This paper describes an off-line signature identification and verification method based on geometric feature extraction and neural network classification. In this method, signature images are simultaneously examined under several scales by superimposing onto them a set of feature extracting grids. Each grid is associated with a trained feed-forward feature network, which generates responses according to the similarity of the input pattern to the stored model pattern. A decision network combines all these responses to generate a collective confidence rating on whether the input is genuine. The system implemented based on this method has been tested with a database containing over 3000 genuine and forgery signature images belonging to 21 signature classes. Experimental results indicate that the system can correctly identify their classes and distinguish a large majority of forgeries.
展开▼
