Abstract

Pattern recognition features derived from a pattern's contour can be made invariant^1,2 to the pattern's scale, position and rotation. Such features, (for example Fourier Descriptors), usually well characterize a pattern with a reduced feature set (in comparison to the feature set generated directly from two-dimensional gray-level patterns) and avoid, through smart segmentation, the problems of pattern shading. Contour features are significantly more sensitive to small changes in the pattern shape than the two-dimensional gray-level features. However, contour feature calculation is easily disrupted by contour discontinuities. Though contour features have been recently described in a hybrid optical-digital image processing system³, we describe an architecture which optically generates in parallel different classes of invariant contour features and is insensitive to small contour discontinuities. The normalization which generates invariant properties of these features is discussed.

© 1985 Optical Society of America