Specialists from Perm Polytechnic University (PNIPU) have trained a neural network to extract objects in images, taking into account the context of the image. Machine vision no longer makes mistakes due to changes in the perspective of the frame and can detect weapons in photographs with many people.
Neural networks are actively used in most different fields, because they can be used to automate many processes. One of them is image recognition. This function is necessary in robotics, in the automotive sector, in medicine and video surveillance systems. Neural networks "see" and understand images using machine vision, when certain types of objects are matched to the image in photographs. However, this technology is still imperfect, because when searching for objects, the context of the image may be ignored. In addition, errors can occur if the angle of the image or its quality changes, since objects are not rigidly attached to the image.
The machine vision system of Perm scientists can highlight the contours of an object and accurately determine its classification, since it uses a two-stage image processing scheme. Unlike a single-stage scheme, it takes into account the context of the image and the distance of objects or changes in perspective. At PNIPU, machine vision was trained to the following algorithm: first, a part of the image with the necessary object is selected. The neural network does not react to an unnecessary fragment of the image. After that, machine vision highlights and segments the necessary objects.
An increase in accuracy of 25% on individual test images occurs due to the artificial restriction of the assignment of categories and localization of objects in the context of the scene of the processed image.
The development of scientists from Perm will help to identify dangerous objects in pictures in a crowd. After all, a regular neural network cannot correctly extract the necessary objects, for example, weapons, on distant or close-up objects. If you first correctly detect the silhouettes of people in the photo, then the detection of weapons will be accurate.
Machine vision for opposite purposes: abroad, an engineer created a drone with a facial recognition system that can chase and attack a person
Active work to improve facial recognition systems is also being carried out outside of Russia. For example, engineer and entrepreneur Luis Venus created a prototype drone that can quickly find a specific person in a crowd, chase and attack him. The device is equipped with facial recognition and guidance systems. The drone is trained to find people in the frame and fly at full speed to the desired object. The engineer spent several hours creating the device.
We check for bombs and weapons, but THERE ARE NO ANTI-DRONE SYSTEMS FOR LARGE EVENTS AND PUBLIC SPACES YET. I was also able to add facial recognition to it and make it attack only someone it knew, it could easily identify a person from 10 meters away. I bet we'll see some kind of terrorist attack using this type of technology in the next few years.
Venus noted that such a drone could be used to attack crowds of people, and systems should be created to combat such devices for civilian spaces.
Read materials on the topic:
Russian researchers have trained a neural network to recognize people's faces from awkward angles
A traffic light with a pedestrian facial recognition camera was tested in Moscow
Kaluga Electromechanical Plant presented arched metal detectors with a facial recognition system
