I would like to find rectangle-shaped objects in images (of size which is approximately known in advance). This is what a typical image looks like:
I am looking for an algorithm which is to some extent non-local (such as MSER) and will tend to output convex or mostly-convex blobs. Local algorithms (like flood-fill, simple intensity-based clustering and so on) tend to break on the "noise" in-between. I would also like to make use of the a-priori knowledge (object orientation will always be plus-minus horizontal, size range is also given). The image is a height map, I am looking for color-based segmentation.
Since I have numerics background, I was thinking about segmenting the image along the lines where the image would "crack" if loaded in tension vertically; that intuitively corresponds to my idea of separation.
I was considering grabcut, but I would have to set seed points, plus (AFAIK) it selects one obejct only, not several at once.
Thanks for pointers.