You can smooth the image in the vertical direction, which reduces the amplitude of the circles but leaves the stripes. Then subtract this from the original image, to leave just the circles and the borders at the top and bottom. You can't achieve perfect separation, but this gets quite a good result.
You can do the smoothing with fspecial and convn (or conv2 if you first convert the image to grayscale). Alternatively, you can use gsmoothn from here, as in the code below:
im = imread('NO AIRGAP2.bmp');
imd = double(im);
sigma = 10; % experiment with this parameter to get best result
imcircles = imd - gsmoothn(imd, [sigma 0 0], 'Region', 'same');
% display the result, converting to grayscale to make display clearer
imshow(mean(imcircles, 3), []);
I suspect there may be a very good solution possible using morphological operations instead of linear ones, but I haven't explored that.
