Plato's Cave: The Conventional View

The Conventional View of Visual Processing

The current view of visual processing is based on a number of prominent milestones in vision science. One was the Hodgkin-Huxley neuron model, which suggests a simple integrate-and-threshold function for the individual neuron. Another was the finding by Huebel & Wiesel of simple, complex, and hyper-complex cells in the visual cortex, suggesting a hierarchy of successive stages of processing to extract ever higher order visual features. This concept is exemplified by recent work like that of Tanaka (1994) to map ever higher level features in higher cortical areas.

Conventional Models of Visual Processing

The conventional view has inspired a number of computational models both in image processing, as in the Blocks World model, and Ballard's Computer Vision, as well as models of human perception such as Marr's Vision model, and Biederman's Geon model of visual coding.

The basic idea in all of these systems is that visual perception involves a feed-forward progression of sensory information through various stages of feature detection. At each stage the information is abstracted and compressed, resulting in an ever more symbolic representation. Abstraction and compression generally go hand-in-hand, because abstraction implies a many-to-one transformation from many possible instances to a single invariant form, so that a system which is capable of such abstraction need only remember the canonical form, not the multiple possible instances of that form. Abstraction, compression, and invariance therefore are all desirable properties of a recognition system.

Return to argument

Return to Steve Lehar