looking at scoreboards made by specialists, while we always get “Excellent” for accuracy, we end up with “Not good” for both cost and availability. Against this backdrop, we have been working on research aiming for the democratization of iris recognition. The new technology drastically improved cost and availability―a significant advance.

I also expect the evolution of iris recognition to be something that fills the missing piece in solving issues in biometric recognition in general. As Mr. Sasaki said, we will materialize possibilities of the key to future digital transformation through prototyping and continue to present how this technology can be useful to society.

Go back to Featured Technologies