Imagesharp: Default PNG quantizer introduces jaggedness

[loraderon]

Prerequisites

• [x] I have written a descriptive issue title
• [x] I have verified that I am running the latest version of ImageSharp
• [x] I have verified if the problem exist in both DEBUG and RELEASE mode
• [x] I have searched open and closed issues to ensure it has not already been reported

Description

When saving png using default settings visible jaggedness is introduced.

Steps to Reproduce

void Main()
{
    using (var image = Image.Load(File.ReadAllBytes("sample.png")))
    using(var stream = File.Open("output.png", FileMode.Create))
        image.SaveAsPng(stream);
}

System Configuration

Windows 10 and in Azure

• ImageSharp version: 1.0.0-dev002459
• Other ImageSharp packages and versions:
• Environment (Operating system, version and so on):
• .NET Framework version: .NET Core 2.2 & .NET Core 3.0 (preview)
• Additional information:

[dlemstra]

This appears to be happening due to the fact that the number of colors in the pallette are being reduced from 83 to 2:

I:\issues\imagesharp\866>identify *.png
55246806-4cbaa700-5246-11e9-9f0f-4e21acab9a97.png PNG 1003x376 1003x376+0+0 8-bit sRGB 83c 7844B 0.000u 0:00.001
55246926-9e633180-5246-11e9-9e53-56faea37decb.png PNG 1003x376 1003x376+0+0 8-bit sRGB 2c 5047B 0.000u 0:00.000

[JimBobSquarePants]

@dlemstra Had a very quick look at the quantizer. Algorithmically it seems ok just a shade aggressive. It would be really useful if someone could compare the color count to the implementation we based ours on here to ensure we haven't made any mistakes.

https://github.com/JeremyAnsel/JeremyAnsel.ColorQuant/blob/master/JeremyAnsel.ColorQuant/JeremyAnsel.ColorQuant/WuAlphaColorQuantizer.cs

[JimBobSquarePants]

Figured it out. We're not using enough accuracy when calculating variance. Will have 69 colors once I submit a fix.

jagged.png PNG 1003x376 1003x376+0+0 8-bit sRGB 69c 9487B 0.000u 0:00.000

[JimBobSquarePants]

We can actually get by with slightly less accuracy (by using far less memory) and produce a good image at 48 colors.