///<summary>Реализует слой WCT.</summary>
///<param name="Content">Карты признаков контентного изображения.</param>
///<param name="Style">Карты признаков стилевого изображения.</param>
///<param name="alpha">Глубина стилизации. [0..1], по умолчанию = 0.6.</param>
public static Tensor WCT(Tensor Content, Tensor Style, float alpha = 0.6f)
{
var content_flat = Content.Flat();
var style_flat = Style.Flat();
var Content_mean = content_flat.HeightMean();
var cF = content_flat.SubHeight(Content_mean);
var contentCov = Tensor.MatMul(cF, cF.Transpose()).Div(Content.Height * Content.Width - 1f);
var Style_mean = style_flat.HeightMean();
var sF = style_flat.SubHeight(Style_mean);
var styleCov = Tensor.MatMul(sF, sF.Transpose()).Div(Style.Height * Style.Width - 1f);
var c_usv = LinAlg.SVD(contentCov);
var c_u = c_usv.Item1;
var c_e = c_usv.Item2;
var c_v = c_usv.Item3;
var s_usv = LinAlg.SVD(styleCov);
var s_u = s_usv.Item1;
var s_e = s_usv.Item2;
var s_v = s_usv.Item3;
var k_c = c_e.W.Length;
for (int i = 0; i < c_e.W.Length; i++)
{
if (c_e.W[i] < 0.00001f)
{
k_c = i - 1;
break;
}
}
var k_s = s_e.W.Length;
for (int i = 0; i < s_e.W.Length; i++)
{
if (s_e.W[i] < 0.00001f)
{
k_s = i - 1;
break;
}
}
var c_d = c_e.Slice1D(0, k_c).Pow(-0.5f);
var step1 = Tensor.MatMul(Slice2dHeight(c_v, 0, k_c), c_d.Diag());
var step2 = Tensor.MatMul(step1, Slice2dHeight(c_v, 0, k_c).Transpose());
var whiten_cF = Tensor.MatMul(step2, cF);
var s_d = s_e.Slice1D(0, k_s).Pow(0.5f);
var targetFeature = Tensor.MatMul(Tensor.MatMul(Tensor.MatMul(Slice2dHeight(s_v, 0, k_s), s_d.Diag()), Slice2dHeight(s_v, 0, k_s).Transpose()), whiten_cF);
targetFeature = targetFeature.AddHeight(Style_mean);
var Out = targetFeature.Unflat(Content.Width, Content.Height);
var OneMinusAlpha = 1f - alpha;
Parallel.For(0, Out.Depth, (int d) =>
{
for (int y = 0; y < Out.Height; y++)
{
for (int x = 0; x < Out.Width; x++)
{
Out.Set(x, y, d, Out.Get(x, y, d) * alpha + Content.Get(x, y, d) * OneMinusAlpha);
}
}
});
return(Out);
}
