In order to arrive at content-aware image resizing, I tried using a couple of energy functions to assign values to different pixels in an input image. The most efficient function I found was the weighted sum of the gradient and of the local entropy in the image.
I think that uniform portions of the image would go away unnoticeably, and I wanted Matlab to think the same. It turns out that using an entropy filter on the image allows me to find out where the details of the image are (because they have high entropy). Combined with the gradient of the image, this gives a pretty good function to use for seam carving. Check the results below! :)
I observed that the my program is quite attracted to faces ... somehow it thinks that they will not be missed from the picture. Areas that have uniformm lighting are quickly spotted and removed from the image, but textures and large-scale shapes will be somewhat distorted. Also, it's interesting to see that in some photographs, the angle of the viewer appears to modify in the rescaled image.
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
|
Remember the parrot from the top of the page? Now it demonstrates cubism! |  |
 |
Cubist angles indeed :) |  |
 |
Cyert Hall has new strage windows! :) |  |
 |
Invisible giant having lunch over here |  |
 |
Matlab eating a grape |  |
 |
Oooh, the weight of the ceiling ! |  |
The images I used are either from www.public-domain-photos.com or my own personal collection.