Census Transform – C++ Implementation

Census Transform is a non-parametric local transform proposed by Zabih and Woodfill[1] in 1994.  Robust and insensitive to various brightness and contrast images is the primary advantage of the transform. The census transform is rather simple

12 58 66          1 1 1

50 70 81   =>  1    0  => 111 1 0 101b  => 245d

35 92 11          1 0 1

For example, we define a 3×3 sliding window to scan the entire image, and compare every pixels with center pixel.

If(PixelIntensity < PixelIntensityCenter)

bit =  1;

else

bit = 0;

then shift the bit to the left.

Do the same job to complete the all elements in the window and move the window to next pixel.

As we can see, the transform only reflect the relative brightness of a pixel to its neighbors. So it is an ideal way to reveal the structure of an image.

imL_original

imL_highcontrast

imL_lowcontrast

////////////////////////////////////////////////////////////////

img33_original

img33_highcontrast

img33_lowcontrast

Pay attention to the low-texture area, the census transform is still powerful to extract the structure inside such areas, which would facilitate stereo matching algorithms to find similarities between image pairs even in low-texture areas.

=======================

I continuously get wrong calculation results from my code at one time, so be careful.

Mat imgIn = imread(imgName, 0);
Mat imgTemp = imgIn;//try to copy the mat, but it is wrong

 #include <iostream>
 #include <sstream>

 #include "opencv2/core/core.hpp"
 #include "opencv2/core/opengl_interop.hpp"
 #include "opencv2/highgui/highgui.hpp"
 #include "opencv2/imgproc/imgproc.hpp"
 #include "opencv2/calib3d/calib3d.hpp"
 #include "opencv2/contrib/contrib.hpp"

 using namespace std;
 using namespace cv;

int main(int argc, char** argv)
 {
 const char* imgName = "imL_re.png";
 Mat imgIn = imread(imgName, 0);
 imshow("input", imgIn);
 Size imgSize = imgIn.size();
 Mat imgTemp = Mat::zeros(imgSize, CV_8U);

 //If (CurrentPixelIntensity<CentrePixelIntensity) boolean bit=0
 //else boolean bit=1
 //450jx375i

 unsigned int census = 0;
 unsigned int bit = 0;
 int m = 3;
 int n = 3;//window size
 int i,j,x,y;
 int shiftCount = 0;
 for (x = m/2; x < imgSize.height - m/2; x++)
 {
   for(y = n/2; y < imgSize.width - n/2; y++)
   {
     census = 0;
     shiftCount = 0;
     for (i = x - m/2; i <= x + m/2; i++)
     {
       for (j = y - n/2; j <= y + n/2; j++)
       {

         if( shiftCount != m*n/2 )//skip the center pixel
         {
         census <<= 1;
         if( imgIn.at<uchar>(i,j) < imgIn.at<uchar>(x,y) )//compare pixel values in the neighborhood
         bit = 1;
         else
         bit = 0;
         census = census + bit;
         //cout<<census<<" ";*/

         }
         shiftCount ++;
       }
     }
    //cout<<endl;

    imgTemp.ptr<uchar>(x)[y] = census;
   }
 }

 imshow("output", imgTemp);
 waitKey();
 return 0;

}

===================
[1]http://www.cs.duke.edu/courses/spring06/cps296.1/handouts/Zabih%20Woodfill%201994.pdf