Selecting optimal strategy for combining per-frame character recognition results in video stream

This paper considers a problem of combining classification results from several observations of the same object. The task is seen as a case of collective decision making by a group of experts with estimated competence levels. Precision of different classification result combination methods is analyzed with different input data model, having per-frame character recognition results combination problem in video stream as an example. Experiments show that the strategy which selects a single most competent expert performs better with input data model without any non-relevant observations (in the context of character recognition in video stream — without characters location and segmentation errors). At the same time experiments show that strategies which combine several most competent experts using product rule or voting procedure outperform single-expect strategy with input data containing non-relevant observations.