这是OpenCV图像处理专栏的第十一篇文章,之前介绍过两种处理白平衡的算法,分别为灰度世界算法和完美反射算法。今天来介绍另外一个自动白平衡的算法,即动态阈值法,一个看起来比较厉害的名字。论文原文链接放在附录。
和灰度世界法和完美反射算法类似,动态阈值算法仍然分为两个步骤即白点检测和白点调整,具体如下:
1、把尺寸为 w × h w\times h w×h的原图像从 R G B RGB RGB空间转换到 Y C r C b YCrCb YCrCb空间。
2、把图像分成 3 × 4 3\times 4 3×4个块。
3、对每个块分别计算 C r Cr Cr, C b Cb Cb的平均值 M r Mr Mr, M b Mb Mb。
4、判定每个块的近白区域(near-white region
)。判别准则为:
- C b ( i , j ) − ( M b + D b × s i g n ( M b ) ) < 1.5 × D b Cb(i, j) − (Mb + Db\times sign(Mb)) < 1.5\times Db Cb(i,j)−(Mb+Db×sign(Mb))<1.5×Db
- C r ( i , j ) − ( 1.5 × M r + D r × s i g n ( M r ) ) < 1.5 × D r Cr(i, j) − (1.5\times Mr + Dr \times sign(Mr )) < 1.5\times Dr Cr(i,j)−(1.5×Mr+Dr×sign(Mr))<1.5×Dr,其中sign为符号函数,即正数返回1,负数返回0。
5、设一个“参考白色点”的亮度矩阵 R L RL RL,大小为 w × h w\times h w×h。
6、若符合判别式,则作为“参考白色点”,并把该点 ( i , j ) (i,j) (i,j)的亮度( Y Y Y分量)值赋给 R L ( i , j ) RL(i,j) RL(i,j)。若不符合,则该点的 R L ( i , j ) RL(i,j) RL(i,j)值为0。
10%
的亮度(Y
分量)值,并选取其中的最小值Lu_min
。RL
,若RL(i,j), RL(i,j)=0
; 否则,RL(i,j)=1
。
R
,G
,B
与RL
相乘,得到R2
,G2
,B2
。 分别计算R2
,G2
,B2
的平均值,Rav
,Gav
,Bav
。Ymax=double(max(max(Y)))
,则Rgain=Ymax/Rav
,Ggain=Ymax/Gav
, Bgain=Ymax/Bav
。Ro= R*Rgain
; Go= G*Ggain
; Bo= B*Bgain
;块的大小取了100,没处理长或者宽不够100的结尾部分,这个可以自己添加。
const float YCbCrYRF = 0.299F; // RGB转YCbCr的系数(浮点类型)
const float YCbCrYGF = 0.587F;
const float YCbCrYBF = 0.114F;
const float YCbCrCbRF = -0.168736F;
const float YCbCrCbGF = -0.331264F;
const float YCbCrCbBF = 0.500000F;
const float YCbCrCrRF = 0.500000F;
const float YCbCrCrGF = -0.418688F;
const float YCbCrCrBF = -0.081312F;
const float RGBRYF = 1.00000F; // YCbCr转RGB的系数(浮点类型)
const float RGBRCbF = 0.0000F;
const float RGBRCrF = 1.40200F;
const float RGBGYF = 1.00000F;
const float RGBGCbF = -0.34414F;
const float RGBGCrF = -0.71414F;
const float RGBBYF = 1.00000F;
const float RGBBCbF = 1.77200F;
const float RGBBCrF = 0.00000F;
const int Shift = 20;
const int HalfShiftValue = 1 << (Shift - 1);
const int YCbCrYRI = (int)(YCbCrYRF * (1 << Shift) + 0.5); // RGB转YCbCr的系数(整数类型)
const int YCbCrYGI = (int)(YCbCrYGF * (1 << Shift) + 0.5);
const int YCbCrYBI = (int)(YCbCrYBF * (1 << Shift) + 0.5);
const int YCbCrCbRI = (int)(YCbCrCbRF * (1 << Shift) + 0.5);
const int YCbCrCbGI = (int)(YCbCrCbGF * (1 << Shift) + 0.5);
const int YCbCrCbBI = (int)(YCbCrCbBF * (1 << Shift) + 0.5);
const int YCbCrCrRI = (int)(YCbCrCrRF * (1 << Shift) + 0.5);
const int YCbCrCrGI = (int)(YCbCrCrGF * (1 << Shift) + 0.5);
const int YCbCrCrBI = (int)(YCbCrCrBF * (1 << Shift) + 0.5);
const int RGBRYI = (int)(RGBRYF * (1 << Shift) + 0.5); // YCbCr转RGB的系数(整数类型)
const int RGBRCbI = (int)(RGBRCbF * (1 << Shift) + 0.5);
const int RGBRCrI = (int)(RGBRCrF * (1 << Shift) + 0.5);
const int RGBGYI = (int)(RGBGYF * (1 << Shift) + 0.5);
const int RGBGCbI = (int)(RGBGCbF * (1 << Shift) + 0.5);
const int RGBGCrI = (int)(RGBGCrF * (1 << Shift) + 0.5);
const int RGBBYI = (int)(RGBBYF * (1 << Shift) + 0.5);
const int RGBBCbI = (int)(RGBBCbF * (1 << Shift) + 0.5);
const int RGBBCrI = (int)(RGBBCrF * (1 << Shift) + 0.5);
Mat RGB2YCbCr(Mat src) {
int row = src.rows;
int col = src.cols;
Mat dst(row, col, CV_8UC3);
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
int Blue = src.at(i, j)[0];
int Green = src.at(i, j)[1];
int Red = src.at(i, j)[2];
dst.at(i, j)[0] = (int)((YCbCrYRI * Red + YCbCrYGI * Green + YCbCrYBI * Blue + HalfShiftValue) >> Shift);
dst.at(i, j)[1] = (int)(128 + ((YCbCrCbRI * Red + YCbCrCbGI * Green + YCbCrCbBI * Blue + HalfShiftValue) >> Shift));
dst.at(i, j)[2] = (int)(128 + ((YCbCrCrRI * Red + YCbCrCrGI * Green + YCbCrCrBI * Blue + HalfShiftValue) >> Shift));
}
}
return dst;
}
Mat YCbCr2RGB(Mat src) {
int row = src.rows;
int col = src.cols;
Mat dst(row, col, CV_8UC3);
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
int Y = src.at(i, j)[0];
int Cb = src.at(i, j)[1] - 128;
int Cr = src.at(i, j)[2] - 128;
int Red = Y + ((RGBRCrI * Cr + HalfShiftValue) >> Shift);
int Green = Y + ((RGBGCbI * Cb + RGBGCrI * Cr + HalfShiftValue) >> Shift);
int Blue = Y + ((RGBBCbI * Cb + HalfShiftValue) >> Shift);
if (Red > 255) Red = 255; else if (Red < 0) Red = 0;
if (Green > 255) Green = 255; else if (Green < 0) Green = 0; // 编译后应该比三目运算符的效率高
if (Blue > 255) Blue = 255; else if (Blue < 0) Blue = 0;
dst.at(i, j)[0] = Blue;
dst.at(i, j)[1] = Green;
dst.at(i, j)[2] = Red;
}
}
return dst;
}
template
inline T sign(T const &input) {
return input >= 0 ? 1 : -1;
}
Mat AutomaticWhiteBalanceMethod(Mat src) {
int row = src.rows;
int col = src.cols;
if (src.channels() == 4) {
cvtColor(src, src, CV_BGRA2BGR);
}
Mat input = RGB2YCbCr(src);
Mat mark(row, col, CV_8UC1);
int sum = 0;
for (int i = 0; i < row; i += 100) {
for (int j = 0; j < col; j += 100) {
if (i + 100 < row && j + 100 < col) {
Rect rect(j, i, 100, 100);
Mat temp = input(rect);
Scalar global_mean = mean(temp);
double dr = 0, db = 0;
for (int x = 0; x < 100; x++) {
uchar *ptr = temp.ptr(x) + 1;
for (int y = 0; y < 100; y++) {
dr += pow(abs(*ptr - global_mean[1]), 2);
ptr++;
db += pow(abs(*ptr - global_mean[2]), 2);
ptr++;
ptr++;
}
}
dr /= 10000;
db /= 10000;
double cr_left_criteria = 1.5 * global_mean[1] + dr * sign(global_mean[1]);
double cr_right_criteria = 1.5 * dr;
double cb_left_criteria = global_mean[2] + db * sign(global_mean[2]);
double cb_right_criteria = 1.5 * db;
for (int x = 0; x < 100; x++) {
uchar *ptr = temp.ptr(x) + 1;
for (int y = 0; y < 100; y++) {
uchar cr = *ptr;
ptr++;
uchar cb = *ptr;
ptr++;
ptr++;
if ((cr - cb_left_criteria) < cb_right_criteria && (cb - cr_left_criteria) < cr_right_criteria) {
sum++;
mark.at(i + x, j + y) = 1;
}
else {
mark.at(i + x, j + y) = 0;
}
}
}
}
}
}
int Threshold = 0;
int Ymax = 0;
int Light[256] = { 0 };
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
if (mark.at(i, j) == 1) {
Light[(int)(input.at(i, j)[0])]++;
}
Ymax = max(Ymax, (int)(input.at(i, j)[0]));
}
}
printf("maxY: %d\n", Ymax);
int sum2 = 0;
for (int i = 255; i >= 0; i--) {
sum2 += Light[i];
if (sum2 >= sum * 0.1) {
Threshold = i;
break;
}
}
printf("Threshold: %d\n", Threshold);
printf("Sum: %d Sum2: %d\n", sum, sum2);
double Blue = 0;
double Green = 0;
double Red = 0;
int cnt2 = 0;
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
if (mark.at(i, j) == 1 && (int)(input.at(i, j)[0]) >= Threshold) {
Blue += 1.0 * src.at(i, j)[0];
Green += 1.0 * src.at(i, j)[1];
Red += 1.0 * src.at(i, j)[2];
cnt2++;
}
}
}
Blue /= cnt2;
Green /= cnt2;
Red /= cnt2;
printf("%.5f %.5f %.5f\n", Blue, Green, Red);
Mat dst(row, col, CV_8UC3);
double maxY = Ymax;
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
int B = (int)(maxY * src.at(i, j)[0] / Blue);
int G = (int)(maxY * src.at(i, j)[1] / Green);
int R = (int)(maxY * src.at(i, j)[2] / Red);
if (B > 255) B = 255; else if (B < 0) B = 0;
if (G > 255) G = 255; else if (G < 0) G = 0;
if (R > 255) R = 255; else if (R < 0) R = 0;
dst.at(i, j)[0] = B;
dst.at(i, j)[1] = G;
dst.at(i, j)[2] = R;
}
}
return dst;
}
图像均为算法处理前和处理后的顺序。
论文原文:http://140.112.114.62/bitstream/246246/200704191001444/1/01465458.pdf
参考文章:https://www.cnblogs.com/Imageshop/archive/2013/04/20/3032062.html
对比前面的灰度世界算法和完美反射算法后,这个算法的效果确实要好很多,原文的内容基本上我的博客就写完了,感兴趣可以再去读读原文。
欢迎关注我的微信公众号GiantPandaCV,期待和你一起交流机器学习,深度学习,图像算法,优化技术,比赛及日常生活等。