Don't know if this will be of use, but I was casually associated several years ago with Dr. Richard E. Haskell, at Oakland U. in Michigan, who was developing feature extraction techniques. See his publications. (Google: 'richard haskell feature extraction')

AFAIR, his approach depended on analysing only two scan lines at a time for feature 'training' and recognition, so the need for computational resources was minimal, and even a simple black-and-white line-scan camera over a moving conveyor belt would be sufficient; recognition could operate in real time. The feature extractors produced were translationally and aspect ratio invariant, but not rotationally invariant; however, rotational invariance could be simulated by 'learning' an object in a whole bunch of different rotational presentations.

I guess the idea for your app would be to learn to recognize 'hard edges' and a 'proper' object, then anything with hard edges that is not coincident with a proper object is 'foreign'. Or something...