CS 791B Computational Photography

Hybrid Images

Asem Othman

Introduction

Hybrid image is a static image with two visual interpretations, which change as a function of the viewing distance. Hybrid images are generated by summing two images at two different spatial scales: low-spatial scale (filtered by a low-pass filter) and the high-spatial scale (filtered by a high-pass filter). In this project, the hybrid image technique has been implemented. In the following, several factors which lead to visually pleasing hybrid images are discussed with an illustrative example. Next, a set of examples is listed.

Example #1 “My daughter”

Neutral Expression	Smiling Expression
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\maryam1.JPG$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\maryam2.JPG$

In this example, there are two images of my daughter (Maryam) with different facial expression, i.e., natural and smiling. The goal is constructing a hybrid image which contains two facial expressions interpretations: smiling (at a far viewing distance) and natural (at a near viewing distance). The hybrid image is created by blending low-spatial frequencies from smiling image and high-spatial frequencies of neutral expression image.

Smiling Image		Low-Pass filter	Filtered Image
	*
Netural Image		High-Pass filter	Filtered Image
	*	(1- )
Low-Pass Filtered Image		High-Pass Filtered Image	Hybrid Image
	+

For all of the images, Gaussian filters were used in the Fourier domain. The standard divisions of Gaussian filters were utilized to tweak the hybrid image by playing the same role of cutoff frequencies in conventional filters. The selection of cutoff frequencies for low-pass filter and high-pass filter control the viewing perception at certain distances. For example, if the cutoff frequency for the high-pass filter is large, then only details with very high frequencies are preserved. Thus, the high-frequency image would soon disappear as viewer start to step back from the image. In this project, I wanted the image with smiling expression to be visible at a medium viewing distance and minimize the trials of constructing the hybrid image. Therefore, the standard divisions of both Gaussian filters were the same (σ_high = σ_low = 20).

Neutral Expression	Smiling Expression	Hybrid Image
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\maryam1.JPG$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\maryam2.JPG$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$

Alignment

Due to the slight movement of the head (it is difficult to keep a 20 months baby steady), the hybrid image is not visually appealing and at different distances it is difficult to determine only one expression. Therefore, 11 points have been marked on the facial features of the two images and have been used to align the two original images.

Annotated Image	Annotated Image

Aligned Image	Aligned Image
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\aligned_maryam2.jpg$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\aligned_maryam1.jpg$

Hybrid Image

$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\output_20_11.bmp$

Gaussian and Laplacian Pyramids

In the following, the Gaussian and Laplacian pyramids are shown to illustrate. For images in the Gaussian and Laplacian pyramid, each image is still a hybrid image (filtered with a band-pass filter). Thus, you can observe that the high-frequency component (i.e., netural face) will disappear one-by-one.

Gaussian Pyramid

Laplacian Pyramid

Example #2 “Asem Washington”

Denzel Washington	Asem Othman
$Description: H:\My_face\EXP1\shares1\den.bmp$	$Description: H:\My_face\EXP1\Target1\asem2.bmp$

Annotated Image	Annotated Image
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\den_anno.bmp$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\asem.bmp$
Aligned Image	Aligned Image
$Description: H:\My_face\EXP1\Test01\asem2.bmp\share1.bmp$	$Description: H:\My_face\EXP1\Test01\asem2.bmp\target.bmp$

High-Pass Filtered Image

(σ_high = 30)

Low-Pass Filtered Image

(σ_low = 20)

Hybrid Image

$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$

Gaussian Pyramid

Laplacian Pyramid

Example #3 “Iron Lion Zion”

Bob Marley	Lion
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\BobMarley.bmp$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\lion2.bmp$

Annotated Image	Annotated Image

Aligned Image	Aligned Image
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\aligned_marley2.bmp$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\aligned_lion2.bmp$

High-Pass Filtered Image

(σ_high = 15)

Low-Pass Filtered Image

(σ_low = 15)

Hybrid Image

$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$

Gaussian Pyramid

Laplacian Pyramid

Effects of Color

Using Color of Low Frequency Component	Using Color of High Frequency Component	Using Color of Both
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$

To demonstrate the effect of color in hybrid image, the images of the lion and Bob Marley have been used. Removing colors of high frequency component only would not affect the perception of the details significantly.

More Examples

Image 1	Image 2	Hybrid Image	Comments
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\panda.bmp$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\Black.bmp$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$ $Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$	Failure: 1- The image (panda) passing through the high-pass filter does not contain enough high-frequency details in the region of interests. Then, the resultant hybrid image may not have significant perception change as the viewing distance varies. 2- The alignment affected the perception. 3- It was also difficult to search for cutoff frequencies that would balance the low and high frequencies of the image
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\lion.jpg$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\angelina.jpg$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp_40_2.bmp$ $Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp_40_2.bmp$	Success: In this example, failure reasons in the above example have been avoided, therefore, the hybird image has two interpretation which easily could be distigushed at different distances.
$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\3.jpg$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\2.jpg$	$Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$ $Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\hybrid\outputtemp.bmp$	Planting Democracy in Tahrir square $Description: C:\Asem\Fall_2011\CS 791B Computational Photography\Assignment_1\Planting-DEMOCRACY-in-Tahrir-Square.png$