Researchers at the Weizmann Institute of Science in Israel are using deep internal learning, in which a machine learning algorithm ascertains the internal structure of a single image from scratch, to edit photos without previous training.
This achievement builds on research from a team at the Skolkovo Institute of Science and Technology in Russia, involving Deep Image Prior (DIP), a technique in which a multi-neural network is trained to replicate a specific image by looking for hierarchies of repeating features.
The Weizmann researchers' Double-DIP process has two DIPs running in parallel, with each converting a random input into an image, and both images superimposed on and compared to a target image.
The DIPs then independently modify their parameters so their combined image comes closer to the target. Dmitry Ulyanov of Moscow’s Skolkovo Institute of Science and Technology said he and his collaborators designed DIP to study the importance of network architecture (versus data).
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2019 SmithBucklin, Washington, DC, USA