编写高效的C#图像处理程序(3) Rgb=>Lab，图像缺陷检测的例子
时间：2011-01-13 来源：xiaotie
1. Rgb24和Lab24
Rgb是电脑上使用较多的彩色空间，Lab是针对人的感知设计的彩色空间，很多情况下进行彩色图像分析，需要在Rgb彩色空间和Lab彩色空间之间进行转化。关于Lab彩色空间的详细介绍和Rgb空间与Lab空间的转换公式见维基百科的对应词条 Lab色彩空间，本文不再叙述。
使用Rgb24和Lab24两个struct定义Rgb彩色空间的像素和Lab彩色空间的像素。
public partial struct Rgb24
{
public static Rgb24 WHITE = new Rgb24 { Red = 255, Green = 255, Blue = 255 };
public static Rgb24 BLACK = new Rgb24();
public static Rgb24 RED = new Rgb24 { Red = 255 };
public static Rgb24 BLUE = new Rgb24 { Blue = 255 };
public static Rgb24 GREEN = new Rgb24 { Green = 255 };
[FieldOffset(0)]
public Byte Blue;
[FieldOffset(1)]
public Byte Green;
[FieldOffset(2)]
public Byte Red;
public Rgb24(int red, int green, int blue)
{
Red = (byte)red;
Green = (byte)green;
Blue = (byte)blue;
}
public Rgb24(byte red, byte green, byte blue)
{
Red = red;
Green = green;
Blue = blue;
}
}
public partial struct Lab24
{
public byte L;
public byte A;
public byte B;
public Lab24(byte l,byte a,byte b)
{
L = l;
A = a;
B = b;
}
public Lab24(int l,int a,int b)
{
L = (byte)l;
A = (byte)a;
B = (byte)b;
}
}
Lab空间参照OpenCV，用一个byte来表示Lab空间的每个通道值，以求提高性能。由于标准的Lab空间中a和b通道是可付的，Lab24中的A、B值减去128，就是标准Lab空间的a，b通道值。
2. Rgb24 <=> Lab24 的实现
OpenCV中Bgr<=>Lab是用C语言实现的，下面将它转换为C#代码：
public sealed class UnmanagedImageConverter
{
/* 1024*(([0..511]./255)**(1./3)) */
static ushort[] icvLabCubeRootTab = new ushort[] {
0,161,203,232,256,276,293,308,322,335,347,359,369,379,389,398,
406,415,423,430,438,445,452,459,465,472,478,484,490,496,501,507,
512,517,523,528,533,538,542,547,552,556,561,565,570,574,578,582,
586,590,594,598,602,606,610,614,617,621,625,628,632,635,639,642,
645,649,652,655,659,662,665,668,671,674,677,680,684,686,689,692,
695,698,701,704,707,710,712,715,718,720,723,726,728,731,734,736,
739,741,744,747,749,752,754,756,759,761,764,766,769,771,773,776,
778,780,782,785,787,789,792,794,796,798,800,803,805,807,809,811,
813,815,818,820,822,824,826,828,830,832,834,836,838,840,842,844,
846,848,850,852,854,856,857,859,861,863,865,867,869,871,872,874,
876,878,880,882,883,885,887,889,891,892,894,896,898,899,901,903,
904,906,908,910,911,913,915,916,918,920,921,923,925,926,928,929,
931,933,934,936,938,939,941,942,944,945,947,949,950,952,953,955,
956,958,959,961,962,964,965,967,968,970,971,973,974,976,977,979,
980,982,983,985,986,987,989,990,992,993,995,996,997,999,1000,1002,
1003,1004,1006,1007,1009,1010,1011,1013,1014,1015,1017,1018,1019,1021,1022,1024,
1025,1026,1028,1029,1030,1031,1033,1034,1035,1037,1038,1039,1041,1042,1043,1044,
1046,1047,1048,1050,1051,1052,1053,1055,1056,1057,1058,1060,1061,1062,1063,1065,
1066,1067,1068,1070,1071,1072,1073,1074,1076,1077,1078,1079,1081,1082,1083,1084,
1085,1086,1088,1089,1090,1091,1092,1094,1095,1096,1097,1098,1099,1101,1102,1103,
1104,1105,1106,1107,1109,1110,1111,1112,1113,1114,1115,1117,1118,1119,1120,1121,
1122,1123,1124,1125,1127,1128,1129,1130,1131,1132,1133,1134,1135,1136,1138,1139,
1140,1141,1142,1143,1144,1145,1146,1147,1148,1149,1150,1151,1152,1154,1155,1156,
1157,1158,1159,1160,1161,1162,1163,1164,1165,1166,1167,1168,1169,1170,1171,1172,
1173,1174,1175,1176,1177,1178,1179,1180,1181,1182,1183,1184,1185,1186,1187,1188,
1189,1190,1191,1192,1193,1194,1195,1196,1197,1198,1199,1200,1201,1202,1203,1204,
1205,1206,1207,1208,1209,1210,1211,1212,1213,1214,1215,1215,1216,1217,1218,1219,
1220,1221,1222,1223,1224,1225,1226,1227,1228,1229,1230,1230,1231,1232,1233,1234,
1235,1236,1237,1238,1239,1240,1241,1242,1242,1243,1244,1245,1246,1247,1248,1249,
1250,1251,1251,1252,1253,1254,1255,1256,1257,1258,1259,1259,1260,1261,1262,1263,
1264,1265,1266,1266,1267,1268,1269,1270,1271,1272,1273,1273,1274,1275,1276,1277,
1278,1279,1279,1280,1281,1282,1283,1284,1285,1285,1286,1287,1288,1289,1290,1291
};
const float labXr_32f = 0.433953f /* = xyzXr_32f / 0.950456 */;
const float labXg_32f = 0.376219f /* = xyzXg_32f / 0.950456 */;
const float labXb_32f = 0.189828f /* = xyzXb_32f / 0.950456 */;
const float labYr_32f = 0.212671f /* = xyzYr_32f */;
const float labYg_32f = 0.715160f /* = xyzYg_32f */;
const float labYb_32f = 0.072169f /* = xyzYb_32f */;
const float labZr_32f = 0.017758f /* = xyzZr_32f / 1.088754 */;
const float labZg_32f = 0.109477f /* = xyzZg_32f / 1.088754 */;
const float labZb_32f = 0.872766f /* = xyzZb_32f / 1.088754 */;
const float labRx_32f = 3.0799327f /* = xyzRx_32f * 0.950456 */;
const float labRy_32f = (-1.53715f) /* = xyzRy_32f */;
const float labRz_32f = (-0.542782f)/* = xyzRz_32f * 1.088754 */;
const float labGx_32f = (-0.921235f)/* = xyzGx_32f * 0.950456 */;
const float labGy_32f = 1.875991f /* = xyzGy_32f */ ;
const float labGz_32f = 0.04524426f /* = xyzGz_32f * 1.088754 */;
const float labBx_32f = 0.0528909755f /* = xyzBx_32f * 0.950456 */;
const float labBy_32f = (-0.204043f) /* = xyzBy_32f */;
const float labBz_32f = 1.15115158f /* = xyzBz_32f * 1.088754 */;
const float labT_32f = 0.008856f;
const int lab_shift = 10;
const float labLScale2_32f = 903.3f;
const int labXr = (int)((labXr_32f) * (1 << (lab_shift)) + 0.5);
const int labXg = (int)((labXg_32f) * (1 << (lab_shift)) + 0.5);
const int labXb = (int)((labXb_32f) * (1 << (lab_shift)) + 0.5);
const int labYr = (int)((labYr_32f) * (1 << (lab_shift)) + 0.5);
const int labYg = (int)((labYg_32f) * (1 << (lab_shift)) + 0.5);
const int labYb = (int)((labYb_32f) * (1 << (lab_shift)) + 0.5);
const int labZr = (int)((labZr_32f) * (1 << (lab_shift)) + 0.5);
const int labZg = (int)((labZg_32f) * (1 << (lab_shift)) + 0.5);
const int labZb = (int)((labZb_32f) * (1 << (lab_shift)) + 0.5);
const float labLScale_32f = 116.0f;
const float labLShift_32f = 16.0f;
const int labSmallScale = (int)((31.27 /* labSmallScale_32f*(1<<lab_shift)/255 */ ) * (1 << (lab_shift)) + 0.5);
const int labSmallShift = (int)((141.24138 /* labSmallScale_32f*(1<<lab) */ ) * (1 << (lab_shift)) + 0.5);
const int labT = (int)((labT_32f * 255) * (1 << (lab_shift)) + 0.5);
const int labLScale = (int)((295.8) * (1 << (lab_shift)) + 0.5);
const int labLShift = (int)((41779.2) * (1 << (lab_shift)) + 0.5);
const int labLScale2 = (int)((labLScale2_32f * 0.01) * (1 << (lab_shift)) + 0.5);
public static unsafe void ToLab24(Rgb24* from, Lab24* to)
{
ToLab24(from,to,1);
}
public static unsafe void ToLab24(Rgb24* from, Lab24* to, int length)
{
// 使用 OpenCV 中的算法实现
if (length < 1) return;
Rgb24* end = from + length;
int x, y, z;
int l, a, b;
bool flag;
while (from != end)
{
Byte red = from->Red;
Byte green = from->Green;
Byte blue = from->Blue;
x = blue * labXb + green * labXg + red * labXr;
y = blue * labYb + green * labYg + red * labYr;
z = blue * labZb + green * labZg + red * labZr;
flag = x > labT;
x = (((x) + (1 << ((lab_shift) - 1))) >> (lab_shift));
if (flag)
x = icvLabCubeRootTab[x];
else
x = (((x * labSmallScale + labSmallShift) + (1 << ((lab_shift) - 1))) >> (lab_shift));
flag = z > labT;
z = (((z) + (1 << ((lab_shift) - 1))) >> (lab_shift));
if (flag == true)
z = icvLabCubeRootTab[z];
else
z = (((z * labSmallScale + labSmallShift) + (1 << ((lab_shift) - 1))) >> (lab_shift));
flag = y > labT;
y = (((y) + (1 << ((lab_shift) - 1))) >> (lab_shift));
if (flag == true)
{
y = icvLabCubeRootTab[y];
l = (((y * labLScale - labLShift) + (1 << ((2 * lab_shift) - 1))) >> (2 * lab_shift));
}
else
{
l = (((y * labLScale2) + (1 << ((lab_shift) - 1))) >> (lab_shift));
y = (((y * labSmallScale + labSmallShift) + (1 << ((lab_shift) - 1))) >> (lab_shift));
}
a = (((500 * (x - y)) + (1 << ((lab_shift) - 1))) >> (lab_shift)) + 129;
b = (((200 * (y - z)) + (1 << ((lab_shift) - 1))) >> (lab_shift)) + 128;
l = l > 255 ? 255 : l < 0 ? 0 : l;
a = a > 255 ? 255 : a < 0 ? 0 : a;
b = b > 255 ? 255 : b < 0 ? 0 : b;
to->L = (byte)l;
to->A = (byte)a;
to->B = (byte)b;
from++;
to++;
}
}
public static unsafe void ToRgb24(Lab24* from, Rgb24* to)
{
ToRgb24(from,to,1);
}
public static unsafe void ToRgb24(Lab24* from, Rgb24* to, int length)
{
if (length < 1) return;
// 使用 OpenCV 中的算法实现
const float coeff0 = 0.39215686274509809f;
const float coeff1 = 0.0f;
const float coeff2 = 1.0f;
const float coeff3 = (-128.0f);
const float coeff4 = 1.0f;
const float coeff5 = (-128.0f);
if (length < 1) return;
Lab24* end = from + length;
float x, y, z,l,a,b;
int blue, green, red;
while (from != end)
{
l = from->L * coeff0 + coeff1;
a = from->A * coeff2 + coeff3;
b = from->B * coeff4 + coeff5;
l = (l + labLShift_32f) * (1.0f / labLScale_32f);
x = (l + a * 0.002f);
z = (l - b * 0.005f);
y = l * l * l;
x = x * x * x;
z = z * z * z;
blue = (int)((x * labBx_32f + y * labBy_32f + z * labBz_32f) * 255 + 0.5);
green = (int)((x * labGx_32f + y * labGy_32f + z * labGz_32f) * 255 + 0.5);
red = (int)((x * labRx_32f + y * labRy_32f + z * labRz_32f) * 255 + 0.5);
red = red < 0 ? 0 : red > 255 ? 255 : red;
green = green < 0 ? 0 : green > 255 ? 255 : green;
blue = blue < 0 ? 0 : blue > 255 ? 255 : blue;
to->Red = (byte)red;
to->Green = (byte)green;
to->Blue = (byte)blue;
from++;
to++;
}
}
}
由于C代码中使用了宏，在改写成C#代码时需要手动内联，以提高性能。上面的代码已经实现手动内联。
3. (A)C#实现与(B)C实现的性能对比（C# vs. OpenCV/PInvoke）
C# 版本（ImageRgb24 代表一幅Rgb24图像，ImageLab24代表一幅Lab24图像，它们之间的变化是调用上文UnmanagedImageConverter中的方法实现的）：
Stopwatch sw = new Stopwatch();
sw.Start();
ImageLab24 imgLab = null;
imgLab = new ImageLab24(img); // img 是一个 ImageRgb24 对象
sw.Stop();
Message = sw.ElapsedMilliseconds.ToString();
OpenCV版本（使用EmguCV对OpenCV的PInvoke封装）
private Image<Lab,Byte> TestOpenCV()
{
Image<Bgr, Byte> imgBgr = new Image<Bgr, byte>(imgMain.Image as Bitmap);
Image<Lab,Byte> imgLab = new Image<Lab,byte>(new Size(imgBgr.Width, imgBgr.Height));
Stopwatch sw = new Stopwatch();
sw.Start();
CvInvoke.cvCvtColor(imgBgr.Ptr,imgLab.Ptr, Emgu.CV.CvEnum.COLOR_CONVERSION.CV_BGR2Lab);
sw.Stop();
MessageBox.Show(sw.ElapsedMilliseconds.ToString() + "ms");
return imgLab;
}
下面针对三副不同大小的图像进行测试，每张图像测试4次，每次测试将上面两种实现各跑一次，前2次，先跑OpenCV/PInvoke实现，后2次，先跑C#实现，单位皆为ms。
图像1，大小：485×342
A: 5 3 5 3
B: 41 5 6 2
图像2，大小：1845×611
A：25 23 23 23
B：23 34 20 21
图像3，大小：3888×2592
A：209 210 211 210
B：185 188 191 185
从测试结果可以看出，C# 和 OpenCV/PInvoke的性能极为接近。
4. 进一步改进性能
偏色、高光检测等不需要多么准确的Rgb=>Lab转换。如果把彩色图像的每个通道用4 bit来表示，则一共有 4096 种颜色，完全可以用查表方式来加速计算。用一个Lab24数组来表示Rgb24到Lab24空间的映射：
Lab24[] ColorMap
首先初始化ColorMap：
ColorMap = new Lab24[4096];
for (int r = 0; r < 16; r++)
{
for (int g = 0; g < 16; g++)
{
for (int b = 0; b < 16; b++)
{
Rgb24 rgb = new Rgb24(r * 16, g * 16, b * 16);
Lab24 lab = Lab24.CreateFrom(rgb);
ColorMap[(r << 8) + (g << 4) + b] = lab;
}
}
}
然后，查表进行转换：
private unsafe ImageLab24 ConvertToImageLab24(ImageRgb24 img)
{
ImageLab24 lab = new ImageLab24(img.Width, img.Height);
Lab24* labStart = lab.Start;
Rgb24* rgbStart = img.Start;
Rgb24* rgbEnd = img.Start + img.Length;
while (rgbStart != rgbEnd)
{
Rgb24 rgb = *rgbStart;
*labStart = ColorMap[(((int)(rgb.Red) >> 4) << 8) + (((int)(rgb.Green) >> 4) << 4) + ((int)(rgb.Blue) >> 4) ];
rgbStart++;
labStart++;
}
return lab;
}
下面测试（C）查表计算的性能，结果和(A)C#实现与(B)C实现放在一起做对比。
图像1，大小：485×342
A: 5 3 5 3
B: 41 5 6 2
C: 3 2 2 2
图像2，大小：1845×611
A：25 23 23 23
B：23 34 20 21
C: 15 15 15 15
图像3，大小：3888×2592
A：209 210 211 210
B：185 188 191 185
C: 136 134 135 135
5. 原地进行变换
还可以进一步提高性能，因为Rgb24和Lab24大小一样，可以在原地进行Rgb24=>Lab24的变换。相应代码如下：
Rgb24[] ColorMapInSpace
...
ColorMap = new Lab24[4096];
ColorMapInSpace = new Rgb24[4096];
for (int r = 0; r < 16; r++)
{
for (int g = 0; g < 16; g++)
{
for (int b = 0; b < 16; b++)
{
Rgb24 rgb = new Rgb24(r * 16, g * 16, b * 16);
Lab24 lab = Lab24.CreateFrom(rgb);
ColorMap[(r << 8) + (g << 4) + b] = lab;
ColorMapInSpace[(r << 8) + (g << 4) + b] = new Rgb24(lab.L,lab.A,lab.B);
}
}
}
private unsafe void ConvertToImageLab24InSpace(ImageRgb24 img)
{
Rgb24* rgbStart = img.Start;
Rgb24* rgbEnd = img.Start + img.Length;
while (rgbStart != rgbEnd)
{
Rgb24 rgb = *rgbStart;
*rgbStart = ColorMapInSpace[(((int)(rgb.Red) >> 4) << 8) + (((int)(rgb.Green) >> 4) << 4) + ((int)(rgb.Blue) >> 4)];
rgbStart++;
}
}
下面测试D（原地查表变换）的性能，结果和(A)C#实现、(B)C实现、(C)查表计算进行比较：
图像1，大小：485×342
A: 5 3 5 3
B: 41 5 6 2
C: 3 2 2 2
D: 2 1 2 1
图像2，大小：1845×611
A：25 23 23 23
B：23 34 20 21
C: 15 15 15 15
D: 13 13 13 13
图像3，大小：3888×2592
A：209 210 211 210
B：185 188 191 185
C: 136 134 135 135
D: 117 118 122 117
6. 为什么用C#而不是C/C++
经常有人问，你为什么用C#而不用C/C++写图像处理程序。原因如下：
（1）C# 打开unsafe后，写的程序性能非常接近 C 程序的性能（当然，用不了SIMD是个缺陷。mono暂时不考虑。可通过挂接一个轻量级的C库来解决。）；
（2）写C#代码比写C代码爽多了快多了（命名空间、不用管头文件、快速编译、重构、生成API文档 ……）；
（3）庞大的.Net Framework是强有力的后盾。比如，客户想看演示，用Asp.Net写个页面，传个图片给后台，处理了显示出来。还有那些非性能攸关的地方，可以大量使用.Net Framework中的类，大幅度减少开发时间；
（4）结合强大的WPF，可以快速实现复杂的功能
（5）大量的时间在算法研究、实现和优化上，用C#可以把那些无关的惹人烦的事情给降到最小，所牺牲的只是一丁点儿性能。如果生产平台没有.net环境，将C#代码转换为C/C++代码也很快。
相关阅读 更多 +
排行榜 更多 +
