Jun Fang Jun Fang - 8 days ago 6
C++ Question

How to efficiently do complex threshold in RGB image using opencv?

I want to translate a matlab code into c++ opencv, it threshold a RGB image: red value>threshold1 AND red/green>threshold2 AND red/blue>threshold3

The matlab code is:

bw=(im(:,:,1)>=red_th&(im(:,:,1)./im(:,:,2))>=red_green_th&(im(:,:,1)./im(:,:,3))>=red_blue_th);


Where
im(:,:,1), im(:,:,2)
and
im(:,:,3)
is the r, g, b value respectively.

I found the matlab code is very efficient comparing with looping all pixels using "for cols and for rows". Hence, I want to find similar efficient method in opencv instead of looping cols and rows.

I read some information about
cv::threshold and inRange
, however, it seems that they cannot satisfy my requirement.

Answer

You can't do this directly with threshold or inRange, but you can easily convert this to OpenCV first splitting the 3 channels, and then using Matrix Expressions:

Mat im = ...
vector<Mat> planes;
split(im, planes); // B, G, R planes

Mat bw = (planes[2] >= red_th) & 
         (planes[2] / planes[1] >= red_green_th) &
         (planes[2] / planes[0] >= red_blue_th);

Since Matlab usually works on doubles, you'd better convert the OpenCV matrices to double (unless they are already so):

Mat im = ...
vector<Mat> planes;
split(im, planes); // B, G, R planes

for(size_t i=0; i<planes.size(); ++i) {
    planes[i].convertTo(planes[i], CV_64F);
}

Mat bw = (planes[2] >= red_th) & 
         (planes[2] / planes[1] >= red_green_th) &
         (planes[2] / planes[0] >= red_blue_th);

Or you can for loops, which can be very fast if you work on pointers (I'm assuming your im is of type CV_8UC3):

Mat3b im = ...
Mat1b bw(im.rows, im.cols, uchar(0));

int rows = im.rows;
int cols = im.cols;
if(im.isContinuous()) {
    cols = rows * cols;
    rows = 1;
}

for(int r=0; r<rows; ++r) {
    Vec3b* ptr_im = im.ptr<Vec3b>(r);
    uchar* ptr_bw = bw.ptr<uchar>(r)
    for(int c=0; c<cols; ++c) { 
        const Vec3b& bgr = ptr_im[c];

        // Take care of division by 0  

        ptr_bw[c] = (bgr[2] >= red_th) &&
                    (bgr[2] / bgr[1] >= red_green_th) &&
                    (bgr[2] / bgr[0] >= red_blue_th);
    }
}