Does the S1 produce sharper images than the S1R?

[scott kirkpatrick]

Advertisement (gone after registration)

To get from 8 exposures with pixel-shifting to 4 times the original number of pixels requires a lot of processing, partly to sharpen the overlapping shifted sensor sites, and partly because the array of 4-exposure de-Bayered sites that the shifting produces are in an array that is oriented at 45 degrees to the original sensor x and y coordinates.  The 45 degree rotation doesn't sound too difficult (it's easy to do with a JPEG), but if the shifts aren't precise or the camera and background move a little while the shots are being collected there will be smearing and the algorithms intended to catch and correct this might make it worse.  Hasselblad did this very early, and only used four shifts.  Olympus has been doing it longest among the cameras that I've used, and they produce JPEGs with only 2.5 to 3X as many pixels, rather than claiming the full 4X.  Highly enhanced resolution with fewer sensors isn't magic.

Edited October 9, 2019 by scott kirkpatrick

[scott kirkpatrick]

As challenged, I shot an identical scene (flat wall of books from about 6 m distance, indoors since it is Yom Kipur and cars don't even start on that day here). with ISOs from L50 up to 800, every 1/3 stop. I put the camera on a tabletop tripod and used Lumix Synch to make adjustments and to shoot. Here's the scene.  With irrelevant detail cropped, it is 5172 b by 3146 pixels (about 18 MPx):  

Welcome, dear visitor! As registered member you'd see an image here…

Simply register for free here – We are always happy to welcome new members!

P1022439 by scott kirkpatrick, on Flickr S1R with SL 35 @/2.0    ISO L50

and here is a crop small enough that it ought to show up at actual size  (1325 x 865):

P1022439 copy by scott kirkpatrick, on Flickr

I really can't ask more than this.  Every bit of text on the books is resolved.  I ran through all the examples at actual size on my editing screen and saw no differences in resolution, only a small amount of added noise in the shadows at ISO 800.  If the extra 1 stop of read noise reduction which Claff shows was obtained by averaging locally, that one stop would require four pixels to be combined.  I see no evidence of something like that.  So the tricks that Panasonic is using must be down beneath the black point somewhere.

 

Edited October 8, 2019 by scott kirkpatrick

[Advertisement]

The Leica Q3 is available for pre-order at B&H Photo Video and Adorama

[scott kirkpatrick]

Here's a comparison between the shot at 200 ISO and the HD shot at the same ISO (the camera saves its first exposure of the series), cropped so that each should display at actual size:

regular:

Welcome, dear visitor! As registered member you'd see an image here…

Simply register for free here – We are always happy to welcome new members!

P1022462 copy by scott kirkpatrick, on Flickr

and HD superresolution:

P1022461 copy by scott kirkpatrick, on Flickr

I haven't sharpened the HD version, which would make the midtone contrast come up a bit so that the two images match better. Note that Bulfinch and Artisan are clear in the second shot, and you can guess that in between is Holt Rinehart and Winston.  But HD is not really twice the resolution, at least not without some further effort.

Edited October 9, 2019 by scott kirkpatrick

[Stuart Richardson]

Yes, and do you see that odd aliasing around the letters...that is the pattern I was mentioning. I have seen it in the landscapes I have taken. I do not think it is the camera being moved, I think it is inherent in the way that the camera combines the images. 

[scott kirkpatrick]

2 hours ago, Stuart Richardson said:

Yes, and do you see that odd aliasing around the letters...that is the pattern I was mentioning. I have seen it in the landscapes I have taken. I do not think it is the camera being moved, I think it is inherent in the way that the camera combines the images. 

Is that really pattern noise or just the result of the actual pixels being larger than the spaces they are mapped into, thus needing some post-sharpening?  I have't used the HR outdoors because of leaves and grass moving.  But it should be good for rocks and ice (but maybe not clouds).

Algorithmically, this is tricky stuff, probably proprietary, and thus each manufacture has to learn some things by trial and error.  Panasonic, for example, offers a correction for moving leaves and grass.  I would imagine that it works well only some of the time and fails in funny ways at other times.  Olympus has been doing this the longest.  Do they show the same effects?

[scott kirkpatrick]

Hmmm.  I think I see what you are looking at.  Download this tiny section

Welcome, dear visitor! As registered member you'd see an image here…

Simply register for free here – We are always happy to welcome new members!

P1022461 copy(2) by scott kirkpatrick, on Flickr

and then blow it up 4 or 8 times actual size: You get this:

Screen Shot 2019-10-14 at 2.45.41 PM by scott kirkpatrick, on Flickr

I think the row of white dots at the edge of the text is the result of moving the two 4-Bayer shot pixels that result from the 8 frames that HD takes into four locations, doubling the resolution of the original grid.  There are probably better ways to do it.

edit:  I believe Hasselblad takes only four shots in one of their HR modes, eliminating the artifacts of de-mosaicing but leaving the same number of pixels.  That seems to fit the spirit of "the essentials" better than what you see in this example.

Edited October 14, 2019 by scott kirkpatrick

[Stuart Richardson]

Advertisement (gone after registration)

Yes, that is what I am referring to. I am attaching two huge crops from a picture I took recently. These are 2 to 1 or 3 to 1 views of the fine detail of a high rez mode shot versus a standard resolution shot. In this case, the false color and strange textures are more visible in the high resolution version than in the standard. This was taken on a good tripod with the S1 and the 180mm APO Elmarit S at f5.6, where it should have enough resolution to outdo the sensor. The crops are from near the center of the frame. This is more or less what I was referring to above...pixel shifting is not really a replacement for true resolution, at least not in this implementation.

Welcome, dear visitor! As registered member you'd see an image here…

Simply register for free here – We are always happy to welcome new members!

[Stuart Richardson]

P.S. These photos look even worse due to the jpeg compression and size changes, so just be aware...that said, the false colors and staircase patterning in certain areas are representative of the original file. I would post the overall picture, but I ran out of space with the paltry 500kb a day.

[scott kirkpatrick]

I did a quick check of what Google can find about how different cameras go about using 4 or 8 shots to get higher resolution.  Imaging resource reviewed most of the contenders about 4 years ago, B&H posted a review article, and there was one blog post.  Both Olympus and Panasonic's G-9 used 8 shot technology.  Pentax K-1 and some others used four shot.  Each article explains the idea that with four shots spaced one pixelsite apart you can collect all four Bayer-filtered exposures and relate that to a single point in the finished image.  Nobody seems to know where the next four shots are taken, but the most plausible answer is this pattern:

X           X

      O           O

X           X

      O            O

with the X's indicating the first four shots and the O's the second set.  But that gives us only twice as many pixel sites, not arranged like the original set but with complete color information for each site.  Olympus then maps these into 2.5 times as many sites as in the original image (in a JPEG) and three times as many in a raw file.  In the S1R and S1, Panasonic is giving us 4 times as many pixel sites in a final raw image, and doesn't produce a JPG..  But the resulting raw files are processed by existing software, which means that the software thinks it is getting a traditional Bayer-ized raw file with four times as many pixels, each pixel carrying only R, G, B or G information.  This means for compatibility the nicely collected color information has to be spread around in the finished HR image.  No wonder there can be color artifacts as well as the patterns of light and dark that I show above near a horizontal black-white edge.

     

Edited October 15, 2019 by scott kirkpatrick

[Stuart Richardson]

Once again, there is no free lunch...still, it is a great tool to have in the toolbox if you have a camera with sensor stabilization. They can do it, so they might as well, and chances are it will only get better over time and throughout the years. As I said above, I think there are a number of situations where the artifacts it creates are more than offset by the extra resolution. But if I had the choice of a 100mp pixel shifted camera or a 50-65mp native resolution camera, with all else being equal, I would take the 50-65mp camera. I think the overall image quality is likely to be greater.