Analysis of error-reject trade-off in linearly combined classifiers