Hybrid Images

aabouche

A hybrid image combines the low frequency data of an image with the high frequency data of another image. The resulting image has charactetistics of the second image when viewed at a short distance. However, increasing viewing distance will cause the hybrid image to look like the first image. When you are close to the image you perceive the high frequency information over the low frequency information. At a distance, the low frequency information becomes the high frequency information in the image because at a distance you cannot see the small details of the original high frequencies.

In order to create the hybrid images, you must isolate the frequencies of the two infput images. I created Gaussian and Laplacian pyramids of eight levels for each image. Below are the image pyramids for the two input images. The top row is the Gaussian pyramid and the bottom row is the laplacian pyramid. Creating the pyramid consists of blurring the original image and subtracting the blurred image from the original image to get the Laplacian filtered image for that level of the pyramid. Then you must downsize the blurred image and begin again for a new level of the pramid. I found that blurring, then downsizing by half, then upsizing to the original size to obtain an image to subtract from the original mage let me to get better Laplacian images. A lot of information was lost when I had simply blurred, subtracted then downsized the image. The resizing method significantly afftected the Laplacian images.


The images below show how you could reconstruct an image from its pyramid. The images are a sum of the last image in their Gaussian pyramid (the first in the sequences) and the Laplacian filtered images (the following 8 images). At each level the image is upsized.

The following sequence uses the gaussian image from the Marilyn image pyramid and the six following Laplacian images from the Marilyn Laplacian pyramid. The last three images are Laplacian filtered images from the top of the Einstein Laplacian pyramid. Thus, the low frequencies from the Marilyn input image are combined with the high frequencies from the Einstein image.

Inputs and Results

The low frequencies from the leftmost images are combined with the high frequencies of the center images to create the rightmost images.