Alpha channel in OpenCV

Question

does OpenCV support alpha-channel? Or is there any way to work with transparent png? I need to merge two images, where the first one is background and the second one is imag

Answer 1

Here is a bash script that I threw together that will perform the ImageMagick conversion given by ypnos on all of the png files in a directory. You can make it recursive by replacing the * in the third line with a find command.

#!/bin/bash

for file in *
do
    if [[ $file =~ (.+)-mask\.png ]]; then
        echo "Ignoring mask $file"
    elif [[ $file =~ (.+)\.png ]]; then
        echo "Generating mask for $file"
        basefn=${BASH_REMATCH[1]}
        convert "$basefn.png" -channel Alpha -negate -separate "$basefn-mask.png"
    fi
done

Answer 2

Updated answer: Use CV_LOAD_IMAGE_UNCHANGED flag to load all four channels (including Alpha) from the image. Then, use mixChannels() and/or split() to separate the alpha channel from others, and threshold on it, as explained below.

Very old answer:

OpenCV does not support alpha channel, only masking. If you want to read in PNG with alpha, use imagemagick first to extract alpha channel:

convert input.png -channel Alpha -negate -separate input-mask.png

Then in OpenCV you can do sth like this:

Mat_<float> mask = imread("input-mask.png", 0);
threshold(mask, mask, 254., 1., THRESH_BINARY);

... to get a real mask (usable as mask matrix in OpenCV operations). Or you can use it in your own operations without the thresholding. To apply the mask it may also be a good idea to extend it to three channels:

std::vector<Mat> marr(3, mask);
Mat_<Vec<float, 3> > maskRGB;
merge(marr, maskRGB);

After that you can test it like this:

imshow("Target", target);
imshow("Mask", mask*255);
imshow("Applied", target.mul(maskRGB));
waitKey();

Note: This is OpenCV 2.0 code.

Answer 3

Input: cam_frame-background, media_frame-foreground,media_frame_alpha - alpha. Output: mixed by alpha transparent. If media_frame_alpha is IPL_DEPTH_8U use "cv::merge", because matrices must have equal dimension, size, channels. Work fast if opencv build with sse or avx. And more faster if use cv:GpuMat, must have nvidia cuda support and build opencv with cuda.

inline cv::Mat frame_mix_alpha(cv::Mat &cam_frame,
 cv::Mat &media_frame,
 cv::Mat &media_frame_alpha)
{
    Mat suba = media_frame_alpha;
    //Mat alpha = cvCreateImage(media_frame_alpha.size(), IPL_DEPTH_8U, 1);
    //cvtColor(media_frame_alpha, alpha, CV_BGR2GRAY);

    //Mat suba;
    //Mat rgba[3] = { alpha, alpha, alpha };
    //merge(rgba, 3, suba);

    Mat subb = ~suba;

    //Mat mincam = min(cam_frame, suba);
    //Mat minmedia = min(media_frame, subb);
    //Mat result = (cam_frame - mincam) + (media_frame - minmedia);
    Mat result = (cam_frame - suba) + (media_frame - subb);

    return result;
}