We present the combination of a decision theoretic and a syntactic approach to image segmentation. It is shown how statistical properties of iconic information can be systematically used to program a special architecture for parallel decision theoretic image segmentation. It is also shown how the probabilistic output of this architecture automatically provides problem dependent primitives for a subsequent syntactic phase. This phase can resolve ambiguities and incomplete segmentation results in cases where objects and background are not clearly distinct by textural and gray level properties alone. Evidence for the performance of the suggested combined approach is provided by examples from different industrial and biomedical applications. © 1989 Springer-Verlag New York Inc.