OK, finally there’s one image of Wednesday, almost plainly out of DxO 5.3. It’s ISO 1600 from a Nikon D200 with Sigma 30/1.4, a combination that is supported by DxO for automatic lens correction. I think you won’t be able to tell from this resolution (even if you click on the image for the higher resolution) how much better the RAW conversion is, especially compared to Adobe Camera RAW. Nikon’s own conversions are slightly better but still inferior. I shall post some examples as we go along with a series of blog posts that will end up as a sort of review of DxO 5.3, but upfront I can already say that the examples on their website don’t lie. The quality improvement is dramatic.
They achieve this by removing noise not after demosaicing but before, directly on the RAW data. This makes sense, because a conventional Bayer sensor uses a pattern of interwoven pixels, half of them green (where the human eye is most sensitive), and a quarter each red and blue. That means that the actual resolution of the sensor for green is half of what you’d expect from the pixel count, and for red and blue it’s only a quarter each. Demosaicing is a process of interpolation, where the full resolution red, green and blue channels are reconstructed from the actual data in the reduced channels and from luminance information from the neighboring pixels of another color. Scary, huh? That’s for precision.
The only sensor type on the market that does not use such patterns at all (some by Kodak and Fuji use different patters or variations) is the Foveon sensor used in Sigma’s cameras, although at the price of much reduced actual pixel resolution. Those pixels upscale well, but an image resolution of 2640×1760 (4.6 megapixels) does not look so sexy today. That impression is further marred by a light sensitivity that’s slightly below today’s standards.
But let’s get back to Bayer sensors and the process of demosaicing. We’ve seen that the actual image is reconstructed from greatly incomplete data. I’d even say much of it is invented. Of course this technology is not new and the algorithms are fairly mature, but it is also clear that at high ISO noise starts to spread out. When neighboring pixels are used to reconstruct the exact color channels per pixel, then noise in those neighboring pixels comes into the equation. This sums up, and the result are ugly blotches of color or strong de-coloration as a consequence of noise reduction algorithms.
This is where DxO sets in. They reduce noise on the original pixels before they get combined in demosaicing. Therefore the noise in one pixel does not spread out to neighboring pixels. What exactly they do and how they do it is not known to me, is most probably patented or secret, but the results are clearly better than everything else that I’ve seen so far, producing very fine grain without much loss of detail and without much loss of color at high ISOs.
It’s really like a camera upgrade, i.e. the improvement is about the amount that you’d expect by going from one camera generation to the next. The Imaging Resource has quite a lengthy interview where these things are explained and where a D700 image is shown, taken at ISO 6400, underexposed by 3 EV and pushed to 51,200 equivalent. Crazy? Yes, but surprisingly there is still usable output.
The Song of the Day is “Tell Me You’ll Wait For Me” from Ray Charles’ 1959 album “The Genius of Ray Charles”. Sorry no video, but Amazon has samples.