00100	A Edges and Regions
00200	B Texture Edges
00300	C Depth Perception
00350	D Selective Attention
00375		High resolutio fovea-zoom
00380			eye tracking
00385			little used in AI- common idea
00390		hypothesis-verification (Shirai)
00400	
00500	A Edges and regions
00600		non-equiv of intensity features to semantic features
00700		inadequacy of current techniques
00800	
00900		A.1	local edge operator
01000		A.2 local edge organization
01100		A.3 more global edge organization
01105	
01110	A.1 local edge operator
01115		design criteria-optimality
01120		duality region-boundary
01125		
01130		sensitivity vs (cost,resolution,utility)
01135		insensitive to smooth gradients-later inhibition,Retinex
01140		2D not 1D-indistinguishable from noise except by extension
01145		range of sizes, for texture,pt features
01150		comp cost-optimality
01155		sensitivity-optimality
01160	
01165		comp cost in terms of N pt support N=nxn
01170		lower bound in cost is k*N arith operations, where k≥2 to
01175			get several moments to find directionality
01180		example k=2, N=8x8 implies 128 arith operations per application
01185		include the effect of applying every bth pixel
01190		allows directly nxn parellelism
01195		
01200		sensitivity
01205		contrast required for 50% false negatives at an edge
01210		for 5% false positives at no edge
01215		signal=(N/2)*h
01220		noise=sqrt(N)*sigma
01225		cutoff threshold: signal = 2* noise
01230		(N/2)*h=2*sqrt(N)*sigma
01235		contrast h = 4*sigma/sqrt(N)
01240		account for a variety of directions by including a factor of sqrt(2)
01245			for vertical and horizontal directions.
01250		Thus the optimal is approximately h = 4*sqrt(2)*sigma/sqrt(N)
01255			or for a 6x6 support, h ≥ .9*sigma
01260		The Hueckel operator gives h = 1.4*sigma
01265		Thus, the Hueckel operator is not terribly far from optimal,
01270		and not much is to be gained by improvements in sensitivity.