Neurological Foundation of Image Processing

A popular computation approach is to process visual images by dividing them into crisp (winner-takes-all) parts in analog to properties of neurophysiological receptive fields. Problem with such symbolic representation is that in a real environment object attributes are seldom invariant. We propose to divide images into rough parts using hierarchical, multi-valued processes.  The bottom-up computation (BUC) is related to prediction where object attributes are approximated by different granules with properties similar to different brain areas: by dots as in the thalamus, by oriented lines as in the primary visual cortex, and by elementary shapes as in V4. There are a large number of possible combinations of elementary granules; therefore objects in BUC are overrepresented. The top-down computation (TDC) fits prediction to hypothesis posed by more complex properties (higher brain areas). If the hypothesis check is positive, TDC verifies the object and eliminates other possible patterns. Such classifications take place in parallel at many functional units. We show an example of such hierarchical system computation on experimentally recorded data from monkey visual area (V4).