CS 194-26 Final Project

Andy Wu, Fall 2020

Gradient Domain Fusion

Poisson Blending

The original images are

Basic copy pasting results in an image with a strong outline and wrong color.

After Poisson blending, the result looks very good. Up close, the border around the penguin is still blurry.

I also tried blending the penguin chick with another snow picture. The result was not as good since the Poisson blending does not take into account the background under the source image. It is very easy to see the blurring when looking at the sky and trees around the chick.

Finally, I inserted a image of a bear into a pool. From a glance, it looks pretty realistic, since the bear takes on the color of the water, but when we look at the shadows cast by the palm tree, we can see the blurriness of the blending.

Mixed Gradients

To fix some of the problems, I also implemented mixed gradients, to account for strong gradients in the target image. We can compare the bear in the pool with the one from before. Using mixed gradients preserves the shadow of the tree, making the blending look much more realistic

Neural Style Transfer

Nekarfront

The original Nekarfront image is

After applying different styles

The results are

Campus

The original campus image is

After applying different styles

The results are

The final image of campus slightly gained the polygonal structure of the Picasso image, but did not fully become black and white. It would probably take more tuning of the style weights to force the image to be black and white.