The theory is:
Put one image over the other as layers.
Blend mode diference on the top one.
Export the resulting image, and convert to grayscale.
Adjust levels or curves. Clean the white with a brush.
Use the grayscale image as a mask on your original image.
But probably newer AI algorithms will do a better job guessing the boundary of some objects.
If anyone wonders, it's mathematically impossible to perfectly restore the second transparent layer if alpha channel of the top layer has any values other than strictly 1 and strictly 0 (and it isn't 100% reliable even then).
Let's assume the background layer is dark gray (RGBA=100,100,100,1.0) and the top layer has three pixels, one gray (RGBA=150,150,150,1....
still one more:
remove smooth luminosity variations (=low spatial frequencies) by applying high pass filtering:
Increase contrast to get black and white. Use curves. Let there be a non-vertical transition zone, do not apply steep tresholding because it generates harsh edges:
There's still some color left. Desaturate to remove it, if needed.
Just to show ...
Photos of papers are almost always unevenly lighted. This makes using Levels very hard because it has to be a compromise between noise, leftover shading and losing lines. However, there's an easy solution.
Like others have already suggested, first you need to pick a color channel with the least noise. In your case, I picked green. Copy it into a new document....