Project 3: Fun with Frequencies and Gradients!
Alex Stennet, cs194-26-agn
Part 1: Frequency Domain
Part 1.1: Warmup
Original:
Sharpened:
To create a sharpened image the process goes as such:
- Separate out the color channels
-
Apply a gaussian filter to these channels
-
Take the difference of the blurred images and the originals
-
Readd the differences to the originals
-
Recombine the channels into a full color image
Part 1.2: Hybrid Images
Combining image of Prof John DeNero and Prof Josh Hug:
Combined image:
The fourier transforms of original John, low-pass John, original Josh, high-pass Josh, combination (respectively):
Part 1.3: Gaussian and Laplacian Stacks
Given the original image:
Gaussian stack (stdev=2,4,8,16,32 respectively):
Laplacian stack:
Applying to the result of 1.2 gives:
Gaussian stack (stdev=2,4,8,16,32 respectively):
Laplacian stack:
Part 1.4: Multiresolution Blending
Orange example images:
Combined:
Given original images:
Combined:
Original images:
Combined:
Stack of lake 1:
Stack of lake 2:
Stack of combined:
Comparing lake and marker, the marker result was much better due to the very little difference in the two images. Essentially the only difference is the pattern and the color of the marker. For the lake blending, at the edge, there is clear blurring (mostly due to the tree line and differing sky colors).
Another example for fun:>
Combined:
Part 2: Gradient Domain Fushion
Part 2.2: Poisson Blending
The goal is to, given 2 images, make a combined image that merges the 2 images according to some mask as well as the images respective gradients. A naive solution to the problem of merging two images is to simply crop; however, this will lead to clear artifacts at the boundary of the crop. This boundary artifact can be resolved by attempting to make the edge smooth and then tweaking the internals of the image to correspond to these altered boundary pixels.
Original Images:
Mask:
Naive merge:
Poisson Blending:
Poisson Blending (with mixed gradients):
.png)
Original Images:
Combined:
