《改进非线性亮度提升模型的逆光图像恢复》论文复现
算法原理
这个论文是中文的,论文网址在这:http://html.rhhz.net/JSJYY/2017-2-564.htm 。算法原理可以直接在这个网址或者下载论文查看。
源码实现
//Method2
#define PI 3.1415926
double log2(double N) {
return log10(N) / log10(2.0);
}
Mat Inrbl(Mat src, double k) {
int row = src.rows;
int col = src.cols;
Mat dst(row, col, CV_64FC3);
Mat dsthsi(row, col, CV_64FC3);
//RGB2HSI
Mat H = Mat(row, col, CV_64FC1);
Mat S = Mat(row, col, CV_64FC1);
Mat I = Mat(row, col, CV_64FC1);
int mp[256] = { 0 };
double mp2[256] = { 0.0 };
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
double h, s, newi, th;
double B = (double)src.at<Vec3b>(i, j)[0] / 255.0;
double G = (double)src.at<Vec3b>(i, j)[1] / 255.0;
double R = (double)src.at<Vec3b>(i, j)[2] / 255.0;
double mi, mx;
if (R > G && R > B) {
mx = R;
mi = min(G, B);
}
else {
if (G > B) {
mx = G;
mi = min(R, B);
}
else {
mx = B;
mi = min(R, G);
}
}
newi = (R + G + B) / 3.0;
if (newi < 0) newi = 0;
else if (newi > 1) newi = 1.0;
if (newi == 0 || mx == mi) {
s = 0;
h = 0;
}
else {
s = 1 - mi / newi;
th = (R - G) * (R - G) + (R - B) * (G - B);
th = sqrt(th) + 1e-5;
th = acos(((R - G + R - B)*0.5) / th);
if (G >= B) h = th;
else h = 2 * PI - th;
}
h = h / (2 * PI);
H.at<double>(i, j) = h;
S.at<double>(i, j) = s;
I.at<double>(i, j) = newi;
dsthsi.at<Vec3d>(i, j)[2] = h;
dsthsi.at<Vec3d>(i, j)[1] = s;
dsthsi.at<Vec3d>(i, j)[0] = newi;
mp[(int)((src.at<Vec3b>(i, j)[0] + src.at<Vec3b>(i, j)[1] + src.at<Vec3b>(i, j)[2]) / 3)]++;
}
}
//Ostu阈值
for (int i = 0; i < 256; i++) {
mp2[i] = (double)mp[i] / (double)(row * col);
}
double mI = 0;
for (int i = 0; i < 256; i++) {
mI += (i / 255.0) * mp2[i];
}
double var = 0;
double ThresHold = 0;
for (int i = 0; i < 256; i++) {
double T = 1.0 * i / 256;
double P1 = 0.0;
double mT = 0.0;
for (int j = 0; j <= i; j++) {
P1 += mp2[j];
mT += (double)(j / 255.0) * mp2[j];
}
if (P1 == 0) continue;
if (((mI*P1 - mT)*(mI*P1 - mT) / (P1*(1 - P1))) > var) {
var = (mI*P1 - mT)*(mI*P1 - mT) / (P1*(1 - P1));
ThresHold = T;
}
}
//
int cnt = 0;
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
if (I.at<double>(i, j) <= ThresHold) {
cnt++;
}
}
}
printf("cnt: %d\n", cnt);
double A = k * sqrt((double)cnt / (double)(row * col - cnt));
printf("A: %.5f\n", A);
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
double D, C;
if (I.at<double>(i, j) <= ThresHold) {
D = A;
C = 1.0 / log2(D + 1);
}
else {
D = (double)(ThresHold * A - ThresHold) / double((1 - ThresHold) * (I.at<double>(i, j))) - (double)(ThresHold * A - 1) / (1 - ThresHold);
C = 1.0 / log2(D + 1);
}
I.at<double>(i, j) = (C * log2(D * (double)I.at<double>(i, j) + 1));
}
}//
//HSI2RGB
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
double preh = H.at<double>(i, j) * 2 * PI;//?
double pres = S.at<double>(i, j);
double prei = I.at<double>(i, j);
double r = 0, g = 0, b = 0;
double t1, t2, t3;
t1 = (1.0 - pres) / 3.0;
if (preh >= 0 && preh < (PI * 2 / 3)) {
b = t1;
t2 = pres * cos(preh);
t3 = cos(PI / 3 - preh);
r = (1 + t2 / t3) / 3;
//g = 1.0 - r - b;
r = 3 * prei * r;
//g = 3 * g * prei;
b = 3 * prei * b;
g = 3 * prei - (r + b);
}
else if (preh >= (PI * 2 / 3) && preh < (PI * 4 / 3)) {
r = t1;
t2 = pres * cos(preh - 2 * PI / 3);
t3 = cos(PI - preh);
g = (1 + t2 / t3) / 3;
//b = 1 - r - g;
r = 3 * prei * r;
g = 3 * g * prei;
//b = 3 * prei * b;
b = 3 * prei - (r + g);
}
else if (preh >= (PI * 4 / 3) && preh <= (PI * 2)) {
g = t1;
t2 = pres * cos(preh - 4 * PI / 3);
t3 = cos(PI * 5 / 3 - preh);
b = (1 + t2 / t3) / 3;
//r = 1 - g - b;
//r = 3 * prei * r;
g = 3 * g * prei;
b = 3 * prei * b;
r = 3 * prei - (g + b);
}
dst.at<Vec3d>(i, j)[0] = b;
dst.at<Vec3d>(i, j)[1] = g;
dst.at<Vec3d>(i, j)[2] = r;
/*
dst.at<Vec3b>(i, j)[0] =(b * 255.0);
dst.at<Vec3b>(i, j)[1] = (int)(g * 255.0);
dst.at<Vec3b>(i, j)[2] = (int)(r * 255.0);*/
//printf("%d %d %d\n", (int)(r*255), (int)(g*255), (int)(b*255));
}
}
//}
return dst;
//return dsthsi;
}
算法效果
原图