page 1  (19 pages)
2to next section

On the Primal Sketch and Feature-Free Profiles

Robyn Owens?

October 10, 1991

1 Introduction

It is common in both machine vision and human vision research to take as a premise the proposition that the early stages of image processing involve the formation of a primal sketch [Marr, 1982]. What this primal sketch is is variously interpreted and represented. Marr himself considered it to be a reduction to tokens (of some unspecified form) of some essential pieces of visual information that, at a later stage, are used by the brain to compute what is in the image. Marr hypothesized that there are tokens for edges, bars, line terminations and other 1- and 2- dimensional image features, including possibly blobs.

In fact, Marr's work derived from much earlier ideas proposed by Ernst Mach, who also suggested that the eye-brain system produces some type of internal sketch in its efforts to interpret the enormous amount of information that falls on the eye.

The notion of producing a sketch was, and still is, seen as a form of data reduction or data compression. It is important, however, that we understand the implicit constraints that this primal sketch proposition imposes. If the primal sketch is to form some type of data compression then we understand implicitly from that notion that the original image can be reconstructed from the compressed data. In other words, what is thrown away is redundant to image interpretation because there exist techniques for its reconstruction.

There does not seem to have been any criticism to this idea of reconstruction mentioned above. Indeed there has been active research into achieving just that [Brady and Horn, 1983; Blake and Zisserman, 1987; Bolle and Vemuri, 1991], even given the fact that there is no real concensus as to what the primal sketch

?This report was prepared whilst the author was on Study Leave from The University of Western Australia, visiting the Department of Engineering Sciences at Oxford University.