AI-based autonomous imager delivers facial recognition for smartphones

Update: June 11, 2021

CEA-Leti has claimed the industry’s first autonomous imager that activates smartphones and small appliances through facial recognition. This imager combines auto-exposure for all lighting conditions, motion detection, feature extraction for event-based functioning, and artificial intelligence-based object detection.

The autonomous imager, called µWAI (micro-WAY), is as small as a 1€ coin and features a novel readout and processing architecture co-designed with an optimized algorithmic pipeline, which is said to provide ultra-low-power wake-up modes and compact silicon implementation for cost savings.

What is an autonomous imager? It is a new class of image sensor with the tight co-design of autonomous image acquisition and image processing, said Antoine Dupret, CEA-Leti’s industrial partnership manager. “As a matter of fact, it is rather a vision sensor since it analyzes the scene to extract the relevant information.”

The tech research institute said it is the first smart image sensor to combine auto-exposure for all lighting conditions (for accurate recognition in variable conditions) and 88-dB dynamic range, together with motion detection, feature extraction for event-based functioning, and AI-based object recognition that triggers highly reliable identification (with 95% accuracy).

CEA-Leti’s autonomous imager device µWAI (Image : CEA-Leti)

AI-based recognition offers two key benefits. “AI has two important features that are exploited in the autonomous imager. First, AI allows achieving high recognition ratio. It eventually further reduces the power consumption by reducing the number of false alarm wake-ups,” said Dupret. “Second, AI is, to some extent, versatile, i.e., the same algorithm and/or hardware can be used for the recognition of different objects by just changing the ‘weights’, the learned features. The autonomous imager can hence be used to recognize other objects.”

Energy savings

Together, the key features of the imager – auto-exposure for all lighting conditions, motion detection, feature extraction for event-based functioning, and AI-based object recognition – enable highly reliable decision-making for a few tens of pJ/pixel/frame, outperforming existing off-the-shelf systems, according to CEA-Leti.

“The figures are derived from the power consumption, the frame rate, and the resolution of the image sensor,” said Dupret. “Considering a VGA image sensor functioning at 15 fps with a power consumption of 100 mW, the energy per pixel and frame is 100e-3/310e3/15=21 nJ/pixel/frame, without any processing.”

The researcher said a typical implementation using a low-power camera plus a processor requires about 10,000 times more energy than an µWAI imager.

Dupret said a COTS implementation requires at the very minimum one image sensor and one microcontroller. “A typical low-resolution image sensor consumes some tens of milliwatts. It requires a microcontroller to adapt its functioning, e.g., to adjust the exposure time. Then, the image has to be analyzed. All in all, the minimum power budget is in the range of the hundreds of milliwatts.”

CEA-Leti touts 3-6 µW operation for the imager, which meets requirements for IoT applications, and it can work with a button cell that lasts five years.

However, the gain is not so dramatic regarding area or volume, said Dupret. “The autonomous imager is intended to wake-up a more complex system. Hence, other processors are needed to further exploit the images.”

The imager also delivers privacy-compliant AI-based recognition because the images are processed within the image sensor itself. “The µWAI image sensor performs on-the-fly data processing within the image sensor (without any frame memory). Hence the content of the scene is not sent outside of the chip and only a set of features are.”

Applications for the µWAI image sensor  include automatic switching and face identification in mobile devices, contact-less smart switching of household appliances, and sport-and-entertainment devices in smart homes. The smart image sensor also can be used for face recognition, people counting, alarm triggering in smart buildings, vehicle-interior situation awareness, driver identification, parking-situation awareness, and a smart-unlocking system in automobiles.

CEA-Leti’s team is working hand-in-hand with STMicroelectronics to develop specific smart-imager products as they look to extend the technology to other use cases, said Dupret. Currently, CEA-Leti is leveraging STMicroelectronics’ state-of-the art CIS technology and BIS pixel.

“The next application will depend on what our industrial partners will target,” said Dupret. “CEA-Leti works hand-in-hand with industrial partners to develop custom-made innovations.”

The µWAI technology will be introduced at CEA-Leti’s digital event, Leti Innovation Days, June 22 and 23, 2021.