CS123
Project: Image Blending
Author: Yan
Li
Login: yanli
Email: yan_li@brown.edu
Goal
of Project
The goal of this project is to implement Poisson Blend described
in P¨¦rez,
et al.
The
source image with a specific mask and the target image can be blended seamlessly
by solving the Poisson Equation.
My
Works
1 Optimizing
the algorithm for solving Poisson Equation.
2 Implementing
the Poisson Blend.
3 Implementing
Blend with mix gradient. (Extra Credit)
4 Implementing
Blend with Laplacian Pyramid. (Extra
Credit)
5 Implementing
Texture Flattening. (Extra Credit)
IMPLEMENTATION
1
Algorithm
Instead of
calculating the gradient for each pixel by convolution, I use ¡°diff¡± operator
in Matlab and compute the X,-X,Y,-Y
gradient, then mix them together and get the result which is same with
convolution.
Instead of
building sparse matrix by carefully index the matrix and retrieve each element
that is not zero, I pre-compute the indices and the values.
For the default
six test cases, the current algorithm for Poisson Blend will cost less than 10
seconds.
2
Poisson Blend
Here I show my results in the following
form. The columns includes the source image, the target image, the mask, the
result of Naïve Blend and the result of Poisson Blend,.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3
Poisson Blend with mix gradient
Instead of always
using the gradient of the source image, I use the larger gradient from either
the source image or the target image. This could generate some transparent
effect between the source image and the target image. The following form
illustrates the difference between Blend with mix gradient and without it. As
you can see, some blend results are enhanced, like the sixth image,
the rainbow has less halo effect. But some blend results are
not realistic such as the third one, the car looks really
transparent.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4
Blend with Laplacian Pyramid
I also
implemented the Laplacian Pyramid Blend. This kind of
blend can preserve the original intensity in source image perfectly, but the
results are not seamless. Obvious seam in the result image can be seen. Here is
the difference:
|
|
|
|
|
|
|
|
|
5
Texture Flattening
Possion equation
cannot only be used for blending. Indeed, we can use those
more important gradient information from the source image to wash out some relatively
useless details. Texture Flattening is such a technique
that can use generate a non-photorealistic effect.
I extracted
the useful gradient information by using edge detector (In my project, I use Sobel Operator to extract edge ). Those guidance gradient on the edge should be taken into
account, while others we just set the guidance gradient to zero.
Here is the
result of Texture Flattening:
|
|
|
|
|
|
|
|
|
|
|
|
6
Other Images
Here are
some other test cases of Poisson Blend:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
