MCUs are Changing the Face Recognition Game
May 20, 2020
Contrary to what you might expect, bringing facial recognition capabilities to computers doesn?t require high-end hardware.
Face Recognition Is Becoming More Cost-Effective
Contrary to what you might expect, bringing facial recognition capabilities to computers doesn’t require high-end hardware. Exploiting computing power has come a long way since the mid-2000s when deep-learning pipelines began demonstrating legitimate image classification with power-intensive GPUs. But today’s embedded developers can design pipelines for efficiency and prioritize just those features most important to bringing functionality to their applications, reducing the processing power needed.
It’s Becoming Easier
Platforms are available today that abstract away a lot of the complexity of machine learning. Building structures from the ground up in facial recognition applications is tedious, and thanks to new development tools, unnecessary. Platforms built on established machine-learning pipelines can very quickly bring high-performance while at the same time provide the required level of customization needed to address different markets.
Access control is one of the biggest use cases for facial recognition in embedded systems. Due to security often being at stake, the system needs to be robust enough to not be fooled by someone holding a photograph up to a camera. Integrated vision with machine learning is important here because they can perform checks on the image to make sure viable data makes its way into the pipeline. This also brings flexibility for the pipeline to account for more than just what’s visible - using ultra-violet sensors, for example, can further increase security by distinguishing reality from a photograph or screenshot.
The Data is Safe
A big concern for users is that most facial recognition applications that we come into contact within our day-to-day lives process the collected data in the cloud. Nobody wants their movements and activities beamed across the Internet and possibly exploited. But there are platforms that perform image processing locally. NXP’s MCU-based EdgeReady platform is an off-the-shelf IoT edge compute solution that performs all analysis and facial recognition locally. This means developers can design products with an eye toward privacy.
It Works in the Dark
Like we mentioned earlier, security is one of the primary applications for facial recognition. So being able to protect your assets or access to protected areas around the clock is important. Technology based on image acquisition might seem limited to operating only when visible light is available, nighttime coverage can be accomplished by augmenting visible light image sensors with secondary devices that work on the infrared spectrum or use time-of-flight data to build 3D maps of objects within range. Doing this eliminates darkness as an obstacle.
It Can be Done with a Lightweight OS
Embedded commercial systems don’t have the same requirements as research-level tools for deep learning, which are often offered as open-source toolkits written for Linux. But MCU-based solutions don’t bear the heavy memory toll or long boot times of Linux installations and can run with much lighter operating systems that use less memory and power.
It’s Power Optimized
Advancements in AI and image processing means facial recognition can be managed on a low-power MCU instead of power-hungry GPUs. Plus, an MCU-based approach comes with the added benefits of multiple power-saving modes supported by most current MCUs. Since an MCU doesn’t need to boot large operating systems like Linux, the main processor can be shut down when it’s not needed. But you can still wake the processor for full capability in a fraction of a second if a sensor determines the necessary level of activity in the field of view.
Teaching Devices Can be Quick
Training early implementations of face recognition in embedded systems, such as tablets and smartphones, required users to strike a series of different poses in order for the neural network to effectively train on a new user’s face. Transfer learning and other techniques are making it possible to show their face to the camera just one once in order to and add them to the approved-user list.
Face Recognitions Applications Are Expanding
More and more devices, including consumer oriented IoT smart things, will be designed with facial recognition as a core feature. What’s more, devices will be able to not only to distinguish faces but discern expressions as well. Devices will be able to read emotional cues like delight, frustration, and anger and potentially respond accordingly.
About the Author
Rick Bye leads product marketing for NXP’s ML/AI edge compute IoT solutions for voice control (Alexa, local commands, etc.) and face recognition. Rick focuses on delivering these capabilities as turnkey solutions that eliminate development barriers and significantly reduce time to market.