Recognition of face disguise can only recognize half by seeing the eyes and mouth [Full text]

Technology dynamic face recognition technology is not new, Facebook used it as early as 2015 to track user information.

Its principle is: First, the system will select more than a dozen points on the face as data collection points, then use artificial intelligence algorithm to analyze one by one, and finally connect the information contained in each point to form a facial star network map. Face image.

Researchers in the United Kingdom and India have jointly developed a Disguised face identification (DFI). Under this system, people wearing masks, hats, glasses, and bearded beards may also be Identify the appearance.

DFI shares the same principles as general face recognition. The difference is that it uses optical flow technology to analyze facial features under the mask.

The stream of light is a series of imprints left on the retina as objects pass by. Pedestrians, cars, flying birds, rotating fans, and the scenery flowing outside the window on the train are optical streams. Scientists often analyze it. To get information on moving objects.

With the optical flow technology, DFI first found 14 points on the human face, with ten concentrated around the eye, one on the nose, and three around the mouth. In general, when the person being monitored wears a hat, a veil, or a long beard, the eyes will be exposed. These eye points are reference points and are important sources of information for identifying people.

14 data acquisition points set by DFI.

The three spots around the exposed nose and mouth become optical spots. DFI infers the shape of the face by capturing these blurred spots of light, comparing their position in the video or in different photos, and finally comparing it with the pictures in the police to determine if the man is a suspect.

However, light flow is only used when the outline of the object can be seen. In order to improve accuracy, the researchers set up two sets of systems to be used together, one set to capture the scene where the person is located, and one set only intercepts the person's head. This can increase the accuracy of DFI to 43% to 56%, but if the background is a little more complicated, it will have to drop about 7%. In addition, if terrorists do not use the veil to switch to helmets, DFI cannot track the stream of light, and there is no way to infer a portrait.

One or three Simple Face Disguise Datasets; two or four Complex Face Disguise Datasets.

"This technology has a long way to go," commented AI commentator Jack Clarke in the Import AI. However, this technology has provided new ideas for the pursuit of suspects and terrorists.