【发布时间】:2016-02-07 18:05:27
【问题描述】:
我正在尝试利用高斯核是可分离的这一事实来实现高性能的高斯模糊,即。 e.您可以将 2D 卷积表示为两个 1D 卷积的组合。
我可以使用以下代码生成两个我认为正确的内核。
/// <summary>
/// Create a 1 dimensional Gaussian kernel using the Gaussian G(x) function
/// </summary>
/// <param name="horizontal">Whether to calculate a horizontal kernel.</param>
/// <returns>The <see cref="T:float[,]"/></returns>
private float[,] CreateGaussianKernel(bool horizontal)
{
int size = this.kernelSize;
float[,] kernel = horizontal ? new float[1, size] : new float[size, 1];
float sum = 0.0f;
float midpoint = (size - 1) / 2f;
for (int i = 0; i < size; i++)
{
float x = i - midpoint;
float gx = this.Gaussian(x);
sum += gx;
if (horizontal)
{
kernel[0, i] = gx;
}
else
{
kernel[i, 0] = gx;
}
}
// Normalise kernel so that the sum of all weights equals 1
if (horizontal)
{
for (int i = 0; i < size; i++)
{
kernel[0, i] = kernel[0, i] / sum;
}
}
else
{
for (int i = 0; i < size; i++)
{
kernel[i, 0] = kernel[i, 0] / sum;
}
}
return kernel;
}
/// <summary>
/// Implementation of 1D Gaussian G(x) function
/// </summary>
/// <param name="x">The x provided to G(x)</param>
/// <returns>The Gaussian G(x)</returns>
private float Gaussian(float x)
{
const float Numerator = 1.0f;
float deviation = this.sigma;
float denominator = (float)(Math.Sqrt(2 * Math.PI) * deviation);
float exponentNumerator = -x * x;
float exponentDenominator = (float)(2 * Math.Pow(deviation, 2));
float left = Numerator / denominator;
float right = (float)Math.Exp(exponentNumerator / exponentDenominator);
return left * right;
}
内核大小由 sigma 计算如下。
this.kernelSize = ((int)Math.Ceiling(sigma) * 2) + 1;
this.sigma = sigma;
给定一个 3 的 sigma,这会在每个方向上产生以下结果。
0.106288522,
0.140321344,
0.165770069,
0.175240144,
0.165770069,
0.140321344,
0.106288522
完美地总结为 1,所以我走在正确的道路上。
事实证明,将内核应用于图像是很困难的,因为出了点问题,虽然我不确定是什么。
我正在使用以下代码来运行 2-pass 算法,该算法在对 Sobel 或 Prewitt 等两个方向上的内核进行卷积以进行边缘检测时完美运行。
protected override void Apply(ImageBase target,
ImageBase source,
Rectangle targetRectangle,
Rectangle sourceRectangle,
int startY,
int endY)
{
float[,] kernelX = this.KernelX;
float[,] kernelY = this.KernelY;
int kernelYHeight = kernelY.GetLength(0);
int kernelYWidth = kernelY.GetLength(1);
int kernelXHeight = kernelX.GetLength(0);
int kernelXWidth = kernelX.GetLength(1);
int radiusY = kernelYHeight >> 1;
int radiusX = kernelXWidth >> 1;
int sourceY = sourceRectangle.Y;
int sourceBottom = sourceRectangle.Bottom;
int startX = sourceRectangle.X;
int endX = sourceRectangle.Right;
int maxY = sourceBottom - 1;
int maxX = endX - 1;
Parallel.For(
startY,
endY,
y =>
{
if (y >= sourceY && y < sourceBottom)
{
for (int x = startX; x < endX; x++)
{
float rX = 0;
float gX = 0;
float bX = 0;
float rY = 0;
float gY = 0;
float bY = 0;
// Apply each matrix multiplier to the
// color components for each pixel.
for (int fy = 0; fy < kernelYHeight; fy++)
{
int fyr = fy - radiusY;
int offsetY = y + fyr;
offsetY = offsetY.Clamp(0, maxY);
for (int fx = 0; fx < kernelXWidth; fx++)
{
int fxr = fx - radiusX;
int offsetX = x + fxr;
offsetX = offsetX.Clamp(0, maxX);
Color currentColor = source[offsetX, offsetY];
float r = currentColor.R;
float g = currentColor.G;
float b = currentColor.B;
if (fy < kernelXHeight)
{
rX += kernelX[fy, fx] * r;
gX += kernelX[fy, fx] * g;
bX += kernelX[fy, fx] * b;
}
if (fx < kernelYWidth)
{
rY += kernelY[fy, fx] * r;
gY += kernelY[fy, fx] * g;
bY += kernelY[fy, fx] * b;
}
}
}
float red = (float)Math.Sqrt((rX * rX) + (rY * rY));
float green = (float)Math.Sqrt((gX * gX) + (gY * gY));
float blue = (float)Math.Sqrt((bX * bX) + (bY * bY));
Color targetColor = target[x, y];
target[x, y] = new Color(red,
green, blue, targetColor.A);
}
}
});
}
这是输入图像:
这是使用 3 sigma 尝试模糊的图像。
如您所见,有些不对劲。这就像我从错误的点或其他地方采样一样。
有什么想法吗?我很欣赏这是一个冗长的问题。
更新
所以根据 Nico Schertler 的建议,我将算法分为以下两遍:
protected override void Apply(
ImageBase target,
ImageBase source,
Rectangle targetRectangle,
Rectangle sourceRectangle,
int startY,
int endY)
{
float[,] kernelX = this.KernelX;
float[,] kernelY = this.KernelY;
int kernelXHeight = kernelX.GetLength(0);
int kernelXWidth = kernelX.GetLength(1);
int kernelYHeight = kernelY.GetLength(0);
int kernelYWidth = kernelY.GetLength(1);
int radiusXy = kernelXHeight >> 1;
int radiusXx = kernelXWidth >> 1;
int radiusYy = kernelYHeight >> 1;
int radiusYx = kernelYWidth >> 1;
int sourceY = sourceRectangle.Y;
int sourceBottom = sourceRectangle.Bottom;
int startX = sourceRectangle.X;
int endX = sourceRectangle.Right;
int maxY = sourceBottom - 1;
int maxX = endX - 1;
// Horizontal blur
Parallel.For(
startY,
endY,
y =>
{
if (y >= sourceY && y < sourceBottom)
{
for (int x = startX; x < endX; x++)
{
float rX = 0;
float gX = 0;
float bX = 0;
// Apply each matrix multiplier to the color
// components for each pixel.
for (int fy = 0; fy < kernelXHeight; fy++)
{
int fyr = fy - radiusXy;
int offsetY = y + fyr;
offsetY = offsetY.Clamp(0, maxY);
for (int fx = 0; fx < kernelXWidth; fx++)
{
int fxr = fx - radiusXx;
int offsetX = x + fxr;
offsetX = offsetX.Clamp(0, maxX);
Color currentColor = source[offsetX, offsetY];
float r = currentColor.R;
float g = currentColor.G;
float b = currentColor.B;
rX += kernelX[fy, fx] * r;
gX += kernelX[fy, fx] * g;
bX += kernelX[fy, fx] * b;
}
}
float red = rX;
float green = gX;
float blue = bX;
Color targetColor = target[x, y];
target[x, y] = new Color(red, green, blue, targetColor.A);
}
}
});
// Vertical blur
Parallel.For(
startY,
endY,
y =>
{
if (y >= sourceY && y < sourceBottom)
{
for (int x = startX; x < endX; x++)
{
float rY = 0;
float gY = 0;
float bY = 0;
// Apply each matrix multiplier to the
// color components for each pixel.
for (int fy = 0; fy < kernelYHeight; fy++)
{
int fyr = fy - radiusYy;
int offsetY = y + fyr;
offsetY = offsetY.Clamp(0, maxY);
for (int fx = 0; fx < kernelYWidth; fx++)
{
int fxr = fx - radiusYx;
int offsetX = x + fxr;
offsetX = offsetX.Clamp(0, maxX);
Color currentColor = source[offsetX, offsetY];
float r = currentColor.R;
float g = currentColor.G;
float b = currentColor.B;
rY += kernelY[fy, fx] * r;
gY += kernelY[fy, fx] * g;
bY += kernelY[fy, fx] * b;
}
}
float red = rY;
float green = gY;
float blue = bY;
Color targetColor = target[x, y];
target[x, y] = new Color(red, green, blue, targetColor.A);
}
}
});
}
我越来越接近我的目标,因为现在出现了模糊效果。不幸的是,它不正确。
如果您仔细观察,您会发现垂直方向有双条带,而水平方向没有足够的模糊。当我将 sigma 提高到 10 时,下图清楚地表明了这一点。
任何助手都会很棒。
【问题讨论】:
-
你不能像这样分离卷积。您必须运行两个单独的通道。第二个卷积应该使用第一个卷积的结果。
-
你能用代码示例扩展一下吗?
-
基本上,调用你的方法两次。在第一次调用中,省略垂直卷积。在第二次调用中,省略水平卷积。
-
你会想要保留这个艺术性的交叉影线过滤器;-)
-
@TaW 一旦我的框架运行起来,我肯定会玩这种东西。
标签: c# image-processing gaussian convolution