Project 1: Hybrid Images
CS 143: Introduction to Computer Vision
Hang Su
Email: hangsu@cs.brown.edu
Sept. 26th, 2011
1. Methodology
Hybrid image [1] is generated by combining a low-filtered image and a high-filtered image, and can be perceived in two different ways according to the distance between the viewer and the image.
In this project, we try a modified way using gaussian/laplasian pyramids to get the two filtered image. The low-filtered image is achieved by firstly generating the gaussian/laplasian pyramid and then adding up the bottom N levels; the high-filtered image is achieved by firstly generating the pyramid as well but then adding up the top M levels (N, M are both configuration parameters chosen vary per image). Below is an example of 8-level gaussian/laplasian pyramid.

After having the two filtered images, simply combining them results the hybrid image (practically, hybrid image is generated by adding up wanted levels in both pyramids directly). Below is an illusatration how hybrid image is generated.
Rocket | (low-filtered) Rocket | Hybrid Image | (High-filtered) Taipei 101 | Taipei 101 |
---|---|---|---|---|
![]() |
![]() |
![]() |
![]() |
![]() |
2. Results
Hybrid Image | Distant View | |
---|---|---|
Heath & The Joker![]() ![]() |
![]() |
![]() |
Tiger & Tiger![]() ![]() |
![]() |
![]() |
Girl & Funny Man![]() ![]() |
![]() |
![]() |
UFO & Tower Bridge![]() ![]() |
![]() |
![]() |
Rocket & Taipei 101![]() ![]() |
![]() |
![]() |
3. Discussion
3.1 Color information
Intuitively, color information might be mainly related with low-frequent components. But actually in most situation color information helps even more for the perception of high-frequent components.
From the example below we can clear see that the top right result(or probably the top left) is the best when considering percepting "The Joker" from a close distance. All of the four results are equally nice result as for the low-frequent perception from a distant place. And we can tell that adding color infomation for low-frequent components (from right column to left column) barely helps for the perception of either image. However, things can be quite different when dealing with different pair of images.
color in both components | color only in high-freq component |
---|---|
![]() ![]() |
![]() ![]() |
color only in low-freq component | both in greyscale |
![]() ![]() |
![]() ![]() |
3.2 Choosing cut-off freqency
Choosing of cut-off freqency is also highly related to the two input images. Most time a rule is the two freqency components should not overlap with each other. [1] As we walking away from the screen, we can perceive one after another from the blow images the "Heath" part; in a far enough distance, all the hybrid images looks exactly the same.
lowest 3 & highest 3 compoments | lowest 3 & highest 4 compoments |
---|---|
![]() ![]() |
![]() ![]() |
lowest 4 & highest 3 compoments | lowest 4 & highest 4 compoments |
![]() ![]() |
![]() ![]() |
4. References
[1] A. Oliva, A. Torralba, P.G. Schyns. Hybrid Images. In Proceedings of SIGGRAPH 2006.
Hang Su
Sept. 26th, 2011