HW02: Hybrid Images
Wentao Liu
wentaol@andrew.cmu.edu
Overview:
This project involves creating hybrid images.
Results: Favorite
First, the images are aligned and transformed using the provided starter code.
The high frequency content of the first image is removed by directly applying a gaussian filter.
The low frequency content of the second image is removed by subtracting the from it the second image with a gaussian filter applied.
The resultant filtered images are then combined by adding the images together.
As shown in the FFT images, this results in an image with the high frequency content of the second image, and the low frequency content of the first image.
As a result the picture looks like the first image from far, but looks like the second image up close.
Cut-off: 80
Filtered Images:
Final Image:
FFT:
Laplacian Pyramid:
Results: Another
The cut-offs are adjusted manually. If the resultant image does not look like the first image from far, then it means that we need a greater spread of frequencies from the first image, and hence the cut-off is increased.
If the resultant image does not look like the second image from up close, then it means that we need a greater spread of frequencies from the second images, and hence the cut-off is decreased.
One technique that increases the chance of success is to use a human face for the first image. I speculate that this works better, because the human brain is trained to recognize human faces from afar, as such it is more sensitive to the low frequency content of faces.
Another technique that applies when both images are faces is to make use of the eyes for alignment. This way, the images will be aligned by both direction and scale, because eyes tend to be the same distance apart.
Cut-off: 140
Filtered Images:
Final Image:
FFT:
Laplacian Pyramid:
Results: Fail
This does not really work because the two images are dissimilar, and is harder to align than faces.
Furthermore the general shape of the first image is not readily recognizable.
As such regardless of what cut-off is chosen, one of the images would fail to be perceived.
Cut-off: 140
Filtered Images:
Final Image:
FFT:
Laplacian Pyramid:
