An invariant pattern representation based on nonuniform sampling in the human visual system

Matching a target pattern to a prototype can be difficult when structural variations are large. Since pattern size and scale are not normally identifying characteristics, eliminating size and scale differences can improve pattern recognition. Rotational differences can be distinguishing however, and cannot be eliminated. This paper presents a pattern representation that is invariant to shift and scale but is otherwise unique. Rotational differences between target and prototype can be determine by cross-correlation. The representation is based on a model of nonuniform sampling in the human retina. The model features a texture-to-grayscale transformation, facilitating the representation of textured patterns