affine should have a background parameter, and not have jaggy edges

[chregu]

I try to rotate an image by arbitrary angle. Basically works fine with just similarity, but the borders look pretty jagged, compared to what for example imagick produces.

Any hints to make them smoother?
With VIPS:

With IMagick:

I'd also like to have a user-definable "background" layer instead of just an fully transparent alpha channel. For now I do this with:

                        $aa = $this->vips->extract_band(3);
                        $this->vips = $aa->ifthenelse($this->vips, $background);

But I assume this doesn't really work nicely, in case I happen to have smooth borders with hopefully also smoothness on the alpha channel for the rotated image. How would I put the rotated image on top of the background image (which can have an alpha channel as well)? Basically a compositeImage operation in IMagick, I assume.

[jcupitt]

I agree, the jagged edges are ugly, they should be fixed.

In the meantime, you can add a one-pixel border around the input image, the same colour as the background for similarity (ie. always black). You're right, a parameter to set the background for affine and similarity would be useful.

You can use flatten to composite an image with transparency against a background colour, would that help? It's the thing jpegsave uses to implement the background parameter there.

8.6 has a new composite operator that can merge a stack of transparent images using a set of PDF blend modes (so over, multiply, saturate etc.), which I think would be useful for you. 8.6 is sadly still not out, perhaps another week or two.

[jcupitt]

I've made this into an enhancement issue for affine, I hope that's OK.

[chregu]

Thanks for the answer, looking forward to the enhancements.

Was also thinking about adding a border before rotating. What's the best way to do this? (I hope my questions aren't too basic, but I couldn't figure that one out)

I assume flatten doesn't help in this case, since the background can have transparency as well, which I'd like to keep. But hopefully composite will solve my problems.

[jcupitt]

Use embed to add borders:

https://jcupitt.github.io/php-vips/docs/classes/Jcupitt.Vips.ImageAutodoc.html#method_embed

The C docs have more detail:

https://jcupitt.github.io/libvips/API/current/libvips-conversion.html#vips-embed

However, I think the right place to fix this would be in libvips. Let me see if I can address this in 8.6, and then you can avoid adding ugly workarounds to your code.

[jcupitt]

There's an issue on the main libvips repo to track this improvement.

[jcupitt]

It seems to be fixed in this branch:

https://github.com/jcupitt/libvips/tree/fix-affine-jaggies

With this command:

$ vips similarity grid.png x.png --angle 30 --background 128

I see:

Would you be able to test this?

[chregu]

Thanks for the quick fix. Looks good in general, only when I use eg. black color with full transparency ([0,0,0,255] in PHP), it looks kinda wrong

But way better than before for other cases (the above one may be a strange case anyway)

[chregu]

btw, can confirm that the composite works, eg:

        $image = $image->composite([$image, $insertImage], [2]);

works nicely, also with alpha channels. Looking forward to see #55 implemented to have consts for those ints.

[jcupitt]

It wasn't premultiplying alpha, which can add edges on resampling. Could you test again?

Funny how fiddly it can be to get all the edge cases out.

[jcupitt]

Oh, test image:

Command:

$ vips similarity grid-rgba.png x.png --angle 30

Result:

[chregu]

The borders are fine now, but the background (apart from the direct border pixels) is now always totally transparent, even with for example ['background' => [0,0,0,255]]. On purpose, so I should composite a background in?

[jcupitt]

Groan, what an idiot, I forgot to convert the background to the pixel format of the premultiplied image. Try again, I now see:

$ vips similarity grid-rgba.png x.png --angle 30 --background "128 0 0 255"

[chregu]

perfect. thanks a lot, looking forward to 8.6

[jcupitt]

Thank you very much for testing this. I'll merge to master and add stuff for this to the test suite.