Changes

Color3d v; Color3d mt = Color3d(0.0, 0.0, 0.0); Color3d st = Color3d(0.0, 0.0, 0.0); for(int y=0; y<target.rows; y++) { for(int x=0; x<target.cols; x++) { v = target.at<Color3d>(y, x); v = mRGB2LMS * v; for(int c=0; c<3; c++) v(c) = v(c) > eps ? log10(v(c)) : log10(eps);  target.at<Color3d>(y, x) = mLMS2lab * v; mt = mt + target.at<Color3d>(y, x); st = st + target.at<Color3d>(y, x) * target.at<Color3d>(y, x); } }  Color3d mr = Color3d(0.0, 0.0, 0.0); Color3d sr = Color3d(0.0, 0.0, 0.0); for(int y=0; y<refer.rows; y++) { for(int x=0; x<refer.cols; x++) { v = refer.at<Color3d>(y, x); v = mRGB2LMS * v; for(int c=0; c<3; c++) v(c) = v(c) > eps ? log10(v(c)) : log10(eps);  refer.at<Color3d>(y, x) = mLMS2lab * v; mr = mr + refer.at<Color3d>(y, x); sr = sr + refer.at<Color3d>(y, x) * refer.at<Color3d>(y, x); } }  for(int y=0; y<target.rows; y++) { for(int x=0; x<target.cols; x++) { for(int c=0; c<3; c++) { double val = target.at<double>(y, x*3+c); target.at<double>(y, x*3+c) = (val - mt(c)) / st(c) * sr(c) + mr(c); } } }  // Transform back from lab to RGB for(int y=0; y<target.rows; y++) { for(int x=0; x<target.cols; x++) { v = target.at<Color3d>(y, x); v = mlab2LMS * v; for(int c=0; c<3; c++) v(c) = v(c) > -5.0 ? pow(10.0, v(c)) : eps;  target.at<Color3d>(y, x) = mLMS2RGB * v; } }  As seen in the call graph of the profile data 98% of the time spent in the program is spent processing in the main file. Within this main multiple calls to the Color3d overloaded operator is made (located in the Color3d file). This makes the program ideal for parallelization as many of the steps used to process the pixel colours can be delegated to the Nvidia GPU. This can be seen in the above code snippets taken from (found in the main, where link above) when calls to Color3d operator are done multiple times in a triple for loop.