Learning Lens Blur Fields
blur-fields.github.io
The lens blur field is a neural representation for modelling optical blur.

cross-posted from: https://programming.dev/post/37278389

Optical blur is an inherent property of any lens system and is challenging to model in modern cameras because of their complex optical elements. To tackle this challenge, we introduce a high‑dimensional neural representation of blur—the lens blur field—and a practical method for acquisition.

The lens blur field is a multilayer perceptron (MLP) designed to (1) accurately capture variations of the lens 2‑D point spread function over image‑plane location, focus setting, and optionally depth; and (2) represent these variations parametrically as a single, sensor‑specific function. The representation models the combined effects of defocus, diffraction, aberration, and accounts for sensor features such as pixel color filters and pixel‑specific micro‑lenses.

We provide a first‑of‑its‑kind dataset of 5‑D blur fields—for smartphone cameras, camera bodies equipped with a variety of lenses, etc. Finally, we show that acquired 5‑D blur fields are expressive and accurate enough to reveal, for the first time, differences in optical behavior of smartphone devices of the same make and model.

  • icelimit@lemmy.ml
    12·
    4 days ago

    This blur persists past digital (automatic) post processing? Or is otherwise still uniquely traceable?

  • CookieOfFortune@lemmy.world
    5·
    4 days ago

    That means we could get better processing by applying this analysis in reverse. Would also reduce this type of fingerprinting.

    • Eagle0110@lemmy.worldEnglish
      2·
      3 days ago

      Uhhh I feel like reversing this wouldn’t be much easier than trying to reverse hash functions lol

      • CookieOfFortune@lemmy.world
        3·
        3 days ago

        We already apply distortion correction based on lens profiles. Probably all cell phone lenses do this. They just aren’t corrected based on individual phones.

    • interdimensionalmeme@lemmy.ml
      23·
      4 days ago

      Have a coordinated volunteer project where people print and photograph special patterned image designed to map the blur and other aberration of their particular lens. With hundreds of thousands of sample, we train a micro-distortion ML model that subtly shifts and distorts the pixels just enough to make positive lens identification impossible. Then have something to auto-apply this filter (and discard originals) on every pictures before they even have a chance of being uploaded to the cloud.

  • communism@lemmy.ml
    1·
    4 days ago

    Is it possible to use some kind of random noise algorithm to modify the image so that devices can’t be uniquely identified like this anymore? Or would that not work?

    • interdimensionalmeme@lemmy.ml
      1·
      4 days ago

      The would have to be enough to obscure the lens’ aberration, that would be an obnoxious amount of noise. Instead I think a better solution is to add micro distortion strategically to make identification ambiguous/inconclusive