ITERATIVE CURVE ORGANIZATION WITH THE EM ALGORITHM

This paper describes how the early visual process of contour organisat ion can be realised using the EM algorithm of Dempster et al. (1977). The underlying computational representation is based on Zucker et al.' s (1988) idea of fine spline coverings. According to our EM approach t he adjustment of spline parameters draws on an iterative weighted leas t-squares fitting process. The expectation step of our EM procedure co mputes the likelihood of the data using a mixture model defined over t he set of spline coverings. These splines are limited in their spatial extent using Gaussian windowing functions. The maximisation of the li kelihood leads to a set of linear equations in the spline parameters w hich solve the weighted least squares problem. We evaluate the techniq ue on the localisation of road structures in aerial infra-red images. We also provide some comparison with the dictionary-based relaxation s cheme of Hancock and Kittler (1990). in a comparative sensitivity anal ysis, the EM spline organisation process is demonstrated to outperform the relaxation operator in terms of noise control. (C) 1997 Elsevier Science B.V.