A new experiment from Google is turning imagery from the company’s Street View service into impressive digital photographs using nothing but artificial intelligence (AI).
Google is using machine learning algorithms to train a deep neural network to roam around places such as Canada’s and California’s national parks, look for potentially suitable landscape images, and then work on them with special post-processing techniques.
The idea is to “mimic the workflow of a professional photographer,” and to do so Google is relying on so-called generative adversarial networks (GAN), which essentially pit two neural networks against one another.
With this Google experiment, the first, “generative” model tries to fix a picture that has previously been messed with on purpose (with things like brightness and contrast changed at random), while the “discriminative” one analyses and compares the original (messed) shot and the fixed one.
The result is a software that is able to understand the principles of good photography (like, for instance, not oversaturating colours), and uses this knowledge to work its way through the scanned images that come from Google Maps.
When the AI system recognises a potentially interesting image, it first crops it, then tweaks things such as saturation and the strength of dynamic range, and applies a filter (Google calls it a “dramatic mask”).
The results are — according to professional photographers — impressive. The photographers Google worked with on the small project ranked about two in every five shots as on par with either semi-pro or straight up professional grade shots.
Take a look for yourself at Google AI’s work below.