An extended-shadow-code based approach for off-line signature verification. II. Evaluation of several multi-classifier combination strategies

For pt.I see Proc. 12th ICPR, p.450-3. In a real situation, the choice of the best representation R(γ) for the implementation of a signature verification system able to cope with all types of handwriting is a very difficult task. This study is original in that the design of the integrated classifiers E(x) is based on a large number of individual classifiers e_k(x) (or signature representations R(γ)) in an attempt to overcome in some way the need for feature selection. In this paper, the authors present a first systematical evaluation of a multi-classifier-based approach for off-line signature verification. Two types of integrated classifiers based on kNN or minimum distance classifiers and 15 types of representation related to the ESC used as a shape factor have been evaluated using a signature database of 800 images (20 writers×40 signatures per writer) in the context of random forgeries