CS 143 / Project 1 / Image Filtering and Hybrid Images
Introduction
The goal of this assignment is to get a hybrid image which appears differently at different distance by combining the high frequency part of one image with the low-frequency part of another. This approach is describled in SIGGRAPH 2006 paper by Oliva, Torralba, and Schyns.
Implementation
To relate the hybrid image and high frequency, low frequency portion of two images, we have H = I1*G1 + I2*(1-G2) where H is hybrid image and I1, I2 are two different images and they are filtered by low-pass filter G1 and high-pass filter (1-G2)(The operation is in frequency domain). The images below show the visualization of one example of low-pass filter(Gaussian filter, left) and high-pass filter(1-Gaussian, right) in frequency domain.
 |  |
In this assignment, for low frequency portion we apply Gaussian filter to keep the low frequency information. For high frequency portion, we get low frequency portion first and subtract it from the original image, then we get the high frequency information.
Results
According to the paper, to generate a hybrid image which is visual appealing, two portions should have a good alignment(obey the rules of perceptual grouping) and two cutoff-frequency should be correctly chosen (two filters in frequency domain have little overlap). The images below are the result hybrid images under various cut-off frequency(left) and according visualization of two filters in frequency domain(right).
| Cutoff-frequency | Hybrid image | Two filters in frequency domain |
Other results
Click the images to see the down-sampled scales.
 |  |