Abstract
Methods are developed that permit the decomposition of two-dimensional convolutional kernels defined on a large area of support into sets of 1 × 3 and 3 × 1 convolutional primitives that are the natural tools of high-speed parallel processors having limited neighborhood data access. Five primitives are shown to be sufficient for 4-neighbor architectures with an additional four primitives for 8-neighbor architectures. Examples of the decomposition process are described for several low-level image analysis kernels, and it is shown that spectral analysis and edge detection can be performed over a wide range of scale. This leads to the conjecture that the convolutional primitives introduced here may be the parallel to the barlike receptors often cited as a model of human visual processing in that both may permit complex processes, such as spectral analysis, to be affected by the massive concatenation of simple primitives.
© 1988 Optical Society of America
Full Article | PDF ArticleMore Like This
John G. Daugman
J. Opt. Soc. Am. A 5(7) 1142-1148 (1988)
A. Enis Çetin and Rashid Ansari
J. Opt. Soc. Am. A 5(8) 1193-1200 (1988)
Göte Nyman, Mark W. Greenlee, and Pentti Laurinen
J. Opt. Soc. Am. A 5(4) 628-635 (1988)