本人最近在写一套计算机视觉处理软件，用的就是C# + Emgu，因为用到的OpenCV方法就那么几个（大概10多个），为了这些为数不多的方法而带着数MB的Emgu DLL，心里很是不爽，于是乎萌生了将这些方法全部用C# unsafe代码重写的想法，反正OpenCV是开源的，算法可以写成一样的，效率上应该差不到哪去。
下面是我自己写的一个图像类：
///<summary>灰度图像处理类，作者：wmesci</summary>
unsafe class Image :CriticalHandle, IDisposable
{
[DllImport("kernel32.dll")]
static extern IntPtr LocalAlloc(int flags, int size);
[DllImport("kernel32.dll")]
static extern IntPtr LocalFree(IntPtr memBlock);
[DllImport("kernel32.dll", EntryPoint = "RtlMoveMemory")]
static extern unsafe void CopyMemory(byte* dst, byte* src, int count);
[DllImport("kernel32.dll", EntryPoint = "RtlMoveMemory")]
static extern unsafe void CopyMemory(byte* dst, IntPtr src, int count);
[DllImport("kernel32.dll", EntryPoint = "RtlMoveMemory")]
static extern unsafe void CopyMemory(byte* dst, byte[] src, int count);
const byte Max = 255;
const byte Min = 0;
public Image(int width, int height)
: base(IntPtr.Zero)
{
if (width <= 0 || height <= 0)
throw new ArgumentOutOfRangeException();
Width = width;
Height = height;
Length = Width * Height;
base.SetHandle(LocalAlloc(0x40, width * height));
Pointer = (byte*)handle.ToPointer();
}
public Image(int width, int height, byte[] dat)
: this(width, height)
{
if (dat != null)
{
CopyMemory(Pointer, dat, Length);
}
}
public Image(int width, int height, byte* dat)
: this(width, height)
{
CopyMemory(Pointer, dat, Length);
}
public Image(int width, int height, IntPtr dat)
: this(width, height)
{
CopyMemory(Pointer, dat, Length);
}
public readonly int Width;
public readonly int Height;
public readonly int Length;
public readonly byte* Pointer;
public byte this[int x, int y]
{
get
{
return *(Pointer + y * Width + x);
}
set
{
*(Pointer + y * Width + x) = value;
}
}
public Image Clone()
{
return new Image(Width, Height, Pointer);
}
public void Add(Image img)
{
Action<int> act = y =>
{
byte* p1 = Pointer + y * Width, p2 = (byte*)img.Pointer + y * img.Width;
for (int x = 0; x < Width; x++, p1++, p2++)
{
double d = *p1 + *p2;
if (d < 0)
*p1 = 0;
else if (d > 255)
*p1 = 255;
else
*p1 = (byte)d;
}
};
Parallel.For(0, Height, act);
}
public void Sub(Image img)
{
Action<int> act = y =>
{
byte* p1 = Pointer + y * Width, p2 = (byte*)img.Pointer + y * img.Width;
for (int x = 0; x < Width; x++, p1++, p2++)
{
double d = *p1 - *p2;
if (d < 0)
*p1 = 0;
else if (d > 255)
*p1 = 255;
else
*p1 = (byte)d;
}
};
Parallel.For(0, Height, act);
}
public void Mul(Image img, double scale)
{
Action<int> act = y =>
{
byte* p1 = Pointer + y * Width, p2 = (byte*)img.Pointer + y * img.Width;
for (int x = 0; x < Width; x++, p1++, p2++)
{
double d = scale * *p1 * *p2;
if (d < 0)
*p1 = 0;
else if (d > 255)
*p1 = 255;
else
*p1 = (byte)d;
}
};
Parallel.For(0, Height, act);
}
public void Threshold(byte threshold)
{
Action<int> act = y =>
{
byte* p = Pointer + y * Width;
for (int x = 0; x < Width; x++, p++)
{
*p = *p > threshold ? Max : Min;
}
};
Parallel.For(0, Height, act);
}
public void AddWeighted(Image img, double a, double b)
{
Action<int> act = y =>
{
byte* p1 = this.Pointer + y * this.Width, p2 = (byte*)img.Pointer + y * img.Width;
for (int x = 0; x < this.Width; x++, p1++, p2++)
{
double d = a * *p1 + b * *p2;
if (d < 0)
*p1 = 0;
else if (d > 255)
*p1 = 255;
else
*p1 = (byte)d;
}
};
Parallel.For(0, this.Height, act);
}
public static void Smooth(Image src, Image dst, int n)
{
int* tmp = (int*)Marshal.AllocHGlobal(src.Width * src.Height * 4).ToPointer();
Action<int> act = y =>
{
byte* p = src.Pointer + y * src.Width;
int d = 0;
for (int i = -n; i <= n; i++)
{
int xx = GetIndex(i, src.Width);
d += p[xx];
}
tmp[y * src.Width] = d;
};
Parallel.For(0, src.Height, act);
act = y =>
{
int i = y * src.Width;
byte* p = src.Pointer + y * src.Width;
for (int x = 1; x < src.Width; x++)
{
int d = tmp[i];
int x1 = GetIndex(x - n - 1, src.Width);
int x2 = GetIndex(x + n, src.Width);
d += (p[x2] - p[x1]);
tmp[++i] = d;
}
};
Parallel.For(0, src.Height, act);
double f = 1.0 / (2 * n + 1);
f *= f;
act = x =>
{
int d = 0;
byte* p = dst.Pointer + x;
for (int j = -n; j <= n; j++)
{
int yy = GetIndex(j, src.Height);
d += tmp[x + yy * src.Width];
}
*p = (byte)(d * f);
p += src.Width;
for (int y = 1; y < src.Height; y++, p += src.Width)
{
int y1 = GetIndex(y - n - 1, src.Height);
int y2 = GetIndex(y + n, src.Height);
d += (tmp[x + y2 * src.Width] - tmp[x + y1 * src.Width]);
*p = (byte)(d * f);
}
};
Parallel.For(0, src.Width, act);
Marshal.FreeHGlobal(new IntPtr(tmp));
}
private static int GetIndex(int i, int max)
{
if (i < 0) return 0;
if (i >= max) return max - 1;
return i;
}
public override bool IsInvalid
{
get { return handle == IntPtr.Zero; }
}
protected override bool ReleaseHandle()
{
LocalFree(handle);
return true;
}
}
（如有可以优化的地方，烦请指正）
用WPF写了个简单的测试程序，其中运行时间使用Stopwatch计算，取其ElapsedTicks值。
先看下运行环境：
OpenCV使用2.2版本。测试图像大小为600*896，预先进行了灰度化，然后再计算处理时间。
下面直接上结果：
1、Add：
Image：imgt.Add(img)
OpenCV：CvInvoke.cvAdd(img, img, img, IntPtr.Zero)
各执行50次，取平均数：
Image花费时间3246，OpenCV花费时间1514
2、Sub：
Image：imgt.Sub(img)
OpenCV：CvInvoke.cvSub(img, img, img, IntPtr.Zero)
各执行50次，取平均数：
Image花费时间3378，OpenCV花费时间1370
3、Mul：
Image：imgt.Mul(img, 1)
OpenCV：CvInvoke.cvMul(img, img, img, 1)
各执行50次，取平均数：
Image花费时间3817，OpenCV花费时间7480
4、Threshold：
Image：imgt.Threshold(120)
OpenCV：CvInvoke.cvcvThreshold(img, img, 120, 255, Emgu.CV.CvEnum.THRESH.CV_THRESH_BINARY)
各执行50次，取平均数：
Image花费时间1645，OpenCV花费时间1361
5、Smooth：
Image：Image.Smooth(img, dst, 3)
OpenCV：CvInvoke.cvSmooth(img, dst, Emgu.CV.CvEnum.SMOOTH_TYPE.CV_BLUR, 7, 0, 0, 0)
各执行50次，取平均数：
6、AddWeighted：
Image：dst.AddWeighted(img, 0.4, 0.4)
OpenCV：CvInvoke.cvAddWeighted(img, 0.4, img, 0.4, 0, dst)
各执行50次，取平均数：
Image花费时间3952，OpenCV花费时间9845
总结一下：
从上表可以看出，Image类和OpenCV基本上是胜率对半。至于为什么，且听下回分解~~~
更新在另一台电脑上运行的结果：
Image惨败！！
源码及测试代码下载地址：http://download.csdn.net/detail/wmesci/3841089
相关讨论帖：http://topic.csdn.net/u/20111124/23/1F236D07-420E-4E2E-83EE-C9C29E689477.html