最近项目需要检测图像是否存在偏色、过亮、模糊等缺陷。由于主要用在视频监控上，对性能要求比较高。有几项检测必须要在Lab彩色下进行，而众所周知Rgb => Lab 计算量较大，C#搞得定搞不定？测试表明，用纯C#编写的Rgb => Lab代码在性能上与C编写的Rgb => Lab代码极为接近。
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均匀彩色空间，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。
Rgb24 与 Lab24 
2 {
3 public static Rgb24 WHITE = new Rgb24 { Red = 255, Green = 255, Blue = 255 };
4 public static Rgb24 BLACK = new Rgb24();
5 public static Rgb24 RED = new Rgb24 { Red = 255 };
6 public static Rgb24 BLUE = new Rgb24 { Blue = 255 };
7 public static Rgb24 GREEN = new Rgb24 { Green = 255 };
8
9 [FieldOffset(0)]
10 public Byte Blue;
11 [FieldOffset(1)]
12 public Byte Green;
13 [FieldOffset(2)]
14 public Byte Red;
15
16 public Rgb24(int red, int green, int blue)
17 {
18 Red = (byte)red;
19 Green = (byte)green;
20 Blue = (byte)blue;
21 }
22
23 public Rgb24(byte red, byte green, byte blue)
24 {
25 Red = red;
26 Green = green;
27 Blue = blue;
28 }
29 }
30
31 public partial struct Lab24
32 {
33 public byte L;
34 public byte A;
35 public byte B;
36
37 public Lab24(byte l,byte a,byte b)
38 {
39 L = l;
40 A = a;
41 B = b;
42 }
43
44 public Lab24(int l,int a,int b)
45 {
46 L = (byte)l;
47 A = (byte)a;
48 B = (byte)b;
49 }
50 }
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：
Rgb24 <=> Lab24 2 {
3 /* 1024*(([0..511]./255)**(1./3)) */
4 static ushort[] icvLabCubeRootTab = new ushort[] {
5 0,161,203,232,256,276,293,308,322,335,347,359,369,379,389,398,
6 406,415,423,430,438,445,452,459,465,472,478,484,490,496,501,507,
7 512,517,523,528,533,538,542,547,552,556,561,565,570,574,578,582,
8 586,590,594,598,602,606,610,614,617,621,625,628,632,635,639,642,
9 645,649,652,655,659,662,665,668,671,674,677,680,684,686,689,692,
10 695,698,701,704,707,710,712,715,718,720,723,726,728,731,734,736,
11 739,741,744,747,749,752,754,756,759,761,764,766,769,771,773,776,
12 778,780,782,785,787,789,792,794,796,798,800,803,805,807,809,811,
13 813,815,818,820,822,824,826,828,830,832,834,836,838,840,842,844,
14 846,848,850,852,854,856,857,859,861,863,865,867,869,871,872,874,
15 876,878,880,882,883,885,887,889,891,892,894,896,898,899,901,903,
16 904,906,908,910,911,913,915,916,918,920,921,923,925,926,928,929,
17 931,933,934,936,938,939,941,942,944,945,947,949,950,952,953,955,
18 956,958,959,961,962,964,965,967,968,970,971,973,974,976,977,979,
19 980,982,983,985,986,987,989,990,992,993,995,996,997,999,1000,1002,
20 1003,1004,1006,1007,1009,1010,1011,1013,1014,1015,1017,1018,1019,1021,1022,1024,
21 1025,1026,1028,1029,1030,1031,1033,1034,1035,1037,1038,1039,1041,1042,1043,1044,
22 1046,1047,1048,1050,1051,1052,1053,1055,1056,1057,1058,1060,1061,1062,1063,1065,
23 1066,1067,1068,1070,1071,1072,1073,1074,1076,1077,1078,1079,1081,1082,1083,1084,
24 1085,1086,1088,1089,1090,1091,1092,1094,1095,1096,1097,1098,1099,1101,1102,1103,
25 1104,1105,1106,1107,1109,1110,1111,1112,1113,1114,1115,1117,1118,1119,1120,1121,
26 1122,1123,1124,1125,1127,1128,1129,1130,1131,1132,1133,1134,1135,1136,1138,1139,
27 1140,1141,1142,1143,1144,1145,1146,1147,1148,1149,1150,1151,1152,1154,1155,1156,
28 1157,1158,1159,1160,1161,1162,1163,1164,1165,1166,1167,1168,1169,1170,1171,1172,
29 1173,1174,1175,1176,1177,1178,1179,1180,1181,1182,1183,1184,1185,1186,1187,1188,
30 1189,1190,1191,1192,1193,1194,1195,1196,1197,1198,1199,1200,1201,1202,1203,1204,
31 1205,1206,1207,1208,1209,1210,1211,1212,1213,1214,1215,1215,1216,1217,1218,1219,
32 1220,1221,1222,1223,1224,1225,1226,1227,1228,1229,1230,1230,1231,1232,1233,1234,
33 1235,1236,1237,1238,1239,1240,1241,1242,1242,1243,1244,1245,1246,1247,1248,1249,
34 1250,1251,1251,1252,1253,1254,1255,1256,1257,1258,1259,1259,1260,1261,1262,1263,
35 1264,1265,1266,1266,1267,1268,1269,1270,1271,1272,1273,1273,1274,1275,1276,1277,
36 1278,1279,1279,1280,1281,1282,1283,1284,1285,1285,1286,1287,1288,1289,1290,1291
37 };
38
39 const float labXr_32f = 0.433953f /* = xyzXr_32f / 0.950456 */;
40 const float labXg_32f = 0.376219f /* = xyzXg_32f / 0.950456 */;
41 const float labXb_32f = 0.189828f /* = xyzXb_32f / 0.950456 */;
42
43 const float labYr_32f = 0.212671f /* = xyzYr_32f */;
44 const float labYg_32f = 0.715160f /* = xyzYg_32f */;
45 const float labYb_32f = 0.072169f /* = xyzYb_32f */;
46
47 const float labZr_32f = 0.017758f /* = xyzZr_32f / 1.088754 */;
48 const float labZg_32f = 0.109477f /* = xyzZg_32f / 1.088754 */;
49 const float labZb_32f = 0.872766f /* = xyzZb_32f / 1.088754 */;
50
51 const float labRx_32f = 3.0799327f /* = xyzRx_32f * 0.950456 */;
52 const float labRy_32f = (-1.53715f) /* = xyzRy_32f */;
53 const float labRz_32f = (-0.542782f)/* = xyzRz_32f * 1.088754 */;
54
55 const float labGx_32f = (-0.921235f)/* = xyzGx_32f * 0.950456 */;
56 const float labGy_32f = 1.875991f /* = xyzGy_32f */ ;
57 const float labGz_32f = 0.04524426f /* = xyzGz_32f * 1.088754 */;
58
59 const float labBx_32f = 0.0528909755f /* = xyzBx_32f * 0.950456 */;
60 const float labBy_32f = (-0.204043f) /* = xyzBy_32f */;
61 const float labBz_32f = 1.15115158f /* = xyzBz_32f * 1.088754 */;
62
63 const float labT_32f = 0.008856f;
64
65 const int lab_shift = 10;
66
67 const float labLScale2_32f = 903.3f;
68
69 const int labXr = (int)((labXr_32f) * (1 << (lab_shift)) + 0.5);
70 const int labXg = (int)((labXg_32f) * (1 << (lab_shift)) + 0.5);
71 const int labXb = (int)((labXb_32f) * (1 << (lab_shift)) + 0.5);
72
73 const int labYr = (int)((labYr_32f) * (1 << (lab_shift)) + 0.5);
74 const int labYg = (int)((labYg_32f) * (1 << (lab_shift)) + 0.5);
75 const int labYb = (int)((labYb_32f) * (1 << (lab_shift)) + 0.5);
76
77 const int labZr = (int)((labZr_32f) * (1 << (lab_shift)) + 0.5);
78 const int labZg = (int)((labZg_32f) * (1 << (lab_shift)) + 0.5);
79 const int labZb = (int)((labZb_32f) * (1 << (lab_shift)) + 0.5);
80
81 const float labLScale_32f = 116.0f;
82 const float labLShift_32f = 16.0f;
83
84 const int labSmallScale = (int)((31.27 /* labSmallScale_32f*(1<<lab_shift)/255 */ ) * (1 << (lab_shift)) + 0.5);
85
86 const int labSmallShift = (int)((141.24138 /* labSmallScale_32f*(1<<lab) */ ) * (1 << (lab_shift)) + 0.5);
87
88 const int labT = (int)((labT_32f * 255) * (1 << (lab_shift)) + 0.5);
89
90 const int labLScale = (int)((295.8) * (1 << (lab_shift)) + 0.5);
91 const int labLShift = (int)((41779.2) * (1 << (lab_shift)) + 0.5);
92 const int labLScale2 = (int)((labLScale2_32f * 0.01) * (1 << (lab_shift)) + 0.5);
93
94 public static unsafe void ToLab24(Rgb24* from, Lab24* to)
95 {
96 ToLab24(from,to,1);
97 }
98
99 public static unsafe void ToLab24(Rgb24* from, Lab24* to, int length)
100 {
101 // 使用 OpenCV 中的算法实现
102
103 if (length < 1) return;
104
105 Rgb24* end = from + length;
106
107 int x, y, z;
108 int l, a, b;
109 bool flag;
110
111 while (from != end)
112 {
113 Byte red = from->Red;
114 Byte green = from->Green;
115 Byte blue = from->Blue;
116
117 x = blue * labXb + green * labXg + red * labXr;
118 y = blue * labYb + green * labYg + red * labYr;
119 z = blue * labZb + green * labZg + red * labZr;
120
121 flag = x > labT;
122
123 x = (((x) + (1 << ((lab_shift) - 1))) >> (lab_shift));
124
125 if (flag)
126 x = icvLabCubeRootTab[x];
127 else
128 x = (((x * labSmallScale + labSmallShift) + (1 << ((lab_shift) - 1))) >> (lab_shift));
129
130 flag = z > labT;
131 z = (((z) + (1 << ((lab_shift) - 1))) >> (lab_shift));
132
133 if (flag == true)
134 z = icvLabCubeRootTab[z];
135 else
136 z = (((z * labSmallScale + labSmallShift) + (1 << ((lab_shift) - 1))) >> (lab_shift));
137
138 flag = y > labT;
139 y = (((y) + (1 << ((lab_shift) - 1))) >> (lab_shift));
140
141 if (flag == true)
142 {
143 y = icvLabCubeRootTab[y];
144 l = (((y * labLScale - labLShift) + (1 << ((2 * lab_shift) - 1))) >> (2 * lab_shift));
145 }
146 else
147 {
148 l = (((y * labLScale2) + (1 << ((lab_shift) - 1))) >> (lab_shift));
149 y = (((y * labSmallScale + labSmallShift) + (1 << ((lab_shift) - 1))) >> (lab_shift));
150 }
151
152 a = (((500 * (x - y)) + (1 << ((lab_shift) - 1))) >> (lab_shift)) + 129;
153 b = (((200 * (y - z)) + (1 << ((lab_shift) - 1))) >> (lab_shift)) + 128;
154
155 l = l > 255 ? 255 : l < 0 ? 0 : l;
156 a = a > 255 ? 255 : a < 0 ? 0 : a;
157 b = b > 255 ? 255 : b < 0 ? 0 : b;
158
159 to->L = (byte)l;
160 to->A = (byte)a;
161 to->B = (byte)b;
162
163 from++;
164 to++;
165 }
166 }
167
168 public static unsafe void ToRgb24(Lab24* from, Rgb24* to)
169 {
170 ToRgb24(from,to,1);
171 }
172
173 public static unsafe void ToRgb24(Lab24* from, Rgb24* to, int length)
174 {
175 if (length < 1) return;
176
177 // 使用 OpenCV 中的算法实现
178 const float coeff0 = 0.39215686274509809f;
179 const float coeff1 = 0.0f;
180 const float coeff2 = 1.0f;
181 const float coeff3 = (-128.0f);
182 const float coeff4 = 1.0f;
183 const float coeff5 = (-128.0f);
184
185 if (length < 1) return;
186
187 Lab24* end = from + length;
188 float x, y, z,l,a,b;
189 int blue, green, red;
190
191 while (from != end)
192 {
193 l = from->L * coeff0 + coeff1;
194 a = from->A * coeff2 + coeff3;
195 b = from->B * coeff4 + coeff5;
196
197 l = (l + labLShift_32f) * (1.0f / labLScale_32f);
198 x = (l + a * 0.002f);
199 z = (l - b * 0.005f);
200
201 y = l * l * l;
202 x = x * x * x;
203 z = z * z * z;
204
205 blue = (int)((x * labBx_32f + y * labBy_32f + z * labBz_32f) * 255 + 0.5);
206 green = (int)((x * labGx_32f + y * labGy_32f + z * labGz_32f) * 255 + 0.5);
207 red = (int)((x * labRx_32f + y * labRy_32f + z * labRz_32f) * 255 + 0.5);
208
209 red = red < 0 ? 0 : red > 255 ? 255 : red;
210 green = green < 0 ? 0 : green > 255 ? 255 : green;
211 blue = blue < 0 ? 0 : blue > 255 ? 255 : blue;
212
213 to->Red = (byte)red;
214 to->Green = (byte)green;
215 to->Blue = (byte)blue;
216
217 from++;
218 to++;
219 }
220 }
221 }
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。
====
补充测试VC 9.0 版本
VC 实现与 C# 实现略有区别，C#版本RGB,Lab使用struct来表示，VC下直接用的三个Byte Channel来表示，然后以 redChannel, greenChannel, blueChannel 来代表不同的 Channel Offset。以 nChannel 代表 Channel 数量。VC下有Stride，C#下无Stride。查表实现也和C#版本有区别，直接使用的是静态的表。O2优化。
E: 非查表实现
void
::ImageQualityDetector::ConvertToLab(Orc::ImageInfo &img)
{
static unsigned short icvLabCubeRootTab[] = {
0,161,203…… };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);int width = img.Width;
int height = img.Height;
int nChannel = img.NChannel;
int redChannel = img.RedChannel;
int greenChannel = img.GreenChannel;
int blueChannel = img.BlueChannel;
int x, y, z;
int l, a, b;
bool flag;for(int h = 0; h < height; h++)
{
byte *line = img.GetLine(h);
for(int w = 0; w < width; w++)
{
int red = line[redChannel];
int green = line[greenChannel];
int blue = line[blueChannel];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;int index = 3 * (((red >> 4) << 8) + ((green >> 4) << 4) + (blue >> 4)) ;
line[0] = (byte)l;
line[1] = (byte)a;
line[2] = (byte)b;line += nChannel;
}
}
}
F: 查表实现
void
::ImageQualityDetector::FastConvertToLab(Orc::ImageInfo &img)
{
static const byte Rgb2LabSmallTable[] = {
0, 129, 128 ……
};int width = img.Width;
int height = img.Height;
int nChannel = img.NChannel;
int redChannel = img.RedChannel;
int greenChannel = img.GreenChannel;
int blueChannel = img.BlueChannel;
for(int h = 0; h < height; h++)
{
byte *line = img.GetLine(h);
for(int w = 0; w < width; w++)
{
int red = line[redChannel];
int green = line[greenChannel];
int blue = line[blueChannel];
int index = 3 * (((red >> 4) << 8) + ((green >> 4) << 4) + (blue >> 4)) ;
line[0] = Rgb2LabSmallTable[index];
line[1] = Rgb2LabSmallTable[index + 1];
line[2] = Rgb2LabSmallTable[index + 2];
line += nChannel;
}
}
}
测试结果：
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
E: 32 30 37 37
F: 15 10 13 11
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
E: 242 240 243 239
F: 70 69 67 67
====
补充测试：C# 下查表实现（Byte数组）
G: C#下直接查找Byte数组，相关代码
static byte[] Rgb2LabSmallTable = new byte[] {
0, 129, 128, … }
private unsafe void ConvertToImageLab24Fast(ImageRgb24 img)
{
Rgb24* rgbStart = img.Start;
Rgb24* rgbEnd = img.Start + img.Length;
while (rgbStart != rgbEnd)
{
Rgb24 rgb = *rgbStart;
int index = (((int)(rgb.Red) >> 4) << 8) + (((int)(rgb.Green) >> 4) << 4) + ((int)(rgb.Blue) >> 4);
rgbStart->Red = Rgb2LabSmallTable[index];
rgbStart->Green = Rgb2LabSmallTable[index+1];
rgbStart->Blue = Rgb2LabSmallTable[index+2];
rgbStart++;
}
}
测试结果：
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
E: 32 30 37 37
F: 15 10 13 11
G: 12 11 13 11
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
E: 242 240 243 239
F: 70 69 67 67
G: 64 64 65 64
====
补充测试：同一种实现下的C#和VC性能对比，附下载
下面消除两种语言的测试区别，C#版本查表时使用指针而非数组，VC下使用无Stride的Rgb24，相关测试代码见 下载链接 。
这又形成了4个测试用例：
H- C#，非查表；I-C#，查表; J-C++，非查表; K-C++，查表
C# 版为 .Net 4.0, VS2010 ，代码中选择快速一项为测试I，不选择为测试H。
C++版 - VS2008。选择快速一项为测试K，不选择为测试J。
测试结果：
图像2，大小：1845×611
H: 31 29 36 32
I: 10 10 10 10
J: 39 33 33 30
K: 9 8 8 8图像3，大小：3888×2592
H: 195 194 194 195
I: 53 52 51 52
J: 220 218 218 222
K: 41 42 41 41
结论：
C#下图像开发是很给力的！还在犹豫什么呢？