Abstract
This paper investigates writer verification using feature parameters based on the knowledge of document examiners, which are automatically extracted from handwritten kanji characters on a digitizing tablet. Criteria of feature selection using the evaluation measure that is obtained by modifying the measure of decidability, d-prime, is established and the criteria are applied to the evaluation measures that are calculated from learning samples. Then two classifiers based on the frequency distribution of deviations of the selected features are proposed and its design method using learning samples is showed. The effectiveness of the proposed method is evaluated by verification experiments with the database including skilled forgeries. The experimental results show that the proposed methods are effective in writer verification.
