Euclidian, FDA and SVM weak classifiers fusion using a posteriori confidence classification (APCC)

The 2-class and multiclass classification systems have important issues when there is overlapping between the samples, insufficient representation of the classes or asymmetrical data representation. Sophisticated classification systems such as SVM and SVM-RBF may have generalization problems, so it is complicated to obtain successful classifiers. In this work it is shown how the use of classification fusion of simpler classifiers may improve the overall classification by using APCC (A Posteriori Confidence Classification). APCC defines the individual reliability of each parameter and each classification system per parameter, and produces a posteriori weight to each classifier according to its output. The developed protocols were tested using simulated data and real data from TPOEM (Temporal Patterns of Oriented Edge Magnitudes) and VPOEM (Volumetric Patterns of Oriented Edge Magnitudes) for facial expression representation. In both cases the use of APCC and classifier fusion allowed to improve the classification accuracy.