Multilevel models for pattern recognition tasks with multiple classes

The problem of choosing binary subtasks for recognition tasks with multiple classes is considered from the points of view of the algebraic and logical approaches to recognition. The limits of their applicability were studied theoretically. The sufficient condition of correctness of algorithms is stated as a result of research of dependency between the first and the second level algorithms. Additionally, the paper proves that the object resolution method is applicable to constructing new objects using the precedent information. As an applied result, two modifications of the ECOC (error-correcting output codes) method are proposed. The first one is based on optimization of the binary subtasks set. The second one develops ideas of the fuzzy object resolution method with classes described by multisets of codes of their precedents. The proposed modifications make it possible to increase the initial method's quality in various situations, which is demonstrated by the example of model and real-world tasks.