Rotation and perspective - resampling

[puttick]

If you rotate an image by a small arbitrary amount (say 1 degree to correct a horizontal), there is inevitably some resampling without a change of image size. Or you might make a small perspective correction on converging verticals, which again involves resampling to expand or contract certain areas relative to each other.

Does this affect image quality? If so, how much?

I've noticed that Photoshop won't let you correct perspective in 16-bit mode. Is this relevant?

[Robert_Lay]

Dear Dr. Puttick,

Using Photoshop CS, I just performed two experiments - one doing arbitrary angular rotations and the other on perspective. In both experiments I started with a 16 bit TIF image.

I did a Rotate Canvas of 1 degree three times in succession. I then did a 3 degree rotation in the opposite direction. After zooming in to the maximum (1600% - where individual pixels are about 1/8" square), I could see absolutely no difference with respect to the unmodified file.

For the second experiment, I followed the instructions for doing a perspective correction on a 16 bit TIF image and everything worked exactly as the Help file said it should.

In conclusion, I would respond to your first issue ("resampling") with a "not to the best of my knowledge".

In regard to your second issue (perspective correction with 16 bit mode), I suggest you try it again and tell us what did not happen relative to your expectations.

[puttick]

Thanks Robert.

Taking your 2nd point first - I am using PS7, which will not perform perspective correction on 16-bit images. I was wondering whether this was for a good reason, perhaps to do with image quality. I did not know that CS allows such correction.

However, consider this example which I have just set up (note this is PS7). I start with a 3000 x 2000 image, at 180dpi. Then I perform a perspective correction on it by moving each of the top corner "handles" in by about 10%. After the image has been "reshaped" back to a rectangle, there must have been resampling in order for the upper part of the image to be stretched laterally to match the lower part. If I now look at the image size, I find it is 3015 x 2010, at 181dpi, and the image' actual output size has not changed. Clearly resampling has taken place, as the image now has 5% more pixels. Now some of these will not lie exactly where their "parents" were, and so some interpolation must have happened. I cannot see any obvious artefacts (jagged diagonals etc). If I correct the resolution back to 180 dpi the image size goes back to 3000 x 2000, but obviously the stretching referred to above has been carried out.

Similarly if you rotate an image, the information contained in some original pixels will have to be remapped (moved) more than others, as you move further away from the centre along the radius. This will in most instances not be by a whole pixel, so interpolation will be needed. In your example, you do not say whether you saved between each rotation. If you did not save, then presumably PS is just undoing what it had done in the previous rotations. I wonder if you saved every time, then reversed the process, saving every time, you might introduce some artefacts.

I find this fascinating. However maybe it's irrelevant, or the effects not visible, unless printing at very large output sizes.

Any other forum members wish to comment?

Thanks
Nigel

[Robert_Lay]

First, I also have PS 7.0.1, so I will check some of these issues in that environment as well.

I can only guess as to what the software does from one step or "action" to the next. However, looking at the concept of "History" and recognizing the extent to which history is preserved and can be repeated, anything is possible.

My gut feeling, having done some programming in the area of image file processing, is that all such translations (rotation, perspective correction, etc.) are done without invoking a "re-sampling". I say that because I think it is only fair to expect "re-sampling" to occur when there is a re-sizing (a significant change in the total number of pixels) called for.

In other words, I interpret re-sampling quite differently than the fairly limited interpolation and/or transformations that must be employed in rotations and perspective corrections.

No, I did NOT do any explicit saves between any steps.

[Robert_Lay]

Just completed the same tests using Photoshop 7.0.1.

As you said, the perspective correction does not seem to be functional in 16 bit Mode - the perspective check box is not even enabled.

The rotation test behavior corresponds with what I got in PSCS - there is no indication of any degradation after a series of rotational steps.

As I indicated previously, I take the "resampling" issue as different from those interpolations that occur during various transformations. It's muddy water at best, but the expressions re-sample, down sample and up-sample have always meant specific things for a specific purpose, and I don't see it being relevant to what we are talking about here.

Just doing some rotations or some perspective corrections will invoke the amount of interpolations necessary to the respective transformation. I truly believe that in the grand scheme of things, especially when the final product is to make a print, that you will NOT see any evidence of these interpolations - especially in regard to rotations. With perspective correction, I can imagine that you would be able to detect some sort of posterization occurring for LARGE perspective corrections. However, for those corrections where you simply didn't point at the horizon and things got a bit tilted, you should be able to correct those with no one but you knowing.

[puttick]

Thanks - I certainly can't see any degradation, even when viewing at "actual pixel" level. Perhaps I am ascribing too much to the importance of the original pixel data - after all it is smudged by the AA filter and then reinterpreted by the camera. One can imagine a 10% (say) change in size of part of the image when correcting perspective (excess perspective can never appear natural if "corrected"), and as you say this is small compared to "upsampling" by 2x to double the linear dimensions of an image.

However it is interesting to try and get deeper into what is going on!

Cheers
Nigel