THE FUSION OF DECISIONS FOR DISTRIBUTED RECOGNITION

摘要

Decision fusion algorithms can be broadly categorised as making use of class labels only (hard decisions) or class labels together with associated confidence estimates (soft decisions). The fusion of hard decisions has the advantage that it is relatively straightforward to implement, requires minimal communication bandwidth and in many cases provides perfectly adequate classification performance. The fusion of soft decisions generally yields higher classification performance at the expense of increased communication bandwidth. This chapter has detailed three basic algorithms for fusing hard decisions including majority voting, weighted voting and maximum a posteriori fusion. The ability to defer a fused decision was also covered. Methods for soft decision fusion either for classifier combining using weighted averaging or multi-sensor fusion using the product rule were examined. The product rule was generalised to allow for sensors that are not conditionally independent. Finally the veto effect and a simple means of correcting for it when using the product rule was described.