CS194-26 Project 2: Fun with Filters and Frequencies

In this project, we first learned to use the difference filter (or the derivatives) combined with a gaussian filter to create a simple edge detector. We then learned that we could combine the filters and the result would be the same. (PART 1) In part 2, we first learned how to sharpen an image by enhancing the high frequencies, and learned how to combine the whole process into one convolution filter called unsharp mask filter. We then combined two images together to give the effect of seeing one object from afar and another object from close-up. This is achieved through combining the low-pass of one image and the high-pass of another image. Lastly, we learned how to blend two images smoothly by using Laplacian stacks and appropriate masks.

Original Cameraman

1.1: derivative filter

We first use the derivatives filter to compute the gradient of the images, and then took the magnitude of the gradients at the pixel by using the formula magnitude = sqrt(d_x^2, d_y^2). This is because d_x and d_y are two directions of the gradient vector, and thus this formula gives the magnitude of the gradient.

1.2: gaussian and derivative

We then apply the gaussian filter first before applying the derivatives filter.
What differences do you see?
We can see that after applying the gaussian filter, the image becomes much smoother compared to 1.1.


We then combine the two filters and then convolve the combined filter with the original image.
Verify that you get the same result as before:
We can see that the results are the same.

2.1: Sharpening

We sharpen the image by separating the high frequencies of the image using high-pass filter and add it back to the original to enhance high frequencies.


I also picked a picture of shang-chi and sharpened it.


I also picked an already sharp image, blurred it with gaussian filter, and then resharpened it. As we can see, the resharpened image is not the same as the original, because blurring image takes away information, while sharpening it does not add any new information. Thus, the resharpened image will be much lower quality and will very different.

2.2: Hybrid images

In this part I created hybrid images by getting the low-pass from one image and the high-pass from another. To do this we need to first align them with each other, and then apply the filters, and then add the two tensors together.

I also picked some of my pictures. My favorite is spongebob x patrick start.

I also did a hybrid between pizza and chocolate cake. It doesn't seem that great tho, probably because the pepperonis are too distinct making the cake hard to be seen.

Extra credit: colors

I experimented with colors on Derek and nugmeg. It's obvious that it works the best when both images are processed withe colors.

2.3: Gaussian and Laplacian stacks

This part, I used Gaussian and Laplacian stacks to blend the orange and apple together. Below is the Laplacian stacks covered with mask thats half 1s and half 0s, and the combined images.

2.4: Multi-resolution Blending

This final part, I picked two other pairs of images, aligned them, and created their blending.
The first is mountain fuji and a volcano.


I also blended Kobe Bryant and sharingan eyes. I used an irregular mask that I hard-coded by find the locations of the eyes. However, this did not work well because the mask is still a square, which does not provide a good blend for eyes. A round one would be better, but also harder to create.