SuperResolution ~ FINALLY got it to work!

Gawds!  Will it ever end?  I have a seemingly bottomless well of curiosity that I dip into when photo-opportunities are slim.  

When I revisited something I'd read some time back, it triggered another round of trying to come to grips with SuperResolution.  I was curious to see if I could get the kind of results people have for years talked about.

Carefully re-reading the DPReview page I realized I could have a closer look at the up-size operators in the Gimp.  Early on I'd used the "cubic" upsize method and found it gave slightly "fuzzy" results in both 2x and 4x area image upsizing.  

Then I found a newer method called "nohalo".  It seemed to work fairly well in 2x area image upsizing, but lacked a bit of "crispness" in 4x area stacked images.  The DPReview article talked of using a "linear" method.  It turns out that the Gimp has that, too.

So, I grabbed an old Sony NEX-7 24mpixel camera that a friend very kindly gave me, put on the sharpest APS-C lens I own (a Sigma 30mm f/2.8 EX DN), set the shutter to "splatter" mode (ie: 10fps), and wandered the streets looking for an "interesting" subject.  A local pizza restaurant seemed to have enough detail of the kind that might lend itself to SuperResolution.

Here is the process/method I used.

Camera -

• Sony NEX-7
• Sigma 30mm f/2.8 EX DN
• Set to "Program" mode
• Set the ISO to 100
  
• Set the shutter to 10fps (focus once, shoot many)

Processing 1 - RawTherapee

• Select four individual images from the "splatter-fest"
  
• Process using the exact same parameters
• Color Management set to "Automatic"
• Color Management "curves" selected
• Lens distortion set to "Automatic"
• Lens distortion "vignetting" de-selected
• Capture Sharpen selected
• Set contrast and Saturation to taste

Processing 2 - Gimp

• Import four images as layers
• Upsize using either "linear" or "nohalo" methods
• Hand align layers *see Note
• Set Opacity of each layer to blend
• Base layer - 100%
• Layer 1 just above the Base - 50%
• Layer 2 just above Layer 1 - 33%
• Layer 3 just above layer 2 - 27%
   
• Flatten image
  
• Duplicate image as another layer and sharpen
• Viewed at 100 or 200percent, set Opacity of sharpening layer to remove possible "over-sharpening" effects

*Note:  Hand aligning layers in the Gimp is somewhat easy.  Using the Base Layer as the reference "registration" layer...

• Keep the Base Layer Opacity at 100%
• Making visible one layer at a time...
• Set layer Opacity to 50% on the layer being aligned so as to see how two images align compared to the base layer
  
• Viewed at 400% and in one corner find an image detail, select the move operator and align the layer over the base exactly using the detail reference point
  
• Select the rotation operation for the layer being worked on
  
• Set the rotation point directly over the newly aligned reference detail
  
• In the opposite corner of the image...
• View at 400%
• Click and rotate the Layer to align another reference detail over the Base Layer
• Deselect the Layer just worked on and repeat this process on another Layer 

Results -

Here is the base scene followed by the results

[Best viewed at 100%] 

 

Comments - 

The SuperResolution process does, in fact, work.  In the past I simply wasn't working carefully enough.  I'd either relied too heavily on the Hugin align image operator, or I wasn't zooming in far enough to see how details were aligning.

I am reminded that this SuperResolution technique does more than add resolution.  It does something that I find to be just as important.  The process averages out noise.  Aligned images are clean and smooth.  Here is what I mean.

 

 

For this little study I deliberately used the early Sony NEX-7 24mpixel camera because it has stronger base ISO noise than the A6000 and A6300.  So I was pleasantly surprised to re-learn the potential usefulness of eliminating noise by stacking and blending images.

One interesting thing about re-working this process is that I think I found the "Linear" ups-size function is ever so very slightly "sharper" after stacking/blending images than "NoHalo."  But I think I have to stand on one toe while the full moon shines and sing three choruses of "Hair" before I'd fully swear by that.  All things being equal, at this point I feel either up-size operator can do the required work and yield pleasing results.

I'm not ready to downsize from 24mpixel sensors just yet.  Though I had thought seriously about picking up a Sony A7S, just for grins.  Northlight has an interesting article about processing an old 11mpixel Canon image that "had me going" for a moment or two.  And that image wasn't stacked/blended, either.  Pretty incredible work, that.

No, I think 24mpixel is kind of a "sweet spot" for me and the kind of work I do and for the kinds of things I enjoy.  It's not too heavy on the computer and RAW files don't take up an enormous amount of space.  And if I wanted to make a huge print of, say, a classic automobile and if I had the luxury of planning ahead, it might be interesting to try image stacking/blending and maybe even use this SuperResolution technique.