[video_player id=鈥漨uo6aqdS鈥� access_level=鈥漞veryone鈥漖
A sensor that mimics the way a human eye responds to light could make digital cameras more efficient. This could pave the way to faster, more efficient machine vision for things like self-driving cars and robots.
Some modern cameras are built around charge-coupled devices (CCDs), which produce a voltage when light falls on them. Continuous light produces a continuous signal. By contrast, retinal cells produce a spike only when first lit up, reacting further only when the light changes. The biological system is more efficient in handling information, as it only sends new data.
Advertisement
John Labram at Oregon State University聽and his team replicated the biological sensor with perovskite, a light-sensitive material which changes capacitance when illuminated. Sandwiched between two electrodes, the perovskite layer produces an electrical spike on being illuminated as it charges and discharges, but no further response until the lighting changes.
The data compression achieved by the retina-like set-up replaces burdensome digital processing often found in smartphones, which takes time and energy, he says.聽鈥淲e have a single pixel doing something that would currently require a microprocessor. This device can be viewed as one part of an ongoing endeavour to make certain types of electronics more like humans,鈥� says Labram.
He says these retina-like sensors could eventually allow AI systems to observe moving scenes and learn in real time. In the shorter term, they could help in smart vision systems such as self-driving cars and robotics, thanks to the rapid handling of the visual data. A paper on the sensor will be published in Applied Physics Letters in November.
鈥淭he idea of using very low-complexity circuitry to implement computation 鈥� such as orientation sensitivity 鈥� would allow on-chip visual processing with relatively low power,鈥� says Anil Anthony Bharath at Imperial College London.
However, Bharath notes that the technology will only be viable if it can be translated into low-cost devices which confer sufficient advantages in real-world computer vision applications.
Labram鈥檚 team is now working on assembling sensor arrays and feeding the output from its sensors into a neural network to create a fully biological vision system.
