Bayesian Network for algorithm selection: Real-world hierarchy for nodes reduction

In order to obtain the best result in image understanding it is desirable to select the best algorithm on a case by case basis. An algorithm can be selected using only image features, however such selected algorithms will often generate errors due to occlusion, shadows and other environmental conditions. To avoid such errors, it is necessary to understand processing errors on a symbolic level. Using symbolic information to determine the best algorithm is however difficult task because the possible combinations of elements and environmental conditions are almost infinite. Consequently it is impossible to predict best algorithm for all possible combinations of objects, environment conditions and context variations. In this paper we investigate selection of algorithms using symbolic image description and the determination of algorithms´ error from high level image description. The proposed method transforms and minimize the high level information contained in the symbolic image description in such manner that will preserve the algorithm selection quality. The transformation takes a high level information label and transforms it into a set of generic features while the minimization uses hierarchy to reduce the specific nature of the information. Both methods of information reduction are used in a Bayesian Network because a BN is well known for using the generalization and hierarchy. As is shown in this paper, such representation efficiently reduces the fine grain high-level symbolic description to a coarser-grain hierarchy that preserves the selection quality but reduces the number of nodes.