Bad image from consecutive histogram equalization AdaptiveTileInterpolation

[leonyu1010]

Hi
I use the follow code to test consecutive histogram Equalizatio + resize. I expect that each cycle is independent and I get near-identical result each time. However, the attached screenshot shows the first cycle is fine, and it degrades in subsequent cycles

Is this a correct way of using the library ? and
Is there a way to "reset" after each cycle so that I can always get the result as good as the first cycle ?

// run the same image and same pipeline multiple times

for (int cycle = 0; cycle < 5; cycle++)
{
    //  pixels  is a SixLabors.ImageSharp.PixelFormats.L16  image   width = 5120, height = 9234
    unsafe
    {
        fixed (void* pixPtr = pixels)
        {
           // the wrappedImage intance is verfied to be correct
            using var wrappedImage = SixLabors.ImageSharp.Image.WrapMemory<SixLabors.ImageSharp.PixelFormats.L16>(pixPtr, image.Width, image.Height);

            var options = new HistogramEqualizationOptions
            {
                Method = HistogramEqualizationMethod.AdaptiveTileInterpolation,
                LuminanceLevels = 4096,
                ClipHistogram = false,
                ClipLimit = 350,
                NumberOfTiles = 8
            };

            // the processing pipeline:  histogram equalization  and reduce image size. 
            var processed = wrappedImage.Clone(ctx =>
            {
                ctx.HistogramEqualization(options);
                ctx.Resize(image.Width / 4, image.Height / 4, KnownResamplers.Bicubic);
            });

           //  save the processed  image to JPEG
            processed.SaveAsJpeg(filename, new JpegEncoder() {Quality = 80, Subsample = JpegSubsample.Ratio444});
        }
    }
}

in the following screenshot,
from left to right: original image, 1st cycle result, 2nd cycle result ... and 5th cycle result

[JimBobSquarePants]

Hi @leonyu1010 where does the pixels buffer come from? Is it immutable?

[antonfirsov]

pixels = load TIFF from disk. so pixels is a buffer of uint16

Could it be that you are loading different layers of the same image between iterations or something like that?
Can you try saving it without processing?

[leonyu1010]

i will try to create a demo repo for this.

meanwhile this screen shows that the wrappedImage before pipeline is correct for each cycle.
left one is the original 12-bit TIFF. the other 5 are wrappedImage.SaveAsJpeg. I thought it makes sense it goes darker under 8 bit. And that is actually the purpose of using histogram equalization here to increase contrast.

[JimBobSquarePants]

@antonfirsov I think the input size is relevant here. I've managed to recreate it using an input png of the same dimensions.

Things I've discovered so far:

• Using a pinned buffer is irrelevant. Any image of the input size seems to be sufficient to trigger the issue.
• The issue is specific to histogram equalization.
• HistogramEqualizationMethod.AdaptiveTileInterpolation is the only one affected.

As a side note. HistogramEqualizationMethod.AdaptiveSlidingWindow is unusably slow for larger images taking over 5 minutes to complete on my SB2.

[leonyu1010]

thanks for the tips. I get very good result using HistogramEqualizationMethod.Global.

[JimBobSquarePants]

PR with fix here #1641