Impressive New Noise Reduction Technique

[chanyr]

Advertisement (gone after registration)

Someone has come up with a method to create noiseless photos at the RAW level. He promises to release a version of dcraw that will do this, this month. Here's a link to an article he wrote, GUILLERMO LUIJK >> ARTICLES >> ZERO NOISE PHOTOGRAPHY , and a forum discussion, Luminous Landscape Forum -> ZERO NOISE technique

 

Is anyone else following this? If this works with the M8, the result is going to be awesome. In effect, the result makes his Canon EOS-350D look like a digital back.

[wstotler]

If this works with the M8, the result is going to be awesome.

 

I was incredibly excited until I read this caveat from the article: "To do this you simply need to shoot twice making use of a tripod. One shot will be as usual, keeping highlights unburnt. The second shot with be done with a severe overexposition."

 

Nice technique for any studio-type work but not so helpful for handheld, low-light (and low speed) situations, where it would be *very* helpful.

 

Thanks for posting the articles! They're well written and will add another item to my toolbox for pano work.

 

Thanks,

Will

[chanyr]

Yes, unfortunately, it requires the use of a tripod and two exposures. So this kinda rules out handheld work. It would work well with deliberate work on a tripod .

 

I wonder if the M8 can be modified to work this way in future if this method proves useful. The M8 is quite re-programmable from what I saw in Mark's M8 tear down article.

[jaapv]

Well, it can do a dark exposure, so why not a second overexposed one. It would be an interesting feature and well in line with the intended use of the camera.

[Guest sirvine]

I'm pretty sure you can acheive this effect with PS3 already. Every time someone challenges him on this point, his response comes back really, really vague.

 

This seems to be just a variant of "blend if" or HDR/tonemap features. The problem is that everyone and their brother thinks HDR means applying the same exposure values to every point in the image (which looks like crap, IMO). But really, HDR is just a method of substituting a properly exposed low or high end, as appropriate to extend the camera's effective DR--with the result that noise is eliminated from the shadows.

[Philinflash]

And then, there is HDR.

 

All of these techniques require some sort of bracketing which is never a bad idea. But it is hard to do.

 

However, there is no automatic bracketing available in the M8. Whilst I recognize there is a lot of resistance to the addition of "bells and whistles," I cannot help but think this would be an appropriate software addition for our cameras in light of its potential real-world usefulness.

 

Philip Kozloff

[joachimeh]

Advertisement (gone after registration)

And then, there is HDR.

 

All of these techniques require some sort of bracketing which is never a bad idea. But it is hard to do.

 

However, there is no automatic bracketing available in the M8. Whilst I recognize there is a lot of resistance to the addition of "bells and whistles," I cannot help but think this would be an appropriate software addition for our cameras in light of its potential real-world usefulness.

 

Philip Kozloff

 

With a built-in prism dividing the original image into 2 identical ones (with one being darkend) the sensor would get (in 2 different sections) 2 absolute identical images, and since noise is random across the chip this method could automatically be applied in camera's software to select pixel by pixel the best noise-to-signal ratio, or stored in 2 different RAW files for offline processing.

With e.g. a 16 MPx chip there would be two 8 Mpx sources, and since texture is affected the result would be a crisp, noise free image. Let's look forward what camera engineers will do with that method, it might not be far off.

[andybarton]

I'm pretty sure you can acheive this effect with PS3 already. Every time someone challenges him on this point, his response comes back really, really vague.

 

This seems to be just a variant of "blend if" or HDR/tonemap features. The problem is that everyone and their brother thinks HDR means applying the same exposure values to every point in the image (which looks like crap, IMO). But really, HDR is just a method of substituting a properly exposed low or high end, as appropriate to extend the camera's effective DR--with the result that noise is eliminated from the shadows.

 

You can do this with film and a scanner.

 

Scan for the highlights on one scan, for the lowlights on a scecond scan, and then blend the two in PS.

 

In fact, I posted a shot here using this technique only the other day.

[Alberti]

With a built-in prism dividing the original image into 2 identical ones (with one being darkend) the sensor would get (in 2 different sections) 2 absolute identical images, and since noise is random across the chip this method could automatically be applied in camera's software to select pixel by pixel the best noise-to-signal ratio, or stored in 2 different RAW files for offline processing.

With e.g. a 16 MPx chip there would be two 8 Mpx sources, and since texture is affected the result would be a crisp, noise free image. Let's look forward what camera engineers will do with that method, it might not be far off.

 

Now we are inventing, why not a co-located pixel source, the chip providing two sets of transistors, whith similarCCD-bucket sizes but one with a pure gray filter in front. Of course it will not be transparent on the whole spectrum and if the M8 chip alrady was made so thin that an UV didn't fit, I guess this is hard.

Alternatively, I have seen a Fuiji (?) chip idea that had the co-location, but with two small lenses, the one being large the other small, so differences in sensitivity arise automatically, and a summing amplifier produces a long slope.

 

I heard that the chip of the M8 originally was 21 Mp Prototype) , now it is 10,5 so I figured that already the chip does a trick like summing to achieve the longer scale we have.

Next step is a FF chip, still 10,5 Mp format, but now with even more summing.... going to near-zeroKelvin noise in the blacks.

[carstenw]

This technique appears to be a two-image HDR with no tonemapping. Photomatix Pro already does so much more than this.

[chanyr]

I was thinking about how this can be done in hardware. The method is to take two exposures. In a way, what is happening is that pixels are amplified at two different levels (two exposures). The best result is then selected from each exposure (final picture is selected pixels amplified at different levels).

 

The way to do this in hardware is to be able to have each pixel amplify the results they're getting individually. I think this is not possible on a CCD chip where if you have a two channel readout you have two amplifiers but this is possible in a CMOS chip where each pixel has individual amplifiers. Thinking about it, this is probably how Canon captures really low noise images with CMOS chips, by individually adjusting the amplification for each pixel.