For more complicated pattern recognition systems several different solutions may be found, each solving the problem differently and thereby making other types of errors. Instead of selecting the best solution, they may be combined. There are various ways to do this. It is important that each of the base systems produces some type of confidence for each decision to be made. Combining may then be based on:
- fixed combining rules like taking the decision with the maximum confidence, or computing the product over the confidences given by the various systems for a particular decision.
- trained rules that train the combining system separately.
In our research we focus on trained combiners and aim to establish the properties of various combiners and to design a strategy for training the overall system: the set of base pattern classifiers and the combiner. Newer research directions in which classifier combining are multiscale image analysis and processing, multiple instance learning, and computer aided detection and diagnosis.