Researchers at the Vienna University of Technology in Austria, supported by the European research project, the Graphene Flagship, have developed an image sensor with an integrated artificial neural network (ANN) capable of learning and classifying images within nanoseconds. The chip is a thousand times faster and uses much less power than conventional machine vision technologies.
The image sensor can simultaneously capture and process images, making object recognition many orders of magnitude faster. The device does not consume any electrical power when it is operating, since the photons themselves provide the energy for the electric current. The sensor is complemented by an ANN, used in this case to recognise an image.
The researchers in Vienna devised sensors containing nine pixels – the ‘neurons’ – placed in a 3x3 array. Every pixel in turn, consists of three photodiodes that provide three outputs. Each photodiode links its pixel to the other 8 pixels.
The current from each photodiode is determined by the intensity of incoming light and the voltage across it. Each neuron sums the individual currents coming from the other 8 neurons, and the combined values are then fed into a computer.
The photodiodes are made of sheets of tungsten diselenide, a layered semiconductor with a tunable response to light. The photodiode sensitivity to light is tuned using an applied voltage.
The device can classify images after a series of training processes, but it can also recognise a characteristic component or structure of an image from input data, without extra information.
The speed sets this device apart from conventional machine vision. Conventional technology is usually capable of processing up to 100 frames per second, with some faster systems capable of working up to 1,000 frames per second. In comparison, this system works with an equivalent of 20 million frames per second.
It has been suggested that the device will be scaled up with today’s technology and find applications in different fields, such