Artificial intelligence can accurately remove tattoos from photos of people鈥檚 faces 鈥� potentially helping face recognition systems, which can be flummoxed by such tattoos, work accurately.
In , at Darmstadt University of Applied Sciences in Germany and his colleagues had identified that face painting and tattoos can impair the performance of face recognition systems. They do this by altering what a computer vision model expects to see in key areas around features such as the cheek, chin and nose.
鈥淪ince facial tattoos are a permanent form of alteration that cannot easily be removed, it is especially interesting to investigate if removing facial tattoos digitally can improve the performance of face verification systems,鈥� says Ibsen.
Advertisement
He and his colleagues have now trained an algorithm to remove facial tattoos by automatically adding ink to 41 images of untattooed faces. The tattoos ranged in size, covering between 5 and 25 per cent of the face in the image.
Those computer-altered tattooed images were then used to train a generative adversarial network (GAN), a type of deep learning algorithm. 鈥淭he idea here is to teach our algorithm to 鈥榯ranslate鈥� from a tattooed facial image to the corresponding facial image without tattoos,鈥� says Ibsen.
The GAN then removed the tattoo on the face 鈥� which it largely did without issue, except for large tattoos that covered the entire face 鈥� and the GAN-altered images were tested against a face recognition system. The rate of errors, where the face recognition system either incorrectly matches two unmatched faces or fails to recognise a matched face, was halved when tattoos were removed by the GAN.
鈥淭he preprint [study] presents a useful technique for training algorithms to remove facial tattoos and improve recognition accuracy,鈥� says at the University of Lincoln, UK. However, Ritchie points out that because the images used to train the algorithm are themselves digitally edited, it may not be as effective with unaltered, real-world images.
Reference:
