Scene understanding from SAR images. An overview

The major task of the scene understanding process is to find the scene which best explains the observed data. In general the observed data, the same radar signals, can be generated by different scenes. To identify the scene the author has to choose between competing hypotheses. The method analyzed within the present paper arises from the fundamental approach of considering the probability theory as a set of normative rules for conducting inference. The scene inversion is a model based approach, and the models carry the thematic information. Model comparison is a delicate task, more complex models can always fit better the data, so the maximum likelihood choice would lead to implausible over-parameterized models that generalize poorly. The author proposes as solution the Bayesian inference that penalizes the unnecessary complicated models and prefers the simpler and precise ones. The paper presents comparatively the problem statement for scene understanding in the light of the Bayesian inference, maximum entropy principle and the methods of simulated annealing