This project creates hybrid images, detailed in this paper, from images of our choice.
I created the low-pass images using a simple gaussian filter, generated by MATLAB's fspecial function, and high-pass images by taking the difference between the original image and a gaussian filtered image. Cutoff frequency was determined by fspecial's sigma parameter, and some hand tweaking to make the images look good. The kernel for the fspecial function was set to 25 in both directions.
While it might be more thematically appropriate to blend Robb's face with a wolf, these pictures lined up more nicely. These pictures are promotional portraits for the third season of Game of Thrones. The two characters are Robb and Bran. Robb was used as the high frequency image and Bran was used as the low frequency image. Cutoff frequencies were 30 for the high, 12 for the low. The filtered images are averaged together at the end.
Bran's FFT shows a bit of blockiness. This could be due to the white border on the image during alignment or due to a gaussian kernel that's not big enough for the image. It seems to not have a huge impact on the final image.
Final Image | Uncompressed
Because Toothless practically is a cat. Cutoff frequencies were 30 for high, 20 for low.
Final Image | Uncompressed
This did not work well at all. I think it's because the outline of the dragon and bird didn't really line up well, and the color contrast between the images are too strong to hide the outline of the dragon. Cutoffs were 30 for high and 16 for low.
Final Image | Uncompressed
Color use on the Game of Thrones portraits worked really well, especially when the color on the side of the faces were aligned. I suspect most of the color comes from the low frequency image, which is why color being on the same side of the face matters for these images.
Arya and the Hound - Final Image | Uncompressed
Bran and Robb - Color | Uncompressed
Generated for Bran and Robb.
Bran and Robb - Gaussian Pyramid | Full Size
Bran and Robb - Laplacian Pyramid | Full Size