Upload & Sell: Off
| p.9 #19 · Post Processing Techniques |
@denoir: Interesting, although I'm not sure how relevant the test image is for real-world subjects, except for certain man made things like the blinds in morfeus' image perhaps.
Very relevant. It simply includes a bunch of spatial frequencies going from coarse detail 10 lp/mm to very fine detail 72 lp/mm. The big question is what resize & sharpen method to use in order to:
1) Keep the detail in the resized image (i.e maintaining a high MTF)
2) Keeping the internal relations of the detail constant.
3) Minimizing aliasing if possible
The test image is really a very simple case. What would be interesting would be to construct a more general one that includes multiple directions (horizontal, vertical, sagittal & tangential) with patterns of different frequencies and then try to find the transfer function (i.e sharpening method) that will yield the best results.
The current resize & sharpen method that I use is an improvised one. It took me a very long time to discover a major flaw in my previous version and there's no guarantee that the current one even approaches anything optimal. I'm getting better results than with bicubic, but that's really trivial to accomplish as bicubic resize is terrible.
I'm seeing and measuring a green cast in your version of the image, whereas the bicubic version is perfectly neutral (grayscale). Something going on with the conversion to Lab and back to RGB?
Heinz (morpheus) image or the test image? In Heinz image it's the LoCA in the blinds that gets preserved while it is lost when using bicubic. The test image is grayscale so there can't be any cast.
Btw, do you normally use "remove: Gaussian blur" for sharpening? I find the lens blur setting much more effective, but perhaps you specifically chose Gaussian for your M9 files?
I find the lens blur setting to be too aggressive - I use smart sharpen primarily to fine tune the highest frequencies (small radius) and don't want the effect to be too strong.