During last couple of months I was going through a very interesting area in machine learning called Neural Style Transfer (original paper here). This is exactly what we were trying to do then. But then we failed because all that we had then was "photoshop merge" which did not work out. Neural style transfer does a very intelligent merge! Neural Style Transfer takes two images, a style image and a content image. Then the content image will be morphed by bringing in gist from style image!! Here are three images. With RD's permission I am inserting RD's clutter image (used as style) and my elephant & calf image (used as content) to create the new "neural style transferred" image.
Style Image
Content Image
Generated image with "Neural Style Transfer" - RD's style, my content (Who has the copyright for this one?!)
The algorithm provides many tweaking options to control the look and feel of the generated image. In particular we can adjust the relative weightage of style and content components.
I think artistically neural style transfer offers many avenues to explore. In particular I would like to explore whether it is possible to match the style from a music on to an image (instead of picture to picture). This goes back to our experiments earlier on matching "ragas" with tones/mood of an image. You may find the related discussion we had long ago here. I think this will be a novel area to explore for style transfer.
Acknowledgement: The image above has been generated using the nice program written by Harish Narayanan for the neural style transfer.
I hope you find this post interesting!
For the image post in the Nature Art Gallery and further discussion please click the thumbnail here below.
