Schemes of Combining Discriminant Functions to Improve the Classification Accuracy for Ensemble of Data Sources

Data classification accuracy is studied in terms of a relation between the error probability and the processed amount of information for different fusion schemes. The fusion schemes for weak discriminant functions are considered on an equimodal dataset and on an ensemble of data from multimodal sources. For the proposed fusion schemes, the error probability redundancy is estimated with respect to the information-theoretic lower bound in the form of a modified rate distortion function with the Hamming distortion metric. The experimental estimates obtained on the datasets of face and signature images demonstrate a decrease in the error probability and its redundancy with respect to the lower bound by increasing the processed amount of information due to the fusion of weak discriminant functions.