Java code for image scene analysis

Have a set of images with different straight-edged shapes lying in different orientations. What I need to end up with is the length and width of every rectangular shape. first image: [login to view URL] second image: [login to view URL] I was told it should be fairly easy to write in java code. These were the steps that were given to me: a) segment the image into “foreground”/“background”. This looks easy, with the caveat that you’ll want to remove the darker parts of the objects so that you get clean boundaries. b) use something like the Hough Transform, with some post-processing to identify all extended edge elements. Another possibility is to directly detect (probably with template matching) the “obvious” long, straight boundaries. Some of the edges on the non-quadrilateral objects might be a bit challenging, but that’s probably a matter of tweaking parameters. c) analyze the collection of extended edge-elements to match them with your model of what your objects look like. This model would include # of sides and size. With a little work, this would allow you to recognize most (perhaps not all) overlapping objects. It’s also possible to use the Generalized Hough Transform to directly model the shapes you are looking at, as long as you have only a small number and the scale does not vary too widely. I suspect that direct template matching would also work. Based on the images you posted, I would probably start by reducing the size of the images a bit - this might help with filling in the darker regions of the objects, and would certainly speed things up. If you can already successfully segment foreground/background, and your only problem is rotation and overlap, then you can jump directly to step c). Apply a Hough transform to find edges, a Generalized Hough Transform to find objects, or use template matching. For completeness (which means you might want to stop reading here…), another technique that would work for the seperable objects is to use a “rotating calipers” approach. The simplest version of this involves projecting the image function onto a series of lines and many orientations - extracting the function w(angle) - the width of the object as a function of projection angle. From w(angle) it is easy enough to extract orientation and size information for all of your objects - but it only works for non-overlapping objects. Hint to developers: if w(angle), given a single segmented, isolated particle does not already exist, I think it might be a good addition to the toolbox. If it’s already there - great! If not, consider adding it.