Technology: Electronic retina picks objects in a flash


By ELISABETH GEAKE An intelligent artificial retina which learns to recognise shapes could allow robots to spot objects, such as bolts on a conveyor belt, anywhere in their field of view almost instantly. Like a biological retina, the system makes an initial decision about whether it recognises an image, before passing it to a more sophisticated processor such as an ordinary computer. It infers common factors from different images, instead of searching for combinations of pixels, as a conventional computer does. This reduces the time needed to pick out objects. If the artificial retina can be scaled up, it could work with general-purpose computers and speed up production lines dramatically. The system is a form of neural network – a computer program which mimics certain actions of the brain. Until now most neural networks have worked entirely on electric signals, but the artificial retina operates on light signals, and so can process images much faster. So no matter how big the retina is made, processing the light takes just a few picoseconds (million-millionths of a second). Paul Horan, a research team leader at Hitachi’s laboratory in Dublin, and John Hegarty and Andrew Jennings of Trinity College, Dublin, made the artificial retina by building optical and electronic devices on a single chip of gallium arsenide. On receiving an input, such as an image from a video camera output (encoded as either a high or a low voltage), the retina converts it into light using asynchronous Fabry-Perot modulators (AFPMs). These are finger-shaped optical elements, each one the width of a human hair, that act as mirrors. The Dublin team has an array of 21 mirrors, covering a square 2.5 millimetres across. When the input voltages are fed into the AFPMs, they alter the mirrors’ reflectivity. If infrared light is shone onto them, the amount of light reflected by each AFPM depends on its input voltage, and the result is a row of dark and bright stripes. The stripes are then ‘autocorrelated’, a process which makes the position of the shape in the field of view unimportant. The pattern of stripes is rotated, by reflection, through 90degree and superimposed on the first AFPM array. The overlapping stripes produce a dot pattern which is always symmetrical about one of the diagonals of the square. For example, if the incoming image is of a letter T, then Ts in different places in the field of view produce the same dot pattern but at different positions along the diagonal. These dots are then shone onto a second array of AFPMs, and the light reflected from them is focused onto a photodetector. The reflectivity of each mirror in the array is set by ‘training’. To get the artificial retina to recognise a T, for example, it is fed examples of Ts and the reflectivity of each finger is adjusted to make the output signal from the photodetector as high as possible. When it is shown real Ts, which may be in different positions or have fuzzy edges, the signal from the photodetector should still be strong enough for the T to be recognised. The maximum reflectivity of the AFPM is 20 times as high as the minimum. Horan says this wide range is needed to give enough different weights. The retina needs fewer than ten examples to learn the shape of an object,
  • 首页
  • 游艇租赁
  • 电话
  • 关于我们