Pure orientation filtering: A scale-invariant image-processing tool for perception research and data compression

Some coding properties of the scale-invariant orientation variable are explored, paying particular attention to its perceptual significance in texture segmentation and compact image coding.

Abstract

A method is described for scale-invariant segregation of image structure solely on the basis of orientation content. This kind of image decomposition is an unexplored image-processing method that is complementary to the well-explored method of filtering in spatial frequency bands; the latter technique is rotation-invariant, whereas the former technique is scale-invariant. The complementarity of these two approaches is explicit in the fact that orientation and spatial frequency are orthogonal variables in the two-dimensional Fourier plane, and the filters employed in the one method depend only on the radial variable, whereas those employed in the other method depend only on the angular variable. The biological significance of multiscale (spatial frequency selective) image analysis has been well-recognized and often cited, yet orientation selectivity is a far more striking property of neural architecture in cortical visual areas. In the present paper, we begin to explore some coding properties of the scale-invariant orientation variable, paying particular attention to its perceptual significance in texture segmentation and compact image coding. Examples of orientation-coded pictures are presented with data compression to 0.3 bits per pixel.