A Biological-Plausable Architecture for Shape Recognition
Final rept. 1 Aug 2005-31 May 2006
NORTH CAROLINA STATE UNIV AT RALEIGH
Pagination or Media Count:
In this project, we explore a new approach to two-dimensional shape recognition. The method draws from literature on the Hough transform and its extensions. The methods is shown to be invariant to zoom, translation, rotation, and partial occlusion, although not zoom and partial occlusion simultaneously. The method is shown to be robust to distortions which smooth the contour shape scale space changes. Furthermore, when the method misclassifies a shape, it chooses a shape which is most similar in a human-intuitive sense to the original. The method is developed and evaluated on a data base of tank silhouettes and a data base of fish silhouettes. The computer-based version of the algorithm is shown to have a reasonable implementation in neural hardware, and a neural-network implementation is described.