Here is all the Neural Network style transfer experiments.

After several times of trying, I found that content weight of 5, style weight of 1, iteration at 800 comes out pleasing images. And if style image has the similar structure of a city like a block, street can fits well with content images.
Details in the style image also demostrated clearly at 800 iteration, especially the ‘body’ parts in the patern.
In order to find out how the color of image matters during the style transfer, I choose this image of Mandbrot set as style patern to different maps of Xiong’an including both plans and satellite photos. Conparasing the fisrt and the second image on the right side, clear boudaries were tend to be kept while opaque ones tend to be modified.

Unlike Vennice and Detroit and other cities on the right that have simple paterns fibering the texture of city, I also choose the first two maps that are of Tokyo with areas differ in densities(central area and urban area) and building scales(high rises and villas), and Shenzhen with residential areas and landscape areas. Let’s how it turns out:

cClusters of the patern were detected as blocks to merge into the texture of each city, the basic structure of city is kept though(blocks, streets and coast lines).

When reverse the content and style, the reflection between each image is more clear. Boundaries to boundries and clusters to clusters.

If the input image has more clarified elements, the results

Conclusions are as follow:

1 Color clusters can be input as layers that reflect in style image;

2 Obvious boundaries obviously kept in the results;

3 New elements can be add into map where it is slightly different;

4 Maps with similar scale matches better.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s