How about detecting vector edge shapes and unifying that result with the existing classifier? Surely a leopard sofa cannot have the same edge vector shape as a real big cat.
That's an idea from 1970-1980s AI, called the "primal sketch" model. The concept was to take an image and try to turn it into a line drawing, then extract the topology and geometry.[1] Further processing might yield a 3D model.
This sort of works on simple situations without too much edge noise. It's been used for industrial robot vision, where what matters are the outside edges of the part. It's not too useful when there's clutter, occlusion, or noisy textures.
More recent thinking is to find surfaces, rather than edges. This works well if you have a 3D imager, such as a Kinect. You can get a 3D model of the scene. Occlusion remains a problem, but texture noise doesn't hurt.
