Pattern classification: Fusion of information

Improvements in the performance of pattern recognition systems can be brought about by fusing as many sources of information as possible. These include multiple cues, multiple opinions of experts, prior contextual knowledge and measurement context. Tools for fusing such diverse sources of information are reviewed. In particular, we show that the use of multiple cues can be formulated as a problem of fusing the opinions of multiple experts employing distinct representations. This contrasts with the methods aimed at improving estimates of decision function values, which can be viewed as fusion of opinions of experts using identical representations. Spatial and temporal context can be incorporated by taking into account measurements on neighbouring objects and modelling the distributions over their relations. This is shown to involve two basic components: a measurement process model and a prior world knowledge model. Depending on formulation, contextual decision making leads either to compound classification or to estimation of Maximum Aposteriori Probability of joint labelling (1).