Commons talk:Featured picture candidates/File:Upernavik cemetary 2007-08-06 original stitch.jpg

HD width, typical monitor resolution

Merge 4×4 pixels to 1, 2.2 Mpixels, FP minimum resolution

Merge 3×3 pixels to 1, 3.9 Mpixel

Merge 2×2 pixels to 1, 8.8 Mpixel

Full resolution 35 Mpixels - maximum per pixel noise, maximum information

HD width downsampled upsampled to full resolution, massive information loss

Downsampled 50% then upsampled to full resolution. No information loss

This image was first downsized to 50% of the original size using Gimp's Cubic interpolation (6,243 x 1,407 = 8.8MP). It was then upsized to the original size, again using Gimp's Cubic interpolation (12,485 x 2,814 = 35.1MP). There are specialist software tools for upscaling images, so possibly that step can be improved. The resulting image was mildly sharpened in Lightroom for a fair comparison with the original (which would have been sharpened in-camera). Please download this and the original and view both at 100%. You will find it very difficult to tell which is the original. This test of information loss is a harsh one. It may be that the image could be reduced to smaller than 50% without losing information but upsizing an image more than 200% can lead to jaggies along high-contrast diagonal lines (which can be seen in the HD downsampling example further up).

Another test of information loss is to save the image with higher and higher JPG compression levels and compare. I assume the original was saved at 100% quality leading to a 17MB file. Compressing the image down to less than 3MB is possible without losing any significant detail. Colin (talk) 17:43, 17 December 2011 (UTC)

Discussion

There are many factors that lead to an image containing less information that one might expect given the pixel dimensions. These include

1. Imprecise focus.
2. Shallow depth of field.
3. Movement during exposure.
4. Atmospheric conditions.
5. Diffraction at small apertures.
6. Soft lens at large apertures.
7. An anti-aliasing filter in front of the sensor.
8. The Bayer colour filter used in most camera sensors.
9. A tiny sensor and/or high ISO leading to noise.
10. Aggressive noise reduction.
11. Chroma subsampling used in JPG compression.
12. JPG compression level.
13. Rotation of the image (e.g. to align stitched segments or fix a wonky horizon)
14. Stretching of the image (e.g. to align stitched segments or alter a projection)
15. Blending of images (e.g. when overlapping stitched segments)

The main argument against downsizing images is that it leads to information loss. However, that information may already have been lost and apparent detail at 100% viewing may be simply noise, JPG artefacts and the effect of sharpening -- all of which are an illusion. The above 35MP panorama is at best a 9MP panorama. Large file sizes give our users problems -- they may be unable to open such an image in their browser or image viewer; they take a long time to download; even the shortest dimension is greater than a monitor screen leading to a lot of panning about. One of the joys of a good panorama is to explore it in detail. I propose this image is a candidate for downsizing 50% without any negative effect and significant positive effects. Colin (talk) 17:43, 17 December 2011 (UTC)

Thanks for your insightful remarks and further examples. I think this is an interesting and important discussion. I agree almost completely with you with this tweak: Theoretically, even with the artifacts produced by the camera with its optics, the sensor, noise and processing algorithms, and the stitching you get the maximum information content by not downsampling. When you start downsampling, you loose information because you merge neigboring cells, in naive or more sophisticated manners (like the cubic spline) as the merge of several cells into fewer cells is an irreversible process. There is no way you came get back to the original after you have downsampled. The trick you have done above with sharpening slightly afterwards is, as you also mention yourself, not really something which brings more information back, it just moves information around in the image. It may be that the information is so spread out or coorelated before you downsample, that the amount of the information loss is very small, so small that it is hard to notice by the eye, such that for practical resons the information loss is so low that it is more than outweighed by the easier handling of a somewhat smaller file, which has almost the same information content. I believe that you downsampling to 8.8 Mpixels is a good example thereof, that the loss in information is more than outweighed by other benefits of handling, bandwidth, memory, server load, and processor load for 95% of all use cases. However, if I were to print the photo in large scale to hang in my living room, I would still prefer the full resolution, as it will lead to a (marginally) better result. But I think you are right that such panos, besides the fulll resolution also should be available in a "more reasonable" resolution, where the loss of information is minimal, and it is much easier to download and handle.

In my professional life I work with radar signal processing and how to best maximize the information and reduce the noise (or clutter) in the radar image. The discussion here has many resemblances with the problems I am faced with in the radar domain.--Slaunger (talk) 22:23, 17 December 2011 (UTC)

I appreciate your nervousness about the irreversible step of downsampling. My argument is that this image has effectively already been downsampled (by the noise reduction algorithm merging adjacent cells, for example) though left at its original size whereupon the camera applied some sharpening to make it look more detailed than it was. Also, I've found when making my own panoramas that some portions of the image get enlarged many times original.

Although this picture doesn't have much noise (left), and noise isn't being used as a reason to downsample, it is interesting to compare audio noise with this. If one records speech onto cassette tape (remember those?) the hiss can spoil one's enjoyment at playback. Applying a 10kHz filter will remove the hiss and not affect speech (though would affect other instruments) making for a better sound. One would generally want to listen to the filtered recording but keep the master tape for when CDs come along and better noise reduction techniques could be applied. Ideally, Commons would have a "for viewing on screen" image that has the ideal size for viewing in detail on screen and a "for printing" image that is closer to the original but probably looks fairly awful at 100%. The latter could even be a 16bit TIFF if that's what the photographer was working with. We'd present the former at FP. Colin (talk) 09:55, 18 December 2011 (UTC)