Image Background Removal

For my gesture recognition uni project I have developed simple code that removes background from the image. For it to work there has to be an image of background scene and an image with an object that partially covers same background scene (a hand with a background of a room scene for example).

function [outIm] = makeMask(bg, im, tol)
%bg - background image
%im - input image
%tol - tolerance

[h,w] = size(bg);
outIm = false(h, w);
for ch = 1:h
    for cw = 1:w
        imPix = im(ch,cw);
        bgPix = bg(ch,cw);
        if ((bgPix == imPix) || (bgPix > imPix && bgPix <= imPix + tol) || (bgPix < imPix && bgPix >= imPix - tol))
            outIm(ch, cw) = 0;
        else
            outIm(ch, cw) = 1;
        end
    end
end
end

As You see, code checks every single pixel and decides if it belongs to background scene. All this works well, however I want to ask is there more efficient way to achieve the same?

Answer 1

[Egon Geerardyn]

You can try vectorizing that code, but that will be most likely destroy any short-circuiting you have in there.

Just by the top of my head, following code should be able to replace your for-loops:

brighterThanBg = (im > bg + tol);
darkerThanBg = (im < bg - tol);
outIm = (brighterThanBg | darkerThanBg);

That will also check every pixel but it will be faster as MATLAB is quite slow on loops. Also, by inverting your logic (you check for background, split into 3 regions, 2 of which require 2 conditions to be satisfied), while this codes only takes care of 2 regions (pixels too bright to be background, pixels too dark to be background), each only requiring a single condition.

[Walter Roberson]

Newer versions of Matlab have improved loop speed considerably, apparently.

[Mazvydas Tadaravicius]

function [outIm] = makeMask2(bg, im, tol)
outIm = ((im > bg + tol) | (im < bg - tol));
end

This function does the job so much faster, have a look at these figs:

>> tic; m1 = makeMask(bgg, ig, 20); toc; tic; m2 = makeMask2(bgg, ig, 20); toc;
Elapsed time is 0.004686 seconds.
Elapsed time is 0.001663 seconds.

>> tic; m1 = makeMask(bgg, ig, 20); toc; tic; m2 = makeMask2(bgg, ig, 20); toc;
Elapsed time is 0.004764 seconds.
Elapsed time is 0.001655 seconds.

>> tic; m1 = makeMask(bgg, ig, 20); toc; tic; m2 = makeMask2(bgg, ig, 20); toc;
Elapsed time is 0.005040 seconds.
Elapsed time is 0.001660 seconds.

Thanks for the tip Egon Geerardyn.

Answer 2

[Siddharth Shankar]

Have you considered using the imabsdiff method. The difference of the two images is basically the object obstructing the background.

[Mazvydas Tadaravicius]

I could use imabsdiff method, however it does not tkae into consideration the change of lighting when object covers background. Also, it is way slower than my method or method suggested by Egon.

>> tic; imabsdiff(bgg, ig); im2bw(ans, graythresh(ans)); toc;

Elapsed time is 0.115902 seconds.

>> tic; imabsdiff(bgg, ig); im2bw(ans, graythresh(ans)); toc;

Elapsed time is 0.030604 seconds.