JPG vs PNG Test became JPG Resizing Test - a temporary topic

[Andrea B.]

And now I've heard from the Internet about calling the bunny dumb. No, politically correct Internet, I was calling myself stupid.

[Andrea B.]

Attached here is the photo which started me on this little misadventure. The CC Passport was photographed to use in making a color profile in Photo Ninja. I wanted to post it to show how sharp the lens is. But every way I resized it made the photo appear not so sharp. So now I will post it as an unresized JPG saved at some quality below 100.

 

D610 + Coastal Optics 105/4.5 + Baader UV/IR-Cut

Converted with application of new color profile in Photo Ninja.

 

The lens is so sharp it picked up the graininess of the paint and the texture of the scratches and the multi-hued reflections of the pebbled surface of the Passport black plastic cover.

[dabateman]

I like your flower test image. Are they still around?

Next would be to test the CO 60, UAT, Nikor 105mm and CO 105mm. Use the same framing. That should be easy as 2 of them are macro.

Then we can see the sharpest Quartz lens.

[Andy Perrin]

Guys, you have to denoise BEFORE you do a nearest neighbor or every frequency over the Nyquist limit will alias and make artifacts.

[Andrea B.]

Yes, but it always weirds me out to soften a photo before resampling it and then resharpening it.

 

And I don't think that is what the forum software is doing. How do they get such a good resize?????

[Andy Perrin]

Do you care or do you just want it to look good? I don’t think we can reverse engineer it by trying random things. I don’t know why denoising first bothers you, because you are throwing away exactly the frequencies that cause the artifacts. It’s the same as the in-camera antialiasing filters, but in software.

[Andrea B.]

Yes, I am being irrational about it!!

I'll try a couple of things.

 

PSE: added Gaussian Blur with 0.5 pixel width to remove some speckles and a bit of noise.

Then resized to 1021 px width with Bicubic Plain.

 

 

[Andrea B.]

PSE: added Gaussian Blur with 0.5 pixel width to remove some speckles and a bit of noise.

Then resized to 1021 px width with Nearest Neighbor.

 

I might not have done enough denoising. I don't really have a good denoiser in PSE.

 

 

That didn't quite work. The curves are a bit bumpety.

Better in some respects than previous.

[Andrea B.]

OK. I need to move on and get some other work done today.

I will continue to try to figure out how best to resize photos for posting here. None of my resizes above look really bad, but they don't look as good as what the forum software could do. So at this point we have the strange conclusion that the best posted photo comes from uploading a JPG saved at 50% quality and resized by the forum software.

 

As a one time deal, I will upload the pink flower TIF so that others can try resizing it for maximum clarity and detail. I can't leave this TIF posted for too long. It's a big hit on the server. But it will be very interesting to see what can be done with other software than what I have.

 

 

Note to Self: Try to find out from Invision devs how they do this.

 

 

hydrangeas.tif

[Andy Perrin]

Andrea, how come your little crops look bigger than they are in the resized photos that they came from?

[Andrea B.]

beats me!

 

I used Macbook Command-Shift-4 which lets you drag over what you want a screen shot of.

It is probably a resolution thing?

 

Let me see if I can figure it out.

[Andrea B.]

My Macbook screen is set to "best for Retina" which is pixel-doubled and 1440 x 900.

 

So when I use Command + to make the browser as magnified as possible,

then I get a larger screen shot than

when I use Command - to make the browser display as shrunken as possible.

 

Here are those screen shots. They both depend on the fact that I'm using 1440 x 900 for the monitor.

 

The largest possible screen shot of the desired area.

 

The smallest possible.

 

 

So now I'll set the Macbook screen to the maximum possible which is, I think?, 1900 x 1200.

And repeat the screen shots.

 

Largest possible.

 

Smallest possible.

 

On the higher res screen the largest possible screen shot didn't change, but the smallest possible screen shot was larger.

Are we all confused now? I certainly am.

[Andy Perrin]

Okay, the answer is, you should not try to make these comparisons using screen shots that depend on your resolution because when the rest of us see them, they don't match the shots they came from. Which matters a lot here! Every time it changes size, some additional processing is happening. So we should be cropping inside photoshop or PN or whatever using the images we take the crops from.

[Andrea B.]

Gotcha. Will do.

 

However, I don't think that it will change the basic result?

 

BRB

[Andrea B.]

Here is an unresized crop from the original unresized photo. The crop was made in Photo Mechanic. It is 661 x 504 pixels.

 

(Please forum software do not smite this crop and resize it!)

 

I just wanted to do this for my own interest. Now I'll go get crops from the 1021 sized photos.

 

[Andrea B.]

PSE resize with Bicubic (plain) to 1021 px width.

 

 

 

Left: Unresized crop made in PSE on 100% version of photo shown here.

Right: Crop shown in preceding post resized downward. Still shows more detail.

[Andrea B.]

There's probably no point in making crops at all if they are just the same size as the reference area in the displayed photo. You can just look at the photo.

It was the enlarged versions of the crops which brought home the point that the forum was making better resizes than I could in NX2, PSE or Photo Ninja.

 

 

 

At the end of this effort today all I can say is that the CO60 on the D810 makes an amazingly detailed photo.

[Andy Perrin]

There's probably no point in making crops at all if they are just the same size as the reference area in the displayed photo. You can just look at the photo.

I thought the point was to show them side by side with the forum one, so we could see the difference without scrolling all around?

[Andrea B.]

Yes, there's that.

 

But I've probably had enough of resizing pink flowers today and should give it a rest.

 

I need to go look at one of the Windows machines around here. Maybe I'm being misled by pixel-doubling thing? Maybe the photos don't look all so different on other screens?

[dabateman]

Yes, there's that.

 

But I've probably had enough of resizing pink flowers today and should give it a rest.

 

Yes clearly time to switch to a blue flower.

 

Joking aside.

Are you sure you are comparing apples to apples. When I crop to 800x600 my file sizes are about 50 to 80 kilobytes. Your upload images are about 1Mbyte. That could be the difference. The forum is not resizeing them but interpolating the data differently.

 

[Andrea B.]

Photo resized to 800 px width and saved at 50% quality.

261.02 KB.

 

 

 

Added some detail slider.

Now at 396.65 KB. Oops!

 

[Andy Perrin]

Here's my shot.

 

Neat Image for denoising (but only the highest frequencies, did not touch the lower ones), followed by bilinear reduction to 800 pixels width, and then Smart Deblur for sharpening.

[Andrea B.]

Andy, thank you for your version.

 

Question: Because resizing reduces noise, how do you decide whether to denoise before or after resizing?

 

Also, does Neat Image denoiser permit the frequency choice or did you apply some kind of masking?

[Andy Perrin]

- Resizing to smaller sizes in general INCREASES high frequencies (which often includes the noise). You can see this especially with nearest neighbor resizes, which are the simplest algorithm. Edges get sharper. However any frequencies over the Nyquist limit will reappear as lower frequency artifacts. (Here is a

that explains it with a 1D audio signal. It's easier to understand in 1D than 2D.)

 

- Other forms of interpolation like bilinear and bicubic incorporate some degree of smoothing as part of the algorithm, which is why they can appear to smooth the image. They often overdo it, leaving the image too soft.

 

- Therefore removing the noise ahead of resizing is the best choice because then you have control over which frequencies are being discarded rather than leaving it to Photoshop or whatever your program is. You should use nearest neighbor or bilinear, not bicubic anything.

 

- Neat Image allows you to select which frequencies are damped. Generally it’s best to leave all but the highest untouched.

 

- Finally, removing noise after resizing is pointless because any frequencies over the Nyquist will already have damaged your image irreversibly. It’s closing the barn door after the horse has left.

[Andrea B.]

Thanks for the explanation!

 

I think there is a *lot* of misinformation about resizing out there in internet-land. Even PSE labels bicubic sharper as "best for reduction".

 

I still need to learn how the forum manages such a detailed resize. It's quite amazing.