Tuesday, March 22, 2016

Basler Improves on Demosaicing Algorithms

Jörg Kunze from Basler AG kindly allowed me to publish few slides from his presentation at the London Image Sensors 2016 conference about his novel Debayering algorithm called PGI.

His implementation is a hardware-efficient single-step 5x5 pixel algorithm, which performs a zipper-free high-quality Bayer-pattern interpolation up to the theoretical frequency limit, color-anti-aliasing, sharpness enhancement and noise reduction together. The pictures look very convincing. Basler has currently a single-lane FPGA implementation with a throughput of 140 MPix/s using 880 Cyclone V logic cells and a quad lane implementation with a throughput of 400 MPix/s using 2600 logic cells. Jörg says, Basler is interested in licensing, cross-licensing or technology exchange.

3 comments:

  1. To adequately show the relative benefits of Bayer demosaicing algorithms, it is necessary to show color bars, not monochrome. What do the alternating red/blue and red/green bars look like near the resolution limit?

    ReplyDelete
    Replies
    1. Hi Dave,

      I completely understand your point of view.

      As I explained in my London talk, my point of view is in this case different. I want a very good visible image quality for a human at a very low computation effort. So I never intended to push red/blue and red/green to their respective resolution limit, because a human eye cannot resolve it either. A human eye resolves bright and dark structures with high resolution. So I deliver the maximum possible B/W high frequency reconstruction inside a low-noise color image and this is what I show.

      Delete
  2. Great work Jorg!
    Is there any quantified improvement, e.g. in MTF50?

    ReplyDelete

All comments are moderated to avoid spam.