The new fully-silical computer vision material advances insensur visual processing technology

Researchers from the University of Massachusetts Amherst have advanced the development of computer vision with new silicon -based equipment which can both capture and process visual data in the analog field. Their work, describe in the newspaper Nature communicationsCould finally add to large -scale computer vision tasks, with a high intensity of data and sensitive to latency.

“It is a very powerful retinomorphic material,” explains Guangyu Xu, an associate professor of electricity and computer Engineering and assistant associate professor of biomedical engineering at Umass Amherst. “The idea of ​​merging the detection unit and the processing unit at the device, instead of separating them physically, is very similar to the way in which human eyes Treat the visual world. “”

Existing computer vision systems often involve redundant data exchange between physically separated detection and computer units.

“For example, the camera of your mobile phone Capture each pixel of data in the field of vision, “explains Xu.

However, this image has more information than the system requires it to identify an object or its movement. Consequently, the time required to transmit and process this additional information introduces a discrepancy to understand the visual information captured, which is often sensitive to the time and high data intensity.

“Our technology is trying to cut this latency between the moment you feel the physical world and the moment when you identify what you capture,” he said.

XU and his team created two integrated tables of tunable silicon photodetectors at the door, or insensur visual treatment networks. Bipolar analog output and low -power operation, a painting Can capture dynamic visual information, such as light changes motivated by events, and spatial characteristics can be captured in static images to identify what the object is.

The scaling of these silicon tables is promising for retinomorphic IT and intelligence detection. For dynamic movements, when they are asked to classify human movements (walking, boxing, waving and applauding) in sophisticated environments, the new analog technology was precise 90% of the time, while digital counterparts were precise from 77.5 to 85%. For static images, their technology classified manuscript numbers with 95%precision, which surpasses methods without insane IT capacities (90%).

A unique feature of these tables is that they are made of silicon, the same material used in computer fleas, unlike visual processors previous inventors which are mainly made of nanomaterials. As such, these tables are more compatible with complementary metal-oxide semiconductors (CMOS), the most commonly used semiconductive technology used to build integrated circuits in a wide range of electronic devices such as computers and memory chips. This compatibility makes them particularly suitable for large -scale computer vision tasks, in which many processes are executed simultaneously, also known as high parallelism.

“Our all-silicion technology lends itself to the integration of CMOS, mass production And a large -scale operation with low variability, so I think it is a major jump in this area, “explains Xu.

XU gives concrete examples of potential applications for this work. The first is autonomous vehicles: “You must always, in real time, treat what surrounds your vehicle and the speed at which they move,” he said. Any reduction in processing time increases the safety of autonomous vehicles.

Another area that benefits is bioimperie. Current technology Can capture much more data than that is really necessary.

“We can perhaps compress the amount of data and give the same biological overview for scientists,” explains XU.