Google's "DeepDream" project.

"DeepDream" is a project initiated by Google as part of artificial intelligence and more specifically pattern recognition and machine learning (DeepLearning).
The programme, already in beta phase, was able to identify shapes in an image that it had learnt to recognise, and reinforce this resemblance. If, for example, a cloud is reminiscent of a dog, the DeepDream programme will reinforce this resemblance by accentuating the eyes or adding a texture reminiscent of the animal's fur.
The result is highly psychedelic images, which seem to have been drawn or painted under the influence of a hallucinogenic drug and are very reminiscent of dreams, or even nightmares.
Technically, it works on the principle of . artificial neural networks The programme must first "learn" by analysing a large number of images. It then builds the database which will then enable it to recognise a particular shape in an unknown image. For the moment, this learning process seems to have been fed mainly by images of animals ; which explains the proliferation of dog heads in the results.
The programme is provided by Google as Open Source. So you can install it and modify it on your own computer. But it does require a good knowledge of computers. Several websites offer interfaces that make it easy to upload images for DeepDream to process. The images are queued up and unfortunately the waiting time is long: around 24 hours.
To find out more :
We tested the system with a few images of our own. The results are shown below.

Photo generated by DeepDream (Google)
In the end, this photo was changed very little, apart from the faces and eyes on the barrel and the strange animal head above the shop window.

Photo generated by DeepDream (Google)
Note the leg, which has taken on a truly animal look, and the strange bug on the left.

Photo generated by DeepDream (Google)
The setting (initially a forest) has been populated with dog heads and strange characters. The most spectacular is the large dog head that appears on the T-shirt.
If you liked this page, share it on your favorite network :