Creating synthetic images with controlled properties?

Hi everyone, Im trying to find out if there is a clever way to mathematically generate a set of different images with some key statistical properties fixed. For example: create a set of b&w images with same average spatial frequency content, same average luminance, but different structure.....

any ideas? any cool references about this? thanks!

Mario

Answer 1

[Image Analyst]

Well here's one demo where I do something like that to generate clouds. Attached below in blue.

But your question is very vague. What is average spatial frequency exactly? A given spatial frequency pattern is the fft and inversing that will give you an image. If you change the fft you get a different image. So what does average mean in that case?

I've seen methods where people lay down random dots and then do morphological operations on them to get different worm-like looking images, but I don't know where any source code for that is. So it really depends on what you want to achieve. What do you want your images to look like?

[Mario Trevino]

Hi IA, thanks for your response. Yes my question is very vague... Im searching for options. Im referring to the radially averaged fft. refining the question: how to create an algorithm to generate different images with similar radially averaged fft and average luminance?.... the inputs could be natural images which are then iteratively processed to achieve that... I donk know.... what do you think? let me check your clouds...

mt

[Image Analyst]

Perhaps you'd be interested in Perlin noise: http://physbam.stanford.edu/cs448x/old/Procedural_Noise%282f%29Perlin_Noise.html

[Mario Trevino]

thanks!

Answer 2

[Walter Roberson]

Some of those are trivial.

For example for the average luminance one, decide what average luminance you would like to use. Then generate random images. For each image, brighten (or darken) the image until the resulting average luminance matches the desired value. This is a completely linear transformation (unless you end up saturating a pixel): if pixel #1 needs to be multiplied by 1.6429 to be bright enough, then pixel 2 would be multiplied by exactly the same thing. Constant ultiplication factor: target_luminance / actual_average_luminance