A Comparison between Top-Down and Bottom-Up Image Analysis in Terms of the Complexity of Searching a Problem Space

In the context of the analysis of remotely sensed data the question arises of how to analyze large volumes of data. In the specific case of agricultural fields in flat areas these fields can often be modeled in terms of geometric primitives such as triangles and rectangles. In this case the options are classical i.e. bottom-up, starting at the pixel level and resulting in a segmented, labeled image or topdown, starting with a model for image partitioning and resulting in a minimum cost estimation of shape hypotheses with corresponding parameters. We report on an investigation of the search effort needed for resolving a simplified segmentation problem of partitioning an image into two segments. Experimental factors are edge length and overlap of monospectral probability distributions of two classes. The method for quantifying the complexity of an approach is to determine the number of possible solutions at each stage in the process and the convergence rate towards a final solution of the segmentation and labeling problem.