rao@enuxha.eas.asu.edu (Arun Rao) (05/03/89)
I'm currently doing a performance evaluation of various methods of geometric object recognition (neural and non-neural). The final objective is to compare the performance of the best existing systems with a new approach we are developing. The pattern set to be recognized consists of simple 2-dimensional contours. All objects are completely visible (no occlusion). The ASCII character set is one example, but the system should be capable of handling arbitrary shapes. It should be insensitive to translation, scaling and rotation. I know of the following methods: (i) Hough transform methods (Dana Ballard). (ii) Fourier/log-polar transform methods (Casasent and Psaltis, Brousil and Smith, many others). (iii) Neocognitron (Fukushima). (iv) Invariance net (Widrow). (v) Hierarchical structure coding (Hartmann). (vi) Similitude invariant system (Prazdny). I am specifically interested in how this problem is dealt with in existing vision systems - most of the above are probably still confined to the laboratory. How good, for example, are OCR systems ? As I understand it, they still require a very strictly defined character set, and do not tolerate misalignment. Any comments would be appreciated. Thanks in advance. - Arun -- Arun Rao ARPANET: rao@enuxha.asu.edu BITNET: agaxr@asuacvax 950 S. Terrace Road, #B324, Tempe, AZ 85281 Phone: (602) 968-1852 (Home) (602) 965-2657 (Office)