A New EdgeDetection Technique and ItsImplementation HARRYWECHSLERAND MASATSUGUKIDODE
Harry Wechsler · 1977
A new edgedetection technique fordigital image processing ispresented. Thetechnique isviewed asa operator inthecontext ofthemorecomplex problem, that ofscene analysis, anda conceptual comparison withsomeprevious edge detectors isdone. Thenewedge detection technique hasbeenim- plemented andits results arecompared with twoother edge detection techniques usingthesamekindofpictures asinput data. Thesoft- wareimplementation ofthis newedge detector written inmachine language takes 32s foran image ofsize128by128picture elements. Theedge detector canbehardware-implemented andsuch animplementation, forwhich theestimated processing timewill be about half asecond, isgiven intheAppendix. I. INTRODUCTION THE SEGMENTATIONproblem, i.e., thetask ofparti- tioning agiven sceneinto meaningful objects, isoneof thecurrent concernsinsceneanalysis, andmuchofthework being doneindigital image processing isaimed atsolving this problem. Thedifficulty ofidentifying the contours ofthe objects present inagiven scene,which iseasyforthehuman eye,isaverycomplex task toautomate, and, although many attempts havebeenmadetosolve it, anadequate solution is still lacking. Theconsensusofpeople working ontheabove problem is that semantic information, i.e., a priori information about possible interpretations andrelationships between the objects present ina given scene,isnecessaryifone isto achieve asuccessful segmentation. Themaindifficulty with a semantic approach isthat theamountofinformation tobe storedandaccessedissolargeastothreaten acombinator- ial explosion. Sucha combinatorial explosion can be avoided ifa roughapproximation can bemadeatthe beginning ofthe segmentation processandiftheapproxima- tion would notdepend on any apriori knowledge aboutthe scene.Theaboveprocedure iscalled a low-level vision algorithm. Thisapproach reduces theamountofcomputa- tion needed toidentify a knowledge-subset.