Unexpected artefacts when quantizing images #2469

Activity

1. BrentCoxEscentual

   added

   needs triage

   on May 31, 2023
2. JimBobSquarePants commented on May 31, 2023

   

   JimBobSquarePants

   on May 31, 2023

   Member

   More actions

   What happens when you quantize without dithering?

   Why are you quantizing then resaving at 32bit?
3. BrentCoxEscentual commented on May 31, 2023

   

   BrentCoxEscentual

   on May 31, 2023

   Author

   More actions

   Hi there,

   I created the the demo program with the absolute minimum code to illustrate the problem that I came across while working with the library and thought it best to raise it.

   As to saving as 32 bit, it was not a conscious decision.

   I am happy to try with out dithering with you could point me in the right direction how I can do that.

   Thanks for your time with this
4. JimBobSquarePants commented on May 31, 2023

   

   JimBobSquarePants

   on May 31, 2023

   Member

   More actions

   I am happy to try with out dithering with you could point me in the right direction how I can do that.

   Instead of using KnownQuantizers.Wu new up an instance of the WuQuantizer and passing QuantizerOptions.Dither as null.

   It's likely the dither as the pattern stems down and left from the widening as it does which is the direction error diffusion goes.

   You can set the strength of the dither (0-1) in those same options.
5. BrentCoxEscentual commented on May 31, 2023

   

   BrentCoxEscentual

   on May 31, 2023

   Last edited by BrentCoxEscentual

   Author

   More actions

   I am happy to try with out dithering with you could point me in the right direction how I can do that.

   Instead of using KnownQuantizers.Wu new up an instance of the WuQuantizer and passing QuantizerOptions.Dither as null.

   It's likely the dither as the pattern stems down and left from the widening as it does which is the direction error diffusion goes.

   You can set the strength of the dither (0-1) in those same options.

   Hi there

   I have made the change as suggested.

   using (var i = Image.Load("Orig.png"))
   {
       i.Mutate(x => x.Quantize(new WuQuantizer(new QuantizerOptions()
       {
           Dither = null
       })));
   
       await i.SaveAsPngAsync("test.png", new PngEncoder()
       {
           CompressionLevel = PngCompressionLevel.BestCompression
       });
   }
   

   The artifact has now disappeared and the image is as expect.

   I have tested all the Dithering types and the following types create the artifacts.

   Burks
   FloydSteinberg
   Sierra2
   Sierra3
   SierraLite
   Stucki

   The rest are fine and don't product any issues that I can see.

   Purely for my education and better understanding, why do you think the dithering would cause this issue?

   Thanks again for you time and help.

   Regards
6. JimBobSquarePants commented on Jun 12, 2023

   

   JimBobSquarePants

   on Jun 12, 2023

   Member

   More actions

   The dithering algorithms work by pushing error (this difference between a quantized pixel and the original) to its surrounding pixels.

   For example, Floyd Steinburg propagates the error to the neighboring pixels as follows:

   • Distribute 7/16 of the error to the pixel on the right.
   • Distribute 3/16 of the error to the pixel on the bottom-left.
   • Distribute 5/16 of the error to the pixel directly below.
   • Distribute 1/16 of the error to the pixel on the bottom-right.

   We use a lookup table that stores up to 5 bits of pixel data per component to match the result of a quantized pixel against it's undithered equivalent. We use 5 bits as that keeps memory low (6 bits would be preferred but that uses over 30 MB compared to 2 1/4 MB). Sometimes we end up with a collision which leads to the artifacts you are seeing.

   This, in the case of your image pushes the artifact caused by that shadow down and to the right. This is why I suggested adjusting the strength of the dithering. Using .5F as a parameter for example would reduce the error by half.

   Here's the output from doing that.

   
7. JimBobSquarePants

   added a commit that references this issue on Jun 12, 2023

   Use more accurate bit for alpha lookup. Fix #2469

   This issue will close once commit ae3900d is merged into the 'main' branch.

   ae3900d
8. JimBobSquarePants commented on Jun 12, 2023

   

   JimBobSquarePants

   on Jun 12, 2023

   Member

   More actions

   I've actually opened up #2473 which allocates more memory towards alpha component comparisons in our lookup table which fixes the issue using the default encoder settings.
9. JimBobSquarePants

   closed this as completedin #2473on Jun 13, 2023