fft2 applies a 2D Fast Fourier transform to the image, which creates a complex (as in complex numbers) image. abs takes the magnitude of the complex image elementwise. log10 scales the result for display.

Then I downloaded the image from the MATLAB app, went into the Photos app and (badly) cropped out the white border.

Despite how dramatically different it looks, it actually contains the same [1] information as the original image. Said differently, you can actually go back to the original with the inverse functions, specifically by undoing the logarithm and applying the inverse FFT.

[1] Almost. (1). There will be border problems potentially caused by me sloppily cropping some pixels out of the image. (2). It looks like MATLAB resized the image when rendering the figure. However, if I actually saved the matrix (raw image) rather than the figure, then it would be the correct size. (3) (Thank you to @itslilith@lemmy.blahaj.zone for pointing this out.) You need the phase information to reconstruct the original signal, which I (intentionally) threw out (to get a real image) when I took the absolute value but then completely forgot about it.

[–] asbestos@lemmy.world 20 points 1 year ago* (last edited 1 year ago)

Can guarantee OP didn’t expect this when he was making the post, damn

[–] itslilith@lemmy.blahaj.zone 12 points 1 year ago (1 children)

don't you lose information by taking the abs(), since to restore the full information you need both complex and imaginary parts of the fft? You could probably get away with encoding Re and Im in different color channels tho

[–] PM_ME_VINTAGE_30S@lemmy.sdf.org 4 points 1 year ago

Correct. I updated my comment.

[–] Otherwise_Direction7@monyet.cc 5 points 1 year ago

Ok, this is really mad cool, I like it

[–] HerbalGamer@sh.itjust.works 4 points 1 year ago

Saving this comment in case I ever need a new phone background.

[–] Jackie_meaiii@lemmy.blahaj.zone 30 points 1 year ago (3 children)

I used the default photos app on my phone to draw a totally normal cow

[–] PM_ME_VINTAGE_30S@lemmy.sdf.org 12 points 1 year ago

So I took your image and ~~ruined my MATLAB account~~ used the most normal part of your totally normal cow as a 3D [1] ~~cockvolution~~ convolution kernel. So in some sense, I dragged the red and purple part all across your image and added up the results. Here's the result:

Figure 2024-01-15 14_46_27

Here's the MATLAB code:

normal_image = imread("totally_normal_image.png");

feature=normal_image(272:350,205:269,:);

feature_expansion = padarray(feature,[0,ceil((79-65)/2),0],'replicate');

for m = 1:1:3
new_normal_image(:,:,m) = conv2(normal_image(:,:,m),feature_expansion(:,:,m));

new_normal_image(:,:,m) = new_normal_image(:,:,m)/max(max(new_normal_image(:,:,m))); 
end

imshow(new_normal_image)

[1] The original image was practically grayscale, so only a 2D convolution is required, i.e. over 2 spatial dimensions. Since you added color, it adds a extra dimension, one per color channel. Which makes it more annoying to work with in MATLAB. I mean, I could have just dumped everything into grayscale, but I need practice with processing color images anyways.