A two-phase connectionist approach to invariant picture interpretation

The paper presents an efficient two-phase approach to picture interpretation based on original connectionist techniques. During the first phase invariant representations of individual objects are obtained based on third-order image correlations and appropriate neural network classifiers are used to provide a probabilistic assignment of labels to objects. The second phase uses relationships between objects to reduce or eliminate ambiguity by means of a relaxation scheme based on stochastic learning automata. Both phases are particularly suited to parallel implementation. Simulation experiments revealed the effectiveness of our approach in solving several problems of small and medium sizes.