zlacker

How resilient is blurring against deconvolution?

>>leni53+(OP)
Can deconvolution create new data? I thought it was just a way to upscale images.

>>leni53+(OP)
> How resilient is blurring against deconvolution?

This depends a lot on the implementation details. If you blur an image using arbitrary-precision real numbers, then blurring is invertible. If you add a bit of random noise, or quantize your pixels into a finite-precision data type, then it becomes essentially one-way, and you cannot recover the original image.

>>ibrarm+U1
If a blur filter uses a convolution, then it's invertible through a deconvolution.

>>ibrarm+U1
Deconvolution is any attempt to recover data that has been passed through some known transformation. It can "create new data" because it is effectively mathematically-educated guesswork.

In the case of upscaling an image, deconvolution involves looking for images which, when scaled down, resemble the original image being upscaled. That kind of pre-image approach can be applied to blur as well (if the blur process is deterministic).

>>klyrs+J3
There will be some information loss from edge effects and quantization noise. But mostly invertible.

>>ibrarm+U1
It can't, but blurring might remove less data than what you want.

>>im3w1l+n6
You're not wrong but invertible filters, noise or no, are simply not anonymizing and should not be used for that purpose.

>>ibrarm+U1
I think you may be confusing deconvolution as the term is used in neural network literature with deconvolution as defined in mathematics/signal processing.

>>enriqu+n2
> you cannot recover the original image

Technically you are correct - you cannot recover the exact original image. The same is the true for saving an image as JPEG. But the question at hand is whether you can still recognize faces, not whether you can restore a byte-for-byte of the original. And whether JPEG or blurring, the answer is generally "yes".

It does depend on the implementation (and whether you know the implementation) how close you can get.

>>Mauran+Gi
Yes, my point is that it depends a lot on the specific blurring. If you just average a square neighborhood of 4 pixels around the center, and add no noise, it is very likely that you can "enhance" the resolution back to almost the original image. Yet, if the blur kernel is much larger (say, a gaussian of width 40 pixels), and you add some noise after te blur, it is very likely that you have completely destroyed the information.

>>leni53+(OP)
It uses StackBlur which seems to be reversible.