- replace hard binary thresholding with smooth linear alpha mapping to reduce halos around foreground edges especially in low resolution images - normalize now uses a quadratic curve instead of hard binarization so that gradient info from the model output is preserved - removebackgroundasync applies a configurable linear interpolation between the set min and max feathering thresholds so you have partial transparency at the edges
133 lines
3.9 KiB
C#
133 lines
3.9 KiB
C#
namespace BGR.Console.Removal.ImageSharp;
|
|
|
|
internal class ImageSharpProcessor : ImageProcessor
|
|
{
|
|
public override async Task<IImage> LoadImageAsync(string path)
|
|
{
|
|
var image = await Image.LoadAsync<Rgba32>(path);
|
|
|
|
if (image.Metadata.DecodedImageFormat is null)
|
|
{
|
|
throw new InvalidOperationException("Image format is not supported.");
|
|
}
|
|
|
|
var stream = new MemoryStream();
|
|
await image.SaveAsync(stream, image.Metadata.DecodedImageFormat);
|
|
stream.Position = 0;
|
|
|
|
return new SharpImage(image.Width, image.Height, stream);
|
|
}
|
|
|
|
public override async Task<ITensor<float>> CreateTensorInputAsync(Stream image, Model model)
|
|
{
|
|
using var resized = await Image.LoadAsync<Rgba32>(image);
|
|
|
|
resized.Mutate(x => x.Resize(model.InputWidth, model.InputHeight));
|
|
|
|
const int batchSize = 1;
|
|
const int channels = 3;
|
|
var tensor = new OnnxTensor(batchSize, channels, model.InputHeight, model.InputWidth);
|
|
|
|
WalkImage(resized.Height, resized.Width, (x, y) =>
|
|
{
|
|
var pixel = resized[x, y];
|
|
tensor.SetValue(0, 0, y, x, model.NormalizeRed(pixel.R));
|
|
tensor.SetValue(0, 1, y, x, model.NormalizeGreen(pixel.G));
|
|
tensor.SetValue(0, 2, y, x, model.NormalizeBlue(pixel.B));
|
|
});
|
|
|
|
return tensor;
|
|
}
|
|
|
|
public override async Task<Stream> GenerateMaskAsync(ITensor<float> maskTensor, int width, int height)
|
|
{
|
|
using var mask = new Image<Rgba32>(width, height);
|
|
|
|
using Image<Rgba32> tempMask = new(maskTensor.Width, maskTensor.Height);
|
|
|
|
const byte opaqueAlpha = 255;
|
|
|
|
WalkImage(maskTensor.Height, maskTensor.Width, (x, y) =>
|
|
{
|
|
var sigmoidValue = CalculateSigmoid(maskTensor.GetValue(0, 0, y, x));
|
|
var normalizedValue = Normalize(sigmoidValue);
|
|
var intensity = ConvertToGreyscale(normalizedValue);
|
|
|
|
tempMask[x, y] = new Rgba32(intensity, intensity, intensity, opaqueAlpha);
|
|
});
|
|
|
|
tempMask.Mutate(x => x.Resize(width, height));
|
|
|
|
WalkImage(height, width, (x, y) => mask[x, y] = tempMask[x, y]);
|
|
|
|
var stream = new MemoryStream();
|
|
await mask.SaveAsync(stream, new PngEncoder());
|
|
stream.Position = 0;
|
|
|
|
return stream;
|
|
}
|
|
|
|
public override async Task<Stream> RemoveBackgroundAsync(Stream image, Stream mask, byte featherMin, byte featherMax)
|
|
{
|
|
image.Position = 0;
|
|
mask.Position = 0;
|
|
|
|
var imageWithBg = await Image.LoadAsync<Rgba32>(image);
|
|
var maskImage = await Image.LoadAsync<Rgba32>(mask);
|
|
using var imageWithBgRemoved = new Image<Rgba32>(imageWithBg.Width, imageWithBg.Height);
|
|
|
|
WalkImage(imageWithBg.Height, imageWithBg.Width, (x, y) =>
|
|
{
|
|
var sourcePixel = imageWithBg[x, y];
|
|
var maskPixel = maskImage[x, y];
|
|
|
|
var alpha = AdjustAlpha(maskPixel.R, featherMin, featherMax);
|
|
|
|
imageWithBgRemoved[x, y] = new Rgba32(sourcePixel.R, sourcePixel.G, sourcePixel.B, alpha);
|
|
});
|
|
|
|
var result = new MemoryStream();
|
|
await imageWithBgRemoved.SaveAsync(result, new PngEncoder());
|
|
result.Position = 0;
|
|
|
|
return result;
|
|
}
|
|
|
|
public override async Task SaveImageAsync(Stream image, string path)
|
|
{
|
|
image.Position = 0;
|
|
var img = await Image.LoadAsync<Rgba32>(image);
|
|
await img.SaveAsync(path, new PngEncoder());
|
|
}
|
|
|
|
private static float Normalize(float value)
|
|
{
|
|
return value * value;
|
|
}
|
|
|
|
private static byte ConvertToGreyscale(float value)
|
|
{
|
|
const float maxIntensity = 255f;
|
|
return (byte)(value * maxIntensity);
|
|
}
|
|
|
|
private static float CalculateSigmoid(float x)
|
|
{
|
|
const float sigmoidScale = 1f;
|
|
const float sigmoidShift = 1f;
|
|
const float sigmoidDivisor = -1f;
|
|
|
|
return sigmoidScale / (sigmoidShift + MathF.Exp(sigmoidDivisor * x));
|
|
}
|
|
|
|
private static byte AdjustAlpha(byte maskValue, byte minVal, byte maxVal)
|
|
{
|
|
if (maskValue <= minVal)
|
|
return 0;
|
|
if (maskValue >= maxVal)
|
|
return 255;
|
|
|
|
float proportion = (maskValue - minVal) / (float)(maxVal - minVal);
|
|
return (byte)(proportion * 255f);
|
|
}
|
|
} |