Artificial intelligence has learned to recognise in the pictures photoshop

Штучний інтелект навчили розпізнавати на знімках фотошоп

A team of American researchers has taught artificial intelligence to recognize photos that have been modified in the graphic editors. It can help in the fight against fakes, says Platforma.

With the development of digital networks of all the bigger becomes the problem of the edited photo. For memes, humor or just eliminate cosmetic defects in the portraits, photoshop is exactly what you need, however, for the media fakes can carry a lot of adversity. Therefore, a new study presented at the American Conference on Computer Vision and Pattern Recognition, it can be very important for the global information system.

The developers at Adobe and the University of Maryland in College Park has offered to fight the fakes with the help of neural networks, which will determine the three most popular operations are used when a significant change in content in the pictures: removing objects, cloning parts of an image, and the combining of images with different pictures.

For this purpose the experts chose a fast regional network – Faster R-CNN. Based on her program RGB-N authenticates the image simultaneously on two channels. The first analysis allows to identify the visible artifacts of editing, unusual contrasts, and other visible signs of retouching. The second channel allows you to notice the uneven distribution of noise in the image – thus the algorithm can see even best done from the point of view of the person fake.

Штучний інтелект навчили розпізнавати на знімках фотошоп

© Zhou Peng et al., / CVPR 2018

To train RGB-N developers used 328 thousand images spaced 2.5 million visible objects – such as a car or a refrigerator. Thanks to this set, the authors created a 42 thousand pairs of fake photos-the original, which was used for training the algorithm.

Штучний інтелект навчили розпізнавати на знімках фотошоп

© Zhou Peng et al., / CVPR 2018

Testing with the edited images showed that the recognition accuracy of the algorithm is impressive 80-90%. If technology will be able to enter into wide use, the problem of fakes may become a little less global.

Share Button