Representation of linear structures using perceptual grouping

Abstract

The robust detection of curvilinear structures from con tours represen ts an importan t cue for most computer vision systems Ho ev w er the analysis of D images of a scene is a dicult task made of under constrained ill posed problems P erceptual Organization is one con tribution of psyc ho visual theories to computer vision It pro vides generic approac hes for the perception of scenes necessary to reduce constrain ts and solv eam biguities The purpose of this paper is to propose an ecien t method based on Gestalt rules to de tect salien tcon tours and extract elemen ts of represen tation from complexe scenes The proposed grouping strategy is hierarc hical and divided in three lev els of organization The rst lev el is based on the saliency net ork framew w ork A qualit y function based on curv a ture cocircularit  con y tin uit y and orien tation is optimized throughout a net orkoflocally con w nected elemen ts The purpose of this lev el is to reduce the complexit yofvisualtasks b y selecting the most salien t linear structures from con tours W e propose a generic formalism for the conception of suc hnet w orks a more stable family of qualit y functions a new algorithm for optimization and criteria for selection of a set of salien t groups after optimization Salien t groups of edge elemen ts pla y a role of focus of atten tion for the second lev el of organization Hypotheses of segmen ts arcs and poin ts of in terest are proposed from eac h salien t group and grouped follo wing rules of parallelism pro ximit  con y tin uit yto create a reduced set of represen tativ e elemen ts of con tours Finaly  w e illustrate the nal lev el of organization with an application to junction detection and matc hing on digital and real scenes The main caracteristic of our approac h is the separation bet een a general strategy of organiza w tion and grouping modules specialized for a dened task Salien t elemen ts are dened b y generic properties A certain amoun tofam biguities and redundancies is necessary to allo w the detection of m ultiscale structures This w ork insists on the manipulation of complexe scenes on usual sys tems It has been applied to v arious t ype of scenes from satellite and medical imaging to indoor and outdoor scenes The qualit y of results conrm the robustness of this approac hin cluttered en vironmen ts Keyw