Decision fusion for writer discrimination

In this work the performance improvement of a one-word-based writer discrimination system is described. This is achieved when decision fusion is employed to combine the classification results (decisions) from N words. Only the binary classification problem (discriminate between two persons) is examined. In the proposed approach the randomized Neyman-Pearson (N-P) test is applied. A special case of feature statistics is considered in order to fit the randomized N-P test and simultaneously achieve minimum classification error for each separate decision. Using only a few words the performance of the discrimination system is radically improved as far as the maximum classification error is concerned