Project 3 Fun with Frequencies and Gradients!
Amy Shan
Here's some pics of some things I did with some math.
1.1 Warmup
1.2 Hybrid Images
Here's what Derek and his cat look like combined as one.
Here's what my friend Ethan and his favorite celebrity look like.
Fourier Analysis on Dwayne "the Ethan" Lee
Here's two more examples:
1.3 Gaussian and Laplacian Stacks
Stacks on stacks on stacks....
Here it is for the Dali painting.
Here it is for Dwethan.
1.4 Multiresolution Blending
Oraple
Cat Dog
Hand Eye (Irregular Mask)
Part 2: Gradient Domain Fusion
In Part 2, we want to transpose pixels from one image to another while preserving the overall look and feel of both images. As mentioned in class, humans perceive images mostly through gradients within the image, and not with the absolute intensity values within the image itself. Using this knowledge we implement the gradient domain fusion technique to minimize the difference in gradients between the source and destination image rather than focusing on preserving absolute quantities in either image. We represent this idea with the following equation, which we use the result of to transpose our pixels.
2.1 Toy Problem
In the toy problem, we focus less on transposing pixels and more on using knowledge of the gradient to completely reconstruct one. We take the x and y gradients of the image and save them into an array, and add in the single intensity value of a corner pixel for a point of reference before using this array to reconstruct the original image.
2.2 Poisson Blending
Featured Result
As seen in the naive picture of cutting and pasting pixels directly from the source onto the destination image, John Cena doesn't blend in very well with the background of the bird photo because of the harsh contrast between the pixels within the mask to those outside the mask. Using the Poisson Blending technique we learned in class, I minimized the gradient between the source and the destination image, which came at the cost of having a slightly more discolored John Cena.
Other Examples
The whale in the lake probably failed because the contrast between the colors and the overwhelmingly white color of the whale belly meant that "spreading out the error" through least-squares resulted in a color gradient heavily skewed towards the darked hued lake.
Blending Style Comparison
Clearly the poisson blending worked a lot better than the multiresolution blending, probably because more specific mask defined in the poisson blending method prevented the residual outline that occurs in the multiresolution version of the same image.