You can use the following technique if the number of images in your sequence is odd.
- Prepare an N (which is odd) channel image from the sequence of images that you have
- Reshape this image into a 1-channel column vector
- Now apply a median filter of size N to this column vector. As the image is a 1-channel column vector, the filter will calculate the median of the channels(of course, there will be some additional calculations that are not useful to us)
- Reshape this filtered image into its original form having N channels and the original number of rows and columns
- Pick the middle channel of this N-channel image, which contains the median image of the sequence
Below I've illustrated the above items with a simple code and its output.
individual channels
channel 0:
[1, 1, 1;
1, 1, 1;
1, 1, 1]
channel 1:
[2, 2, 2;
2, 2, 2;
2, 2, 2]
channel 2:
[3, 3, 3;
3, 3, 3;
3, 3, 3]
channel 3:
[4, 4, 4;
4, 4, 4;
4, 4, 4]
channel 4:
[5, 5, 5;
5, 5, 5;
5, 5, 5]
output for N = 3
3-channel image data:
[1, 2, 3, 1, 2, 3, 1, 2, 3;
1, 2, 3, 1, 2, 3, 1, 2, 3;
1, 2, 3, 1, 2, 3, 1, 2, 3]
1-channel column vector image data:
[1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3]
median of the 1-channel column vector image data:
[1; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 3]
reshaped filtered image data:
[1, 2, 2, 2, 2, 2, 2, 2, 2;
2, 2, 2, 2, 2, 2, 2, 2, 2;
2, 2, 2, 2, 2, 2, 2, 2, 3]
median image data:
[2, 2, 2;
2, 2, 2;
2, 2, 2]
output for N = 5
5-channel image data:
[1, 2, 3, 4, 5, 1, 2, 3, 4, 5, 1, 2, 3, 4, 5;
1, 2, 3, 4, 5, 1, 2, 3, 4, 5, 1, 2, 3, 4, 5;
1, 2, 3, 4, 5, 1, 2, 3, 4, 5, 1, 2, 3, 4, 5]
1-channel column vector image data:
[1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2
; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5]
median of the 1-channel column vector image data:
[1; 2; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3
; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 4; 5]
reshaped filtered image data:
[1, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3;
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3;
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 5]
median image data:
[3, 3, 3;
3, 3, 3;
3, 3, 3]
Code:
// number of channels (= number of images in the sequence)
// N MUST BE ODD
const int N = 5;
// channel data
uchar ch0[] = {1, 1, 1, 1, 1, 1, 1, 1, 1};
uchar ch1[] = {2, 2, 2, 2, 2, 2, 2, 2, 2};
uchar ch2[] = {3, 3, 3, 3, 3, 3, 3, 3, 3};
uchar ch3[] = {4, 4, 4, 4, 4, 4, 4, 4, 4};
uchar ch4[] = {5, 5, 5, 5, 5, 5, 5, 5, 5};
// images
Mat m0 = Mat(3, 3, CV_8U, ch0);
Mat m1 = Mat(3, 3, CV_8U, ch1);
Mat m2 = Mat(3, 3, CV_8U, ch2);
Mat m3 = Mat(3, 3, CV_8U, ch3);
Mat m4 = Mat(3, 3, CV_8U, ch4);
// prepare image sequence
Mat channels[] = {m0, m1, m2, m3, m4};
// put the images into channels of matrix m
Mat m;
merge(channels, N, m);
// reshape data so that we have a single channel column vector as the image
Mat n = m.reshape(1, m.rows * m.cols * m.channels());
// apply median filter to the 1-channel column vector image. filter size must be the number of channels
Mat med;
medianBlur(n, med, N);
cout << N << "-channel image data:" << endl;
cout << m << endl;
cout << "1-channel column vector image data:" << endl;
cout << n << endl;
cout << "median of the 1-channel column vector image data:" << endl;
cout << med << endl;
// reshape the filtered 1-channel column vector image into its original form having N channels
med = med.reshape(N, m.rows);
cout << "reshaped filtered image data:" << endl;
cout << med << endl;
// split the image
split(med, channels);
// extract the middle channel which contains the median image of the sequence
cout << "median image data:" << endl;
cout << channels[(N+1)/2 - 1] << endl;