Upsampled Conversion - An Effective, Practical Technique

[Noel_Carboni]

From time to time I'll try to post interesting, effective, and/or just plain gee whiz techniques here.

Today's entry: Raw Conversion to an Upsampled Resolution

We've got supercomputers nowadays. Let's use them to get better results:

When converting a raw image with Adobe Camera Raw engine (e.g., using Photoshop, or Lightroom) consider converting to an upsampled resolution.

There is nothing that demands your image be converted to the pixel dimensions of your camera's imager, though that's usually the default. In fact, converting to an upsampled resolution can actually deliver a bit more (undistorted) detail in the result and make for better data on which to do subsequent editing.

In Photoshop's Camera Raw converter, here's where you find the menu to choose the size:



For example, for my 10.1 MP Canon EOS-40D, I typically choose an upsampled resolution of 3413 x 5120 pixels (17.5 megapixels):



Photoshop will remember this setting for your next conversion once you've done one this way.

If you've got sufficient computer resources to handle the larger file size, the resulting image data turns out to be VERY good input for editing.

Try it and let us know what you think.

-Noel

[Doug_Kerr]

Hi, Noel,

Wouild you care to suggest a "theoretical" (even, "conceptual") premise for the improved performance evidently experienced with this approach?

In audio processing, we use "oversampling" to allow us to avoid the need for an analog low pass reconstruction filter with a cutoff at about the Nyquist frequency, and instead construct a digital low-pass reconstruction filter (which then needs its own analog reconstruction filter, but at a significantly higher cutoff frequency, which makes its implementation easier).

In disrete imaging, we don't often have an overt low pass reconstruction filter, but rely on such things as the Gaussian spot profile in a display CRT to produce the same function.

So that notion of oversampling doesn't apply.

[Noel_Carboni]

The short answer: If you believe a raw image contains more information than a converted image in general (of course, this is debated from time to time), this allows the raw converter a way to save information that might otherwise be lost.

Also, processing an image that's got too many pixels (i.e., a fuzzy image) can yield fewer artifacts than processing one that's stuffed to the gills with detail.

I find converting to an upsampled size, processing, sharpening (using sophisticated techniques) can yield a very good looking and sharp upsampled image. Perhaps the sharpness is fooling the eye, but it just looks good.

-Noel

[Doug_Kerr]

Hi, Noel,

Originally Posted by Noel_Carboni

The short answer: If you believe a raw image contains more information than a converted image in general (of course, this is debated from time to time)

Well, indeed, the true-raw dataset (for a CFA-array sensor system) contains more information that a demosiaced image derived from it.

But I'm not sure that upsampling the demosaiced image (if that is indeed what the process you described does - and I have to review how you described it to make sure I'm not jumping to the wrong conclusion) would "conserve" this higher-information-content nature.

Perhaps it is the still-mosaiced ("raw") data set that is upsampled. That would increase the amount of information in the "bit-count" sense, but not the amount in the true information theory sense. It is not as if, for example, we are actually sampling the image on the sensor at a finer pitch.

Also, processing an image that's got too many pixels (i.e., a fuzzy image) can yield fewer artifacts than processing one that's stuffed to the gills with detail.

Even though I don't have a good conceptual model of that in my mind at the moment, it certainly sounds reasonable.

I appreciate your reminding us of this interesting approach.

Best regards,

Doug

[Noel_Carboni]

I really don't know how the algorithms inside Camera Raw do the job; perhaps they scan the image in terms of the output pixels, drawing on the input data as needed to render each output pixel. It seems clear it's doing more than just a 1:1 photosite to pixel conversion then upsizing the result.

I have to admit I was a bit surprised to find that this approach actually did deliver better data at first, but in practical terms it's a bit like installing a larger sensor in the same camera body. The more sophisticated sharpening techniques seem to be able to "bite" on the upsampled data and return a decent looking result, and the quality of the result is my gauge for finding the best techniques.

-Noel

[larry_angier]

Hi Noel and all,

I've been using this technique successfully for many years.

Though PS does a very good job in up-rezzing the image, I figure that going straight to the horse's mouth and doing it at the start, I have an advantage time wise and quality wise, however slight.

Theoretically, it would seem that starting with the image at a larger size seems to make sense for larger prints, especially since one is using the data at its source. It just seems that this approach is the better way to go, right or wrong.

However, In practice and once the photograph is on the wall, I don't think it will make that much difference if you start at the beginning or at the end unless you are near-sighted like me and look at your 3x5 prints at just a few inches or a foot-or-two away.

For me, I just want to make the image at the largest I will ever use and be done with it as a master image and then resize downward from that file as I repurpose the image for print, web, smaller prints.

For efficiency, I tend to work once then resize downward from the master image.

[Dennis_Vied]

I usually use LR to open my raw images, and I know it uses ACR, so I went into LR to see if that capability is available to open (import) the images at a higher resolution, and so far, I haven't found it.

It's an interesting concept, but I wonder if it is really getting something for nothing. When you upsample in PS, you can then do the same processing, it seems to me, except that you are arguably not upsampling from the raw data now, you're upsampling from the converted file. The question is, does it make a difference, and you obviously think it does. I'm not one to argue with results.

Another factor to consider is that when you open a 12 bit raw file in 16 bit mode, that, in effect, is an upsampling as data has to be created to stuff the 16 bit file. And that is what upsampling is, creating data where none existed before. I have no argument with the fact that this produces better results, but I believe the upsampling is creating data where none existed before, rather than extracting more data from a raw file than is possible with native resolution. In other words, I don't think you're getting something for nothing.