00100	Neurophysiological studies suggest that cells in the visual cortex of
00105	mammals  extract  edges.     A  priori,  intensity  edges  occur   at
00110	discontinuities  in  surface  slope  or  reflectance.     Images  are
00115	patchworks of areas which are projections of obscuring objects.  Some
00120	edges  correspond  to  boundaries  between  objects, some to interior
00125	edges of objects, others to surface markings, others to  shadows  and
00130	reflections. All signify features of potential interest, but there is
00135	no  direct  connection  between  intensity  boundaries  and   spatial
00140	interpretations (which are our perceptual goal). Thus, it is possible
00145	to find intensity edges on a local basis, even though some edges with
00150	low  contrast  will  be  missed.   It  is  not  possible  to find the
00155	"meaningful edges" on a local basis,  ie  to  make  a  "perfect  line
00160	drawing".  Nor is it easy to find intensity edges; present techniques
00165	require extensive calculation and are not really  adequate.  We  feel
00170	that   computational   complexity   is   inherent.   We  expect  that
00175	performance approaching human visual acuity inherently requires great
00180	computation  cost,  and that substantial economies in computation are
00185	possible only by sampling strategies which ignore much of the  image,
00190	except  for  those  extended  features  which  show  up  on  a coarse
00195	sampling. For most  problems  of  interest,  however,  those  reduced
00200	abilities are presumably adequate.
     

00100	Edges are discontinuities along a smooth curve, ie locally straight.  
00200	Edges can be distinguished from spots only in a two dimensional context.
00300	Thus local edge operators must be two-dimensional (over a small area)
00400	and directional.  Mammalian edge operators, Hubel-Wiesel cells, are directional;
00500	there are cells for many directions (angular resolution ??) and
00600	corresponding to retinal positions (spatial resolution and guess as
00700	to how many).  Presumably they are more dense
00800	near the fovea.  However it is not clear that these cells are
00900	distributed proportional to retinal cells.  Teleologically, straight
01000	line approximations to edges are less useful over large visual angles.
01100	Physiological evidence shows many more retinal cells than ganglion
01200	cells connecting them to the brain.  Only a portion of the fovea appears
01300	to be connected one-one to the visual cortex. It is known that shape
01400	discrimination is crude in the periphery of the field of view.
01500	What is done with the outputs of cells which are not directly connected
01600	to visual levels of the brain?  Do ganglion cells multiplex spatial
01700	information in the time domain, or is there primitive feature extraction
01800	in the retina?
01900	
02000	Hueckel [] created a local edge operator for computers, which computes
02100	directional moments and solves for a best edge.  The signal is primarily
02200	determined by the gradient, and thus the operator has an unfortunate
02300	sensitivity to smooth shading.
02400	
02500	Local edge detection as described here is a low-level operation in that
02600	no information from higher levels is used.  The output of local edge
02700	operators does provide a context for other operations.  
02800