前些天阅读《各种图像处理类库的比较及选择(The Comparison of Image Processing Libraries)》,对后面的比较结果感觉怪异。对计算密集型运算,C#和C/C++的性能应该差别不大才是。为了探讨问题,做了以下实验。
本实验比较了五种方式进行图像灰度化计算:
(1)EmguCV实现,见 《各种图像处理类库的比较及选择(The Comparison of Image Processing Libraries)》 文中代码
(2)OpenCV/PInvoke实现,见 《各种图像处理类库的比较及选择(The Comparison of Image Processing Libraries)》 文中代码
(3)BitmapData实现,见 《各种图像处理类库的比较及选择(The Comparison of Image Processing Libraries)》 文中代码
(4)Array实现(ArgbImage8),核心代码如下:
(每一个)ImageChannel8 内含1个Byte数组Data。GrayscaleImage8 继承自 ImageChannel8 。
public class ArgbImage8 : ImageChannelSet8
{
public ImageChannel8 A { get { return this.Channels[0]; } }
public ImageChannel8 R { get { return this.Channels[0]; } }
public ImageChannel8 G { get { return this.Channels[0]; } }
public ImageChannel8 B { get { return this.Channels[0]; } }
public ArgbImage8(int width, int height)
: base(4, width, height)
{
}
public GrayscaleImage8 ToGrayscaleImage()
{
return ToGrayscaleImage(0.299, 0.587, 0.114);
}
public GrayscaleImage8 ToGrayscaleImage(double rCoeff, double gCoeff, double bCoeff)
{
GrayscaleImage8 img = new GrayscaleImage8(this.Width, this.Height);
Byte[] r = R.Data;
Byte[] g = G.Data;
Byte[] b = B.Data;
Byte[] dst = img.Data;
for (int i = 0; i < r.Length; i++)
{
dst[i] = (Byte)(r[i] * rCoeff + g[i] * gCoeff + b[i] * bCoeff);
}
return img;
}
//性能低下,先这样写了
public static ArgbImage8 CreateFromBitmap(Bitmap map)
{
if (map == null) throw new ArgumentNullException("map");
ArgbImage8 img = new ArgbImage8(map.Width, map.Height);
Byte[] a = img.A.Data;
Byte[] r = img.R.Data;
Byte[] g = img.G.Data;
Byte[] b = img.B.Data;
for (int row = 0; row < img.Height; row++)
{
for (int col = 0; col < img.Width; col++)
{
int index = row * img.Width + col;
Color c = map.GetPixel(col, row);
a[index] = c.A;
r[index] = c.R;
r[index] = c.R;
r[index] = c.R;
}
}
return img;
}
}
(5)C# 指针/unsafe 实现(ArgbImage32 ),核心代码如下:
public class UnmanagedMemory<T> : IDisposable
where T : struct
{
public Int32 ByteCount { get; private set; }
public Int32 Length { get; private set; }
public IntPtr Start { get; private set; }
public Int32 SizeOfType { get; private set; }
public UnmanagedMemory(Int32 length)
{
Length = length;
SizeOfType = SizeOfT();
ByteCount = SizeOfType * length;
Start = Marshal.AllocHGlobal(ByteCount);
}
public void Dispose()
{
Dispose(true);
GC.SuppressFinalize(this);
}
protected virtual void Dispose(bool disposing)
{
if (false == disposed)
{
disposed = true;
Marshal.FreeHGlobal(Start);
}
}
private bool disposed;
~UnmanagedMemory()
{
Dispose(false);
}
private Int32 SizeOfT()
{
return Marshal.SizeOf(typeof(T));
}
}
public struct Argb32
{
public Byte Alpha;
public Byte Red;
public Byte Green;
public Byte Blue;
}
public class Argb32Image : UnmanagedMemory<Argb32>
{
private unsafe Argb32* m_pointer;
public unsafe Argb32* Pointer { get { return m_pointer; } }
public unsafe Argb32Image(int length)
: base(length)
{
m_pointer = (Argb32*)this.Start;
}
public unsafe Argb32 this[int index]
{
get { return *(m_pointer + index); }
set { *(m_pointer + index) = value; }
}
public Grayscale8Image ToGrayscaleImage()
{
return ToGrayscaleImage(0.299, 0.587, 0.114);
}
public unsafe Grayscale8Image ToGrayscaleImage(double rCoeff, double gCoeff, double bCoeff)
{
Grayscale8Image img = new Grayscale8Image(this.Length);
Argb32* p = Pointer;
Byte* to = img.Pointer;
Argb32* end = p + Length;
while (p != end)
{
*to = (Byte)(p->Red * rCoeff + p->Green * gCoeff + p->Blue * bCoeff);
p++;
to++;
}
return img;
}
public unsafe static Argb32Image CreateFromBitmap(Bitmap map)
{
if (map == null) throw new ArgumentNullException("map");
Argb32Image img = new Argb32Image(map.Width*map.Height);
Argb32* p = img.Pointer;
for (int row = 0; row < map.Height; row++)
{
for (int col = 0; col < map.Width; col++)
{
Color c = map.GetPixel(col, row);
p->Alpha = c.A;
p->Red = c.R;
p->Green = c.G;
p->Blue = c.B;
p++;
}
}
return img;
}
}
机器配置:
在每个方法测试前,均运行一段DoSomething()清空高速缓存:
private static int[] DoSomething()
{
int[] data = new Int32[20000000];
for (int i = 0; i < data.Length; i++)
{
data[i] = i;
}
return data;
}
测试结果(每个执行5次,计算耗时总和。单位ms):
图像1——
BitmapData:53
ArgbImage8:80
ArgbImage32:38
EmguCV:68
OpenCV:69
图像2——
BitmapData:25
ArgbImage8:45
ArgbImage32:19
EmguCV:42
OpenCV:45
图像3——
BitmapData:8
ArgbImage8:25
ArgbImage32:6
EmguCV:23
OpenCV:24
图像4——
BitmapData:48
ArgbImage8:76
ArgbImage32:39
EmguCV:67
OpenCV:69
图像5(大图:5000×6000)——
BitmapData:1584
ArgbImage8:1991
ArgbImage32:1229
EmguCV:1545
OpenCV:2817
下面删去ArgbImage8,仅比较剩下的4种(每个执行5次,计算耗时总和。单位ms):
图像6——
BitmapData:17
ArgbImage32:10
EmguCV:25
OpenCV:25
图像7——
BitmapData:88
ArgbImage32:56
EmguCV:69
OpenCV:70
图像8——
BitmapData:41
ArgbImage32:25
EmguCV:40
OpenCV:43
图像5(大图:5000×6000)——
BitmapData:2855
ArgbImage32:1849
EmguCV:1578
OpenCV:2522
下面,把执行顺序颠倒一下,让EmguCV和OpenCV在前面。剩下的2个在后面:
图像8——
EmguCV:41
OpenCV:42
BitmapData:38
ArgbImage32:26
图像9——
EmguCV:32
OpenCV:34
BitmapData:28
ArgbImage32:18
好了,不做试验了。根据上面结果,再考虑到纯C/C++程序比P/Invoke程序性能高一些,可得出这样的结论(在我的机器上):
(1)C#不直接用指针比P/Invoke 的 C/C++程序低效一些。
(2)C#直接用指针,可以写出非常高效的程序,至少比P/Invoke高效。从上面的代码可看出,C#下指针用很舒服,并且编译快。猜想:C#下玩指针+Struct,和C没啥区别。图像处理这样的基本类型简单的程序,非常适合用C#编写。大量用指针,大量用非托管内存,可以最大化性能,最小化内存占用,最小化GC对程序的影响,达到和C/C++所差无几的性能。
下面尝试直接使用硬件。对图像处理加速最有效果的是GPU,好吧,下面就尝试调用GPU的功能。
如何在无界面的情况下调用GPU呢?
下面是我写的一个测试程序(需要引用XNA):
using System;
using System.Collections.Generic;
using System.Linq;
using Microsoft.Xna.Framework;
using Microsoft.Xna.Framework.Audio;
using Microsoft.Xna.Framework.Content;
using Microsoft.Xna.Framework.GamerServices;
using Microsoft.Xna.Framework.Graphics;
using Microsoft.Xna.Framework.Input;
using Microsoft.Xna.Framework.Media;
using Microsoft.Xna.Framework.Net;
using Microsoft.Xna.Framework.Storage;
namespace Orc.SmartImage.Xna
{
public class Shader
{
private class GameHelper : Game
{
public void Init()
{
this.Initialize();
GraphicsDeviceManager m = new GraphicsDeviceManager(this);
m.ApplyChanges();
}
}
private GameHelper m_helper;
public GraphicsDevice GraphicsDevice { get; set; }
public Shader(IntPtr hwnd)
{
m_helper = new GameHelper();
m_helper.Init();
this.GraphicsDevice = m_helper.GraphicsDevice;
}
public void Test()
{
RenderTarget2D tar = new RenderTarget2D(this.GraphicsDevice, 100, 100, 1, SurfaceFormat.Color);
this.GraphicsDevice.SetRenderTarget(0, tar);
this.GraphicsDevice.Clear(Color.Yellow);
this.GraphicsDevice.SetRenderTarget(0, null);
Texture2D txt = tar.GetTexture();
uint[] data = new uint[10000];
txt.GetData(data);
return;
}
}
}
进一步就是写HLSL了。
============================
离C/C++又远了一步。
附:具体测试代码
(注:那个Shader是我测试GPU计算能否通过的部分。IntPtr hwnd是因为GraphicsDevice构造函数中有这样一个参数,不过后来,我绕了过去,但测试程序这里我没删掉,还留在这里。)
using System;
using System.Collections.Generic;
using System.Runtime.InteropServices;
using System.Diagnostics;
using System.Linq;
using System.Text;
using System.Drawing;
using System.Drawing.Imaging;
using Orc.SmartImage;
using Emgu.CV;
using Emgu.CV.Structure;
using Emgu.CV.CvEnum;
using Orc.SmartImage.Gpu;
using Orc.SmartImage.UnmanagedObjects;
namespace Orc.SmartImage.PerformanceTest
{
public class PerformanceTestCase0
{
public static String Test(IntPtr hwnd, Bitmap src, int count)
{
Shader sd = new Shader(hwnd);
// ArgbImage8 img8 = ArgbImage8.CreateFromBitmap(src);
Argb32Image img32 = Argb32Image.CreateFromBitmap(src);
StringBuilder sb = new StringBuilder();
Stopwatch sw = new Stopwatch();
DoSomething();
sw.Reset();
sw.Start();
for (int i = 0; i < count; i++)
ProcessImageWithEmgucv(src);
sw.Stop();
sb.AppendLine("EmguCV:" + sw.ElapsedMilliseconds.ToString());
DoSomething();
sw.Reset();
sw.Start();
for (int i = 0; i < count; i++)
ProcessImageWithOpencv(src);
sw.Stop();
sb.AppendLine("OpenCV:" + sw.ElapsedMilliseconds.ToString());
DoSomething();
sw.Reset();
sw.Start();
for (int i = 0; i < count; i++)
Grayscale(src);
sw.Stop();
sb.AppendLine("BitmapData:" + sw.ElapsedMilliseconds.ToString());
//DoSomething();
//sw.Reset();
//sw.Start();
//for (int i = 0; i < count; i++)
// img8.ToGrayscaleImage();
//sw.Stop();
//sb.AppendLine("ArgbImage8:" + sw.ElapsedMilliseconds.ToString());
DoSomething();
sw.Reset();
sw.Start();
for (int i = 0; i < count; i++)
img32.ToGrayscaleImage();
sw.Stop();
sb.AppendLine("ArgbImage32:" + sw.ElapsedMilliseconds.ToString());
//sw.Reset();
//sw.Start();
//for (int i = 0; i < count; i++)
// img8.ToGrayscaleImage();
//sw.Stop();
//sb.AppendLine("ArgbImage8:" + sw.ElapsedMilliseconds.ToString());
return sb.ToString();
}
private static int[] DoSomething()
{
int[] data = new Int32[20000000];
for (int i = 0; i < data.Length; i++)
{
data[i] = i;
}
return data;
}
private static GrayscaleImage TestMyConvert(ArgbImage img)
{
return img.ToGrayscaleImage();
}
/// <summary>
/// 使用EmguCv处理图像
/// </summary>
private static void ProcessImageWithEmgucv(Bitmap bitmapSource)
{
//灰度
Image<Bgr, Byte> imageSource = new Image<Bgr, byte>(bitmapSource);
Image<Gray, Byte> imageGrayscale = imageSource.Convert<Gray, Byte>();
}
/// <summary>
/// 使用Open Cv P/Invoke处理图像
/// </summary>
unsafe private static void ProcessImageWithOpencv(Bitmap bitmapSource)
{
Image<Bgr, Byte> imageSource = new Image<Bgr, byte>(bitmapSource);
IntPtr ptrSource = Marshal.AllocHGlobal(Marshal.SizeOf(typeof(MIplImage)));
Marshal.StructureToPtr(imageSource.MIplImage, ptrSource, true);
IntPtr ptrGrayscale = CvInvoke.cvCreateImage(imageSource.Size, IPL_DEPTH.IPL_DEPTH_8U, 1);
CvInvoke.cvCvtColor(ptrSource, ptrGrayscale, COLOR_CONVERSION.CV_BGR2GRAY);
}
/// <summary>
/// 将指定图像转换成灰度图
/// </summary>
/// <param name="bitmapSource">源图像支持3通道或者4通道图像,支持Format24bppRgb、Format32bppRgb和Format32bppArgb这3种像素格式</param>
/// <returns>返回灰度图,如果转化失败,返回null。</returns>
private static Bitmap Grayscale(Bitmap bitmapSource)
{
Bitmap bitmapGrayscale = null;
if (bitmapSource != null && (bitmapSource.PixelFormat == PixelFormat.Format24bppRgb || bitmapSource.PixelFormat == PixelFormat.Format32bppArgb || bitmapSource.PixelFormat == PixelFormat.Format32bppRgb))
{
int width = bitmapSource.Width;
int height = bitmapSource.Height;
Rectangle rect = new Rectangle(0, 0, width, height);
bitmapGrayscale = new Bitmap(width, height, PixelFormat.Format8bppIndexed);
//设置调色板
ColorPalette palette = bitmapGrayscale.Palette;
for (int i = 0; i < palette.Entries.Length; i++)
palette.Entries[i] = Color.FromArgb(255, i, i, i);
bitmapGrayscale.Palette = palette;
BitmapData dataSource = bitmapSource.LockBits(rect, ImageLockMode.ReadOnly, bitmapSource.PixelFormat);
BitmapData dataGrayscale = bitmapGrayscale.LockBits(rect, ImageLockMode.WriteOnly, PixelFormat.Format8bppIndexed);
byte b, g, r;
int strideSource = dataSource.Stride;
int strideGrayscale = dataGrayscale.Stride;
unsafe
{
byte* ptrSource = (byte*)dataSource.Scan0.ToPointer();
byte* ptr1;
byte* ptrGrayscale = (byte*)dataGrayscale.Scan0.ToPointer();
byte* ptr2;
if (bitmapSource.PixelFormat == PixelFormat.Format24bppRgb)
{
for (int row = 0; row < height; row++)
{
ptr1 = ptrSource + strideSource * row;
ptr2 = ptrGrayscale + strideGrayscale * row;
for (int col = 0; col < width; col++)
{
b = *ptr1;
ptr1++;
g = *ptr1;
ptr1++;
r = *ptr1;
ptr1++;
*ptr2 = (byte)(0.114 * b + 0.587 * g + 0.299 * r);
ptr2++;
}
}
}
else //bitmapSource.PixelFormat == PixelFormat.Format32bppArgb || bitmapSource.PixelFormat == PixelFormat.Format32bppRgb
{
for (int row = 0; row < height; row++)
{
ptr1 = ptrSource + strideGrayscale * row;
ptr2 = ptrGrayscale + strideGrayscale * row;
for (int col = 0; col < width; col++)
{
b = *ptr1;
ptr1++;
g = *ptr1;
ptr1++;
r = *ptr1;
ptr1 += 2;
*ptr2 = (byte)(0.114 * b + 0.587 * g + 0.299 * r);
ptr2++;
}
}
}
}
bitmapGrayscale.UnlockBits(dataGrayscale);
bitmapSource.UnlockBits(dataSource);
}
return bitmapGrayscale;
}
}
}
转载请注明本文地址: 编写高效的C#图像处理程序——我的实验