Neural network taught to recognise human faces from uncomfortable angles
Russian scientists have taught the neural network to accurately recognise the faces of people who, using uncomfortable angles, are trying to hide from CCTV cameras – as noted by senior researcher at the South Ural State University of Russia Alexei Ruchai, TASS reports
“We trained the VGGface neural network to recognise the so-called 2.5-dimensional images. People who are trying to take advantage of uncomfortable angles cannot get away from such artificial intelligence,” said the source of the news agency.
He explained that face recognition systems have long been part of everyday life, but they work with flat, i.e., two-dimensional images. The computer sees only a matrix of pixels, and its task is to learn how to distinguish facial features in it and learn to recognise them among thousands of others. “If a person’s face is far from the camera or too close to it, if it is rotated at an angle or the lighting falls ‘wrong’, then the neural network may give a false result. We have solved this problem,” the scientist said.
The found solution is based on replacing a two-dimensional picture with a 2.5-dimensional one, where the depth of the image is also added (the distance from the camera to the surface of the object, that is, the person’s face). To measure the depth, scientists used an estimate of the curvature of the infrared grid: infrared illumination.
